Every insurance policy is a bet. You pay a small certain amount today (the premium) in exchange for a large uncertain payment from the insurer if something bad happens. The insurer is taking the opposite side of that bet - collecting your premium today in exchange for a contingent liability that may or may not crystallise.
For the insurer to stay in business, that bet has to be priced so that - across millions of customers - premiums collected exceed claims paid. In expected value terms, the policy must have negative expected value for you and positive expected value for them. That is the entire industry in one sentence.
This is not a moral problem. It is the mechanism by which an insurer can credibly promise to pay you £200,000 if your house burns down. The negative EV is the price of that promise. What matters is understanding the maths underneath, so you can spot when the price is fair and when it is not.
## Step One: The Actuarial TableActuarial tables are the mathematical core of every insurance product. At their simplest, they are tables of empirical probabilities - for life insurance, the probability that a person of a given age, sex, and health profile will die within the next year. The UK Office for National Statistics publishes the national life tables that every UK life insurer uses as a starting point; the ONS National Life Tables are public data.
For a 40-year-old non-smoker in good health, the published one-year mortality probability is roughly 0.0015 - about 1.5 deaths per 1,000 people in that bracket per year. An insurer offering a £200,000 one-year term policy needs at minimum:
Expected payout = 0.0015 × £200,000 = £300
That is the pure premium - the actuarially fair price of the bet. The actual premium will be higher because the insurer also has to cover operating costs, regulatory capital, profit margin, and the cost of being wrong about the probability. The difference between the actuarially fair price and the price you actually pay is called the loading, and it is where insurer profit comes from.
- **Insured profile**: 40-year-old non-smoker, no major conditions - **Mortality probability (1 year)**: ~0.0015 (ONS national life table) - **Sum insured**: £200,000 - **Pure premium (EV of payout)**: 0.0015 × £200,000 = £300 - **Typical actual premium**: ~£480/year - **Loading (insurer margin + costs)**: ~£180 = 60% of pure premiumThe table above is illustrative - actual UK term-life pricing in 2026 varies by underwriter and product - but the structure is universal. Every personal-lines insurance product (motor, home, travel, pet) is built the same way. The insurer estimates the probability of a claim, multiplies by the expected claim size, and adds a loading.
The interesting questions are not about the formula. They are about the inputs. How does an insurer know the right probability for you specifically?
## Step Two: Risk Pooling and the Law of Large NumbersAn insurer cannot predict whether you specifically will die next year, crash your car, or claim on your travel policy. The 0.0015 mortality rate above does not mean a 0.0015 chance for you - it is an average across a population of similar people. Some are healthier than the average and will not die; some are quietly sicker and will. The insurer does not know which.
What saves the insurer is the law of large numbers - a result from probability theory which says that as the number of independent trials increases, the average outcome converges to the expected value. With one customer, an insurer faces enormous variance: they either pay £200,000 or they pay £0. With 100,000 customers in the same risk class, the number of claims is extremely predictable.
If the true mortality rate is 0.0015, then with 100,000 policies sold the insurer expects 150 claims. The standard deviation of the number of claims is √(100,000 × 0.0015 × 0.9985) ≈ 12.2. So the actual number of claims will land somewhere between roughly 126 and 174 in 95% of years. Total payouts land between £25.2m and £34.8m on a pure-premium revenue of £30m. Profit is reliable; ruin is rare. This is the same reason a casino is happy to take many small bets but would never take one bet for its entire bankroll - see position sizing for the formal version of the argument.
For insurance to work, the risks must be:
When any of these break down, insurance breaks down with them. Earthquake insurance and pandemic-business-interruption cover are difficult to write because losses are correlated - when one customer claims, many do, and the law of large numbers fails. This is the risk vs uncertainty distinction in action: routine motor and home claims are risk (probabilities are known and stable); pandemic-induced commercial losses are uncertainty (probabilities cannot be reliably estimated from prior data).
## Step Three: Adverse SelectionImagine an insurer selling identical £500/year health policies to everyone. Healthy 30-year-olds will look at the price, decide it is bad value, and decline. Chronically ill 60-year-olds will look at the same price, see a bargain, and rush to buy. The pool of people who actually buy is therefore worse than the average of the population the insurer based its price on. Claims come in higher than expected; the insurer raises premiums; even more healthy people leave; the cycle accelerates. This is adverse selection, and it has destroyed insurers that ignored it.
The academic foundation here is George Akerlof's 1970 paper The Market for Lemons, which won him a share of the 2001 Nobel prize in economics. Akerlof's insight: when buyers know more about themselves than sellers do, markets can collapse entirely. Insurers prevent this collapse with three tools:
Underwriting questions. A life insurer will ask about your medical history, smoking, weight, profession, and family history. A motor insurer will ask about driving record, claims history, garaging, and annual mileage. The questions are designed to extract the private information that you have and they do not.
Risk-class pricing. Once they have extracted information, insurers price you into a risk class with similar customers and charge the actuarial rate for that class. A 40-year-old smoker pays substantially more than a 40-year-old non-smoker because the data shows they should.
Mandates and pools. When adverse selection cannot be priced around, the only fix is to force the healthy to buy too. This is why every developed country mandates motor third-party insurance, and why countries with private health insurance (the United States most prominently) typically pair private cover with mandates or subsidised pools.
The screening process produces a base rate for each risk class. Customers often complain that insurer pricing 'unfair' on individual grounds - I have never made a claim, why am I paying so much? The answer is that insurers are pricing your risk class, not you, because that is the only thing the law of large numbers lets them price reliably.
## Step Four: Moral HazardAdverse selection is about who buys. Moral hazard is about how people behave once they have bought. It is the second great information problem in insurance, and it has its own Nobel-winning literature (Bengt Holmström, 2016 prize, on contract theory).
The basic mechanism is simple: insurance reduces the cost of bad outcomes, which reduces the incentive to avoid them. Drivers with comprehensive cover may park more carelessly. Skiers with travel insurance may take a riskier slope. Restaurant owners with business-interruption cover may invest less in fire suppression than if every pound of loss came out of their own pocket. None of this is fraud - it is just the rational response of a person whose private cost of bad outcomes has been partly externalised to the insurer.
Moral hazard is invisible at the individual level but visible in the aggregate. Insurers manage it with three primary tools:
Excess (deductible) and co-insurance. If you pay the first £500 of every claim, you still have skin in the game. The bigger the excess, the more residual risk you carry, and the cheaper the policy.
Limits and exclusions. Most travel policies cap medical evacuation at a defined limit. Most home policies exclude wear and tear, gradual damage, and 'unoccupied' losses. These narrow what counts as a covered loss and reduce the insurer's exposure to behaviours they cannot easily monitor.
Bonus-malus and experience rating. Motor no-claims discounts are the textbook example: every claim-free year reduces your premium, every claim raises it. Customers face direct future-cost consequences of claiming, partially restoring the incentives that insurance otherwise dilutes.
The deeper point is that insurance is not just a money transfer - it is a contract that has to be designed to keep the underlying probabilities stable. If insurance changes behaviour enough that the probability of a claim doubles, the actuarial pricing collapses. The whole machine has to be engineered to prevent that.
## Step Five: Investment FloatThere is one more piece of the picture, and it is a big one. Premiums are paid up front. Claims are paid later - sometimes years later for liability and life lines. In the gap, the insurer holds your money. That money is called the float, and it can be invested.
For short-tail business (motor, home), the float is held for months. For long-tail business (life, annuities, employer's liability), it is held for decades. Warren Buffett's letters to Berkshire Hathaway shareholders are the most accessible exposition of the economics of float. Berkshire's insurance subsidiaries have, in many years, generated negative underwriting cost - meaning Berkshire was effectively paid to hold investable float - while also earning the investment return on that float. That double layer is why insurance has historically been a structural source of compounded capital for the patient.
For the customer, the existence of investment float is mostly invisible. But it explains why insurers are willing to write some lines of business at very thin underwriting margins (or even small underwriting losses): the float economics make the deal work even when the policy itself does not.
## Why You Should Buy Insurance Even Though It Has Negative EVOnce you understand the maths, the natural question is: why play the game? If insurance has negative expected value by construction, why does any rational person buy it?
The answer is that expected value is not the only thing that matters when stakes are large relative to wealth. The mathematical concept here is utility - the subjective value you assign to outcomes - and the key empirical observation is that for most people utility is concave in wealth. The loss of £200,000 hurts roughly fifty times as much as the gain of £200,000 helps, because losing your house ruins your life in a way that a windfall does not symmetrically improve it.
This interacts with loss aversion - the well-documented behavioural pattern where losses loom about twice as large as equivalent gains in subjective experience. But even setting loss aversion aside, the objective concavity of utility justifies paying a premium to eliminate catastrophic outcomes.
Formally: a rational decision-maker should buy insurance whenever the utility-adjusted expected value of insuring exceeds the utility-adjusted expected value of self-insuring. For high-frequency low-severity risks (lost phone, dented bumper), self-insurance usually wins - the loading on small policies is high in percentage terms, and the loss is easily absorbed. For low-frequency high-severity risks (house fire, premature death with dependants, long-term disability), insurance usually wins - the loading is small relative to the catastrophic downside it removes.
- **Phone insurance (£10/month, £400 phone)**: Usually negative-EV with high loading. Self-insure. - **Extended warranty on consumer electronics**: Famously high-margin product line. Self-insure. - **Annual travel insurance for a healthy adult**: Cheap loading, covers a true tail risk (medical evac). Buy. - **Term life insurance with dependants**: Catastrophic downside, cheap when young & healthy. Buy. - **Buildings insurance on your only home**: Single largest tail risk most households face. Buy. - **Pet insurance**: Mixed - depends on excess, exclusions, and your cash buffer.The rule of thumb is straightforward: insure against losses you cannot absorb; self-insure against losses you can. The smaller the loading and the bigger the tail, the better the buy.
## Where Insurer Pricing Goes WrongThe pricing machine described above is impressive, but it is not perfect. Three failure modes recur often enough to be worth knowing about.
1. Stale historical data. Actuarial pricing assumes the future will look like the past. When the underlying probability is shifting - climate-driven flood frequency, motor accident rates as cars get safer, mortality during a novel pandemic - published rate tables lag. Customers in shifting risk classes can be over- or under-priced for years before the tables catch up.
2. Crude risk classification. Insurers price the class, not the individual. Within any class, some customers are genuinely lower-risk than the class average; others are higher. If you have private information that you are a lower-than-average risk in your class (no claims history, advanced driving certificate, garaged car), you are subsidising the worse risks in your class. Telematics motor policies are an attempt to break the class down further and price closer to the individual.
3. Anchoring on renewal premiums. UK motor and home insurers have historically loaded renewal premiums on existing customers - what the FCA called the 'loyalty penalty'. The 2022 FCA rules now prohibit charging existing customers more than equivalent new ones, but anchoring bias still does work for the insurer: customers who do not shop around accept the renewal quote and overpay. The empirical advice is universal: get three quotes at every renewal, and treat the renewal letter as a starting bid, not a final price.
A fourth (less common) failure mode is screening that misses a key signal. A medical underwriter who does not ask about a relevant family-history factor is effectively pricing without the information that would put you in a higher-risk class - a false-positive paradox-style mismatch between what is being measured and what actually matters.
## The Mental Model in One Paragraph Insurance is a probability-pricing business. Insurers estimate the probability of a covered loss for each class of customer, multiply by the expected loss size, and add a loading for costs, capital, and profit. The law of large numbers makes the aggregate cost of claims predictable even though no individual claim is. Adverse selection and moral hazard are managed with underwriting, pricing tiers, excesses, exclusions, and experience rating. Investment float is a second profit engine on top of underwriting margin. For you as a buyer, the only number that matters is whether the loss being insured against is one you cannot absorb. If it is, the negative expected value is the price of removing variance you cannot live with - and that is a rational thing to pay for. **Q: If insurance has negative expected value, is buying it irrational?** No. Expected value alone is the right criterion only for repeated decisions with manageable stakes. For one-off events with potentially catastrophic outcomes, accepting a negative-EV contract that eliminates variance can be the rational choice. This is the same logic that justifies wearing a seatbelt on a single journey: the EV of a seatbelt-fastening minute is trivially negative in isolation, and decisively positive when you average over the rare but catastrophic outcome it prevents. **Q: How big is the loading on a typical UK personal-lines insurance policy?** Loading varies enormously by line. Term life insurance for healthy young adults is typically priced at a 30-50% loading over the pure premium. Buildings insurance is similar. Phone insurance and extended warranties on consumer electronics are at the other extreme - loading commonly exceeds the pure premium itself, which is why consumer protection advice almost always says self-insure. Travel insurance varies widely by underwriter and policy. **Q: Why do insurers care about my postcode for motor or home insurance?** Postcode is one of the most predictive variables in actuarial models because it correlates with theft and accident rates, flood and subsidence exposure, and rebuild costs. The insurer is not making a moral judgement about your area - they are using a stable empirical signal that the published claims data supports. Telematics policies attempt to replace this proxy with direct measurement of how you actually drive. **Q: What is the difference between insurance and gambling?** Mathematically they are the same kind of contract - a contingent payoff for a premium - but the direction of variance is opposite. Insurance reduces the variance of your overall financial position by transferring tail risk to the insurer. Gambling increases variance by adding new uncertainty to a position that did not previously have it. The same math, used for opposite ends. This is why most personal-finance advice treats them as moral opposites even though they are mathematical cousins. **Q: Are there situations where insurance has positive expected value for the buyer?** Rarely, and never by design. The clearest example is government-subsidised cover where the state absorbs part of the actuarial cost - UK flood reinsurance …[truncated for length] --- # When Is Travel Insurance Worth It? The EV Case Source: https://expectedvalue.co.uk/blog/travel-insurance-expected-value/ # When Is Travel Insurance Worth It? The EV Case Travel insurance is usually negative-EV, yet often the right call. The expected-value rule for when a risk is worth insuring - and when to self-insure. **Answer** Travel insurance is a negative expected value bet: on average you pay more in premium than you ever claim back. You buy it anyway because one medical emergency abroad can cost six figures - a loss you could never recover from. The rule: insure catastrophes you cannot absorb, and self-insure small, affordable risks like a cracked phone screen.Almost every insurance policy you will ever buy is a losing bet in pure money terms. The insurer sets the price so that, averaged across all its customers, it collects more than it pays out. That gap is its margin, and it comes out of your pocket. So the honest question is not whether travel insurance is good value - on average it is not - but whether the rare, ruinous outcome it protects against is one you could survive without it.
This is where expected value stops being the whole story. Expected value (the average outcome if you could repeat the same gamble thousands of times) tells you what happens on average. It does not tell you what happens to you, once, on the single trip you are actually taking.
## Why is travel insurance usually a negative-EV bet?An insurer has to charge more than the average claim, or it goes bust. Staff, marketing, capital and profit all sit on top of the expected payout. That loading is why buying insurance is, on average, a slow leak of money.
Here is an illustrative single-trip policy - the numbers are made up to show the mechanism, not quoted from any insurer. Say the policy costs £40. The distribution of what you might claim looks something like this:
| Outcome | Chance | Payout | Contribution to expected payout |
|---|---|---|---|
| No claim | 96.97% | £0 | £0.00 |
| Minor claim (delay, baggage, curtailment) | 3% | £70 | £2.10 |
| Serious injury or illness abroad | 0.02% | £30,000 | £6.00 |
| Catastrophe (repatriation + long ICU stay) | 0.0125% | £200,000 | £25.00 |
Add those contributions and the total expected payout is roughly £33, against a £40 premium. Your expected value as the buyer is about minus £7 per trip. Repeat that across a hundred trips and you would expect to be several hundred pounds down. On the maths alone, you should never buy it.
## So why buy a bet you expect to lose?Because the losing bet removes a number you could not survive. Look again at that last row: a roughly 1-in-8,000 chance of a £200,000 bill. On its own, that single risk has an expected cost of 0.0125% multiplied by £200,000, which is about £25. Tiny, on average. But there is no NHS in a Florida hospital, and an air-ambulance repatriation from the United States can genuinely run into six figures.
Now ask what a £200,000 loss would actually do to you. For almost anyone, it is not a setback - it is the end of the financial game. Savings gone, house remortgaged or sold, years of earnings committed to a debt. You cannot shrug it off and let the average work itself out over your next hundred holidays, because there is no hundred-holiday average for a household that has just been wiped out.
That is the difference between expected value and expected utility (the value of an outcome to you, which for most people rises far more slowly than the money itself, and falls off a cliff near ruin). Money you do not have is worth vastly more than money you can spare. Paying £40 to delete a £200,000 tail is a bad deal in pounds and an excellent deal in utility. We cover the full mechanism in expected value vs expected utility.
**The one-line version** Insurance converts a small, certain cost (the premium) into protection against a large, uncertain, non-recoverable loss. It is worth doing precisely when the loss is one you could not absorb - not when it is one you could. ## What makes a loss worth insuring?The deciding factor is not how likely the loss is, or even its expected size. It is whether the worst plausible outcome is recoverable. A risk worth insuring has a tail that could knock you out of the game entirely.
This is the same idea as ergodicity (whether the average across many parallel gamblers matches the average of one gambler's path through time). Expected value is an average across thousands of imaginary versions of your trip. You only ever live one. If one bad outcome ends that single path - bankruptcy, a debt you can never clear - then the comforting ensemble average never reaches you. A loss that can be absorbed is ergodic in the way that matters: you take the hit, recover, and carry on. A loss that ruins you is not. We unpack this in ergodicity explained.
So insurance earns its keep on exactly the risks that are rare, large, and non-recoverable. It is a terrible deal on risks that are common, small, and easily absorbed - which is most of the insurance the market tries to sell you.
## When should you skip insurance and self-insure?Take phone or gadget cover. Say it costs £9 a month, or £108 a year. The worst realistic outcome is a cracked screen at around £140, or a full replacement near £900. Suppose there is a 1-in-6 chance you damage the screen in a given year. The expected payout is roughly 17% multiplied by £140, plus a slim chance of a bigger loss - call it £30 a year in total.
That is an expected value of about minus £78 every year. Far worse, proportionally, than the travel policy. And crucially, the worst case here is affordable. A £140 repair, or even a £900 replacement, is annoying - it is not ruin. There is no catastrophe tail to hedge, so the utility argument that rescues travel insurance simply is not available. You are paying a hefty premium to smooth out a bump you could pay for in cash.
The rational move is to self-insure: skip the cover, keep the £108, and pay the occasional repair yourself. Over a typical phone's life you would expect to be well over £300 ahead. The same logic sinks most extended warranties and explains why a higher voluntary excess usually pays: you happily absorb the small, frequent stuff to stop bleeding premium on it.
Watch your own psychology here too. The instinct to insure a £140 screen is loss aversion - the pull of avoiding any loss at all - not rational risk management. Rational risk-aversion caps catastrophes; loss aversion insures inconveniences.
## The rule: insure the catastrophe, not the inconvenienceYou can compress all of this into one decision. Buy insurance against a risk only when both of these hold:
- **The worst plausible loss is non-recoverable.** - It would force you into debt you could not clear, wipe out your savings, or change your life. If you could write the cheque and move on, it fails this test. - **You cannot cheaply prevent or self-fund it.** - There is no simple way to avoid the risk, and you do not hold enough of a buffer to absorb it comfortably.If either test fails, self-insure and keep the premium - the negative expected value is then a pure cost with no catastrophe to justify it. A useful heuristic: if losing it would change your life, insure it; if losing it would just annoy you, do not.
## What does this mean for your next trip?The entire case for travel insurance rests on one line item: emergency medical treatment and repatriation abroad, where costs are effectively uncapped and no NHS is standing behind you. That is the catastrophe. The cancellation, delay and baggage cover bundled alongside it is the inconvenience - pleasant to have, but on its own it would be as poor a buy as phone insurance.
Two practical points for UK travellers. First, a GHIC (the Global Health Insurance Card, the UK's post-Brexit replacement for the EHIC) gives you access to state-provided healthcare in the EU on the same terms as a local - but it does not cover repatriation, private treatment, or travel outside the EU, so it is not a substitute for insurance. Second, the UK's Foreign, Commonwealth & Development Office is blunt about this: it advises taking out travel insurance for any overseas trip, and warns that being uninsured can cost you or your family many thousands of pounds. See the official gov.uk guidance on foreign travel insurance and the GHIC details.
So insure the medical catastrophe on every trip that involves one - a fortnight in the US, a ski week in the Alps, anywhere a serious injury means a private bill. Do not agonise over the baggage limit or lose sleep because the policy is, on average, a small loss. That small average loss is exactly the point: it is the price of deleting the one outcome you could not survive.
## Frequently asked questions **Q: Is travel insurance a rip-off if it's negative expected value?** No. Almost all insurance is negative expected value by design - the insurer's margin sits on top of the expected payout. That does not make it a rip-off. You are not buying a good average return; you are buying the removal of a rare, ruinous loss. For catastrophic medical costs abroad, paying a small certain premium to delete a six-figure tail is rational even though the maths says you lose on average. **Q: When is travel insurance NOT worth buying?** When the only realistic losses are affordable ones. If you are taking a cheap domestic trip you could rebook out of pocket, and there is no uncapped medical exposure, the policy is mostly inconvenience cover - and inconvenience cover is a poor buy for the same reason phone insurance is. The catastrophe tail is what justifies the premium; without it, self-insure. **Q: Does a GHIC replace travel insurance?** No. A GHIC (Global Health Insurance Card) gives access to state healthcare in EU countries, but it does not cover repatriation, private treatment, or any travel outside the EU. Repatriation - flying you home with medical support - is often the single largest cost in a serious incident, and it is exactly what a GHIC leaves out. Carry both. **Q: How do I decide whether to insure any risk, not just travel?** Apply two tests. First: would the worst plausible loss be non-recoverable - debt you could not clear, savings wiped out, a changed life? Second: can you not cheaply prevent or self-fund it? Insure only when both are true. If the worst case is merely annoying and affordable, self-insure and keep the premium. ## SourcesWorked figures in this article are illustrative models chosen to show the underlying mechanism, not prices quoted from any insurer.
## Related guides --- # Expected Value Thinking: Better Decisions Under Uncertainty Source: https://expectedvalue.co.uk/blog/expected-value-explained/ # Expected Value Thinking: Better Decisions Under Uncertainty Expected value thinking: the most important concept in decision-making under uncertainty. What it is, how to calculate it, when to apply it. ## What Is Expected Value? **Answer**Expected value is the long-run average outcome of a decision, weighted by each outcome's probability. A bet with 50% chance of +£100 and 50% chance of -£50 has an expected value of +£25 per play. Every important decision under uncertainty is implicitly an EV calculation; making them explicit beats most heuristic alternatives over time.
Expected value (EV) is the weighted average of all possible outcomes of a decision, where each outcome is weighted by its probability of occurring. It tells you what you'd get on average if you could repeat a decision infinitely many times.
The formula is straightforward:
EV = Σ (Probability of Outcome × Value of Outcome)
Or in plain English: multiply each possible result by how likely it is, then add them all up.
This deceptively simple concept is the foundation of rational decision-making under uncertainty. It's how insurance companies price policies, how venture capitalists evaluate startups, and how poker professionals decide whether to call a bet. Once you internalise expected value thinking, you'll never look at decisions the same way again.
## A Simple Example: The Coin Flip GameSuppose someone offers you this game: flip a fair coin. Heads, you win £200. Tails, you lose £100. Should you play?
Let's calculate the expected value:
The expected value is positive - £50 per flip on average. You should play this game every single time it's offered to you. Not because you'll win every flip (you won't), but because mathematics guarantees that over many repetitions, you'll average £50 profit per game.
This is the key insight: a positive expected value decision is correct even when the individual outcome is uncertain. You're not predicting what will happen; you're identifying which choice has the mathematical edge.
## Real-World EV Calculations - **Annual premium**: £350 - **P(major claim) per year**: ~0.5% - **Average major claim payout**: £45,000 - **EV of insurance to you**: 0.005 × £45,000 - £350 = -£125/year - **Why buy it anyway?**: Ruin avoidance - a £45k loss could be catastrophicThe insurance example reveals something crucial: expected value alone doesn't always determine the right decision. Insurance has negative EV for the buyer (that's how insurers make money), but it's still rational to buy because it eliminates the small probability of a catastrophic loss. Insurance also illustrates a deeper point: EV calculations only work when you can actually estimate the probabilities - when you can't, you're dealing with uncertainty rather than risk, and the risk vs uncertainty distinction becomes essential.
This is where utility theory enters the picture. A £45,000 loss hurts far more than a £45,000 gain helps. When stakes are large relative to your wealth, you should be willing to pay a premium to reduce variance - even at the cost of negative EV.
The rule of thumb: maximise EV when you can absorb the variance; buy insurance (accept negative EV) when a bad outcome could ruin you. For a mathematical framework that tells you exactly how much to risk, see our guide to the Kelly Criterion.
For a deeper look at why this matters - and the cases where expected value gives the wrong answer entirely - see our guide to ergodicity and time averages.
## EV in Investing: Why the Stock Market Rewards PatienceConsider two investment options:
Option A: Savings account - guaranteed 4% annual return, no variance.
Option B: Global equity index fund - expected return of ~8% per year, but with annual volatility of roughly ±15%.
In any given year, the index fund might return +25% or -10%. But the expected value is clear: 8% beats 4%. Over 30 years, the difference is staggering.
The positive-EV choice (equities) triples your outcome over three decades. The key requirement is that you can tolerate the short-term variance without being forced to sell at the bottom. This is why emergency funds exist - they let you stay in positive-EV investments during temporary downturns.
## EV in Career Decisions - **Current job (certain)**: £75,000/year salary - **Startup: P(failure) = 70%**: £55,000/year for 2 years, then job hunt - **Startup: P(moderate success) = 25%**: £55k salary + £200k equity over 4 years - **Startup: P(big win) = 5%**: £55k salary + £1.5M equity over 5 years - **4-year EV of startup**: 0.70×£110k + 0.25×£420k + 0.05×£1.775M = £271k - **4-year EV of current job**: £300,000 - **EV difference**: -£29k (current job wins on pure EV)Interesting - in this hypothetical, the current job actually has slightly higher expected value. But the analysis doesn't end there. Several factors might tip the scales:
The EV framework doesn't give you a mechanical answer to every life decision. What it does is force you to be explicit about your assumptions - the probabilities, the payoffs, and what you're optimising for. That clarity alone is worth the exercise.
## EV in insurance: should you buy the warranty?Insurance is the EV problem that most people get wrong on autopilot, because the framing has been engineered against you. Every extended-warranty pitch, every "protect your phone" upsell, every airport travel-insurance kiosk runs on the same trick: vivid loss aversion paired with hidden probabilities. Walking the numbers through an EV calculation usually reveals the answer.
Take a £30 extended warranty on a £400 laptop. The retailer's data (typically not disclosed but reconstructable from manufacturer reliability reports) shows roughly a 5% chance the device needs a covered repair within the warranty window, with average repair cost around £180. The EV calculation runs:
The warranty has negative expected value, which is exactly what you would expect for any retail insurance product - the retailer needs to charge more than the actuarial cost of payouts to make a profit. This is true for almost all extended warranties, almost all device-protection plans, and most travel insurance bought at airport kiosks.
The same maths runs the other way for genuinely catastrophic, low-probability risks. House insurance has negative EV in pure pound terms - the average homeowner pays more in premiums over a lifetime than they collect in claims - but the right side of the equation is no longer just "money". A total house loss is unrecoverable from savings for most people, so the question is not "is the EV positive?" but "can I absorb the worst outcome?". That is the cleanest practical example of where EV stops being the right framework on its own - the topic of the next section.
A useful heuristic: buy insurance only for losses you cannot personally absorb, and walk away from insurance offered on losses you can. A broken £400 laptop is annoying but recoverable; a burnt-down house is not. The £30 warranty is mathematically wrong, the £600 annual house premium is psychologically right.
## My bets are positive EV - so why am I losing?This is the most common complaint about expected value, and it deserves a straight answer. You can make five genuinely good decisions - all of them positive EV - and lose on every single one. Nothing went wrong with the maths. Expected value is a statement about the average of many repetitions, and in a short sequence the average hasn't had time to assert itself. The name for the spread around the average is variance.
The two numbers answer different questions. Expected value tells you which way a repeated decision drifts; variance tells you how wide the swings are on the way. A high-variance positive-EV bet can put you well down for a long stretch, and a low-variance negative-EV product (most insurance, for example) can feel painless for years. Judging a decision by a handful of outcomes confuses the weather with the climate. The formula page works through the arithmetic, and the law of large numbers is the reason the long-run average eventually dominates: the more independent repetitions you get, the closer the running average moves towards the expected value.
## Expected value in fractional odds: the bookmaker's marginUK bookmakers still quote fractional odds, and expected value translates directly. Odds of 5/1 mean five pounds of profit for every one pound staked if you win. A fair price on a coin flip would be evens (1/1): stake £1, win £1, expected value exactly zero. What you'll actually find on a two-way market is a price like 10/11 on both sides: stake £11 to win £10. Run the arithmetic: half the time you win £10, half the time you lose £11, so the average is minus 50p for every £11 staked - about minus 4.5% of your stake, on a market with no uncertainty about the fairness of the coin.
That gap between fair odds and offered odds is the margin (or overround). Convert each side's odds to an implied probability and a fair book sums to 100%; a real book sums to more, and the excess is the bookmaker's expected share of every pound staked, whichever side wins.
- **1/1 (evens)**: Decimal 2.00 · implied 50.0% - **10/11**: Decimal 1.91 · implied 52.4% - **5/6**: Decimal 1.83 · implied 54.5% - **2/1**: Decimal 3.00 · implied 33.3% - **5/1**: Decimal 6.00 · implied 16.7% - **10/11 both sides of a coin flip**: Implied total 104.8% - the extra 4.8% is the marginTwo practical consequences. First, a bet is only positive EV when your probability estimate beats the implied probability in the price by more than the margin - being roughly right isn't enough. Second, the margin applies before any judgement about the event itself, which is why consistently beating the book is hard: you start every bet a few percent behind. The bookmaker-margin calculator lets you measure the overround on any real market. The one structural exception is bookmaker sign-up promotions: a free bet hands you the margin instead of charging it, which is the basis of matched betting - a low-variance way to capture that one-off edge.
## How common UK gambles compare on expected valueOnce you can price a bet, the everyday products sort themselves quickly. Lottery tickets are structurally negative EV - a lottery pays out less than it takes in by design, which is what funds the prizes and the good causes; the full arithmetic is in why lottery tickets always lose. Premium Bonds are the interesting middle case: the advertised prize-fund rate describes the average pound, while the typical holder earns less than that average because the jackpot skews the distribution - mean and median part company. Bookmaker and casino prices carry the margin you've just seen. And if you're wondering whether positive expected value exists anywhere at all for an ordinary punter, this guide covers the few places it genuinely appears - none of them are get-rich schemes, and all of them still come with variance.
## When EV is the wrong tool: expected utility theoryEV assumes that each unit of outcome (each pound, each minute, each life-year) is worth the same as every other unit. In a casino with infinite repeats and small stakes that assumption is fine; in real life it often isn't. A £100,000 windfall changes a 25-year-old's trajectory in a way a second £100,000 added to a £10m portfolio doesn't.
Expected utility theory (EUT) replaces "the average pounds" with "the average happiness", and the difference matters whenever the outcomes are large relative to your existing wealth or large in irreversibility. A 50/50 bet to either win £1m or lose £900k has a positive EV (+£50k), but most people would refuse it - losing £900k crosses an irreversible threshold (lose the house, change of life) that winning £1m cannot fully compensate for. EUT captures that intuition directly; EV does not.
Practically, this is why the Kelly criterion uses log-utility for bet sizing (covered in our Kelly criterion guide), why investment advice rotates risk allocation with age, and why pensions are taken as annuities rather than lump sums by most retirees. Our expected value vs expected utility primer covers the formal distinction; for everyday use the rule of thumb is: use EV when the outcomes are small relative to your wealth, and you can afford to repeat the bet many times. Use utility thinking when the outcomes are large, irreversible, or you only get one shot.
This is also where the insurance heuristic from the previous section comes from. House insurance has negative EV but positive expected-utility because the worst-case loss falls in the steep, lifechanging part of your utility curve. Extended-warranty insurance has negative EV and negative expected-utility because the worst-case loss is small enough that the curve is still flat there. EV gives you the cash answer; utility tells you whether to care.
Try the numbers yourself: the expected value calculator walks you through entering outcomes and probabilities, summing the EV, and comparing two competing decisions side by side. It is the fastest way to make the EV framework concrete on a decision you're actually facing.
## Common Mistakes in EV ThinkingMost failures of EV thinking are not arithmetic failures - they are cognitive-bias failures. The 12 cognitive biases that wreck probability estimates pillar covers the underlying psychology; the four most common bias-driven EV mistakes are:
**1. Ignoring the probability side of the equation** People fixate on the size of outcomes and neglect their likelihood. A lottery jackpot of £50 million sounds life-changing, but at odds of 1 in 45 million, the EV of a £2 ticket is approximately -£0.89. The big number is irrelevant if the probability is vanishingly small. **2. Treating one-shot decisions as if they're repeated games** EV is most powerful in repeated decisions. For true one-off events (selling your house, choosing a university), you need to supplement EV with considerations about variance and irreversibility. **3. Using wrong probability estimates** Garbage in, garbage out. If you estimate a 90% chance of your business succeeding when the base rate for new businesses is 20%, your EV calculation will be wildly optimistic. Always start with base rates and adjust from there. Building calibrated probability estimates - where your stated 70% confidence actually corresponds to a 70% hit rate - is the foundation of useful EV analysis. We cover the most common probability pitfalls in Thinking in Probabilities: Why Your Brain Is Bad at Risk. **4. Forgetting about hidden costs and externalities** The true cost of a decision includes time, stress, opportunity cost, and impact on other areas of your life. A positive-EV side project that destroys your sleep and relationships might have negative *total* value. **4. Stopping at the first-order outcome.** EV math is right; the inputs are often wrong because people only score the immediate consequence. The true EV of a decision includes how the system *responds* to that decision - competitors, regulators, your own future self. See second-order thinking: how to see around corners for a practical framework on extending EV inputs past the obvious. ## How to Apply EV Thinking DailyYou don't need a spreadsheet for every decision. Here's how to integrate EV thinking into everyday life:
Ask: "What are the possible outcomes, and how likely is each?" - Just framing a decision this way forces you out of gut-reaction mode and into analytical thinking.
Look for positive-EV bets you're avoiding due to loss aversion. - Many people avoid investments, career moves, or conversations because the downside feels scary, even when the expected value is clearly positive. Our probabilistic framework for career decisions walks through this in detail.
Seek repeatable positive-EV situations. - The power of EV is strongest in repeated games. If you can find situations where you have an edge and can play many times (networking, content creation, investing), the law of large numbers works in your favour. One real-world venue where EV maths is laid out explicitly is the prediction markets - see how prediction markets work for a tour of how contract prices map directly to probabilities.
Update your probabilities when new information arrives. - EV calculations are only as good as your inputs. Be willing to revise estimates as reality provides feedback. Our guide to thinking in probabilities covers practical techniques for better calibration.
## The complete expected-value toolkitExpected value is a small idea with a lot of branches. These guides go deeper on the pieces that matter most, and the interactive calculators let you put real numbers in.
**Try the numbers yourself** Two of our calculators are pure expected-value tools: the salary-negotiation EV …[truncated for length] --- # Nassim Taleb's Ideas: Black Swan, Antifragile & More Source: https://expectedvalue.co.uk/blog/nassim-taleb-ideas-explained/ # Nassim Taleb's Ideas: Black Swan, Antifragile & More A guide to Nassim Taleb's key ideas - Black Swans, Antifragile, Skin in the Game, fat tails, barbell strategy and the Lindy effect - and where to start.This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
**Answer**Nassim Taleb's core argument across Black Swan, Antifragile, and Skin in the Game is that rare high-impact events drive most of what matters, that systems get stronger from controlled stress (antifragile) rather than just surviving it (robust), and that anyone making predictions should bear personal cost when they are wrong. Take asymmetric bets, avoid ruin, and ignore forecasters who do not have skin in the game.
Where to buy these books (Amazon.co.uk):
Nassim Nicholas Taleb's ideas - Black Swans, Antifragile, Skin in the Game, fat tails, the barbell strategy, the Lindy effect - sit underneath a great deal of how serious people now think about risk, uncertainty and decision-making. Taleb himself wrote five books over almost two decades to lay them out, collectively called the Incerto. This guide explains each of the headline concepts in plain English, shows how they fit together, points to the right book to start with depending on what you want from him, and ends with the criticisms that you should weigh alongside them.
## Who is Nassim Nicholas Taleb?Taleb is a Lebanese-American mathematical statistician, former options trader, and probability theorist. He spent two decades on Wall Street trading options - specifically, the kind of options that pay off in rare crashes - before retiring to write and to hold an academic post at NYU's Tandon School of Engineering. The trading background matters: every idea in the Incerto traces back to a working life spent betting on rare events.
His five-book Incerto series - Fooled by Randomness (2001), The Black Swan (2007), The Bed of Procrustes (2010), Antifragile (2012) and Skin in the Game (2018) - is one extended argument: that we systematically underestimate the role of rare, large events; that we mistake the absence of evidence for evidence of absence; that institutions which hide their exposure to those events while collecting fees in the meantime are causing slow-motion catastrophes; and that the only honest way to behave under deep uncertainty is to keep skin in the game and to position yourself so that the rare event helps you instead of destroying you.
You don't have to read the Incerto chronologically - Taleb himself says the books are 'one continuous logical argument' and that any book is a starting point - but they do build on each other.
**Key terms in one paragraph** A Black Swan is a rare, high-impact, hard-to-predict event. A 'fat tail' is a probability distribution where extreme outcomes are far more likely than a normal-distribution model would suggest. To be Antifragile is to gain from disorder rather than just survive it. Skin in the Game means bearing the downside of decisions you make for others. The Barbell Strategy means combining extreme safety with extreme risk and avoiding the middle. The Lindy Effect says that for non-perishable things, the longer they have already lasted, the longer they are likely to last. ## What is a Black Swan event, in Taleb's words?A Black Swan, in Taleb's definition, has three properties. It is an outlier (lying outside the realm of regular expectations), it carries an extreme impact, and - crucially - it is only rationalised after the fact, never before. The 9/11 attacks, the 2008 financial crisis, the rise of the internet, COVID-19, the discovery of penicillin: all Black Swans by Taleb's test, all explained convincingly only in retrospect.
Two consequences follow. First, our forecasting tools - built on bell-curve assumptions and historical regularities - systematically miss what matters most. The variance you see in 100 years of stock returns tells you almost nothing about the variance you'll see in year 101 if year 101 is a Black Swan year. Second, our explanations of past events are largely after-the-fact storytelling. Taleb calls this the 'narrative fallacy'.
The practical takeaway is not that you should try to predict Black Swans (you can't, by definition). It's that you should structure your life and your portfolio so that one Black Swan can't end you, and so that the rare positive Black Swans - a startup that takes off, a relationship that compounds, a piece of work that finds an audience - get a chance to find you. This is closely linked to our guide to the false positive paradox and the wider problem of risk vs uncertainty in Knightian terms.
## What does Fooled by Randomness argue?Taleb's first major book argues that successful traders, fund managers, executives and entrepreneurs are far more often the lucky ones than they (or we) realise. He gives the example of a trader who is 90% likely to make $100k a year and 10% likely to lose $1m: the expected value is sharply negative, but the trader will look like a genius for years before the bad year arrives. Survivorship bias - the simple fact that we only see the winners - does the rest.
The book's deepest practical idea is that the quality of a decision and the quality of an outcome are different things. A bad decision can produce a good outcome (the gambler who 'wins big' on a -EV bet) and a good decision can produce a bad outcome. To get better at decisions, you have to evaluate the process, not the result. This is the foundation that decision journals and the pre-mortem sit on top of, and it pairs naturally with our guide to expected value.
Read this first if you want one Taleb book that sharpens how you think about probability in your own life rather than the global financial system.
## What does antifragile actually mean?If The Black Swan diagnoses the problem, Antifragile proposes the response. Taleb introduces a three-way distinction:
Taleb's argument is that the language we use about risk only has the first two - we talk about 'risk management' as keeping things robust - and we have therefore been blind to the antifragile category. Once you can see it, the goal of policy, investing, parenting and personal development changes: you stop trying to protect yourself from volatility and start trying to position yourself so that volatility makes you stronger.
The most cited practical idea from Antifragile is via negativa: the asymmetry between adding good and removing bad. Removing one toxic relationship, one bad habit, one unpaid commitment is reliably positive; adding a new positive one is more variable. Antifragile systems improve mostly by subtraction.
## What is Taleb's Barbell Strategy?The Barbell Strategy is the operational core of Antifragile. The argument: in domains with fat tails, the middle of the risk spectrum - moderate-risk investments, moderate-risk careers, moderate-risk decisions - is where you collect ordinary returns while remaining exposed to the rare ruinous event. You're better off splitting your exposure between extreme safety on one side (cash, treasuries, a stable salary) and extreme upside on the other (concentrated bets with bounded downside).
For an investor, the canonical Taleb barbell is roughly 85-90% in cash or short-term government bonds and 10-15% in highly speculative bets where the most you can lose is what you put in but the upside is uncapped. This dominates a 'balanced' 60/40 portfolio for someone who actually believes in fat tails - the 60/40 is exposed to the same fat-tail crash as a 100% equity portfolio, just slightly less so, while sacrificing the option for serious upside.
The barbell extends well beyond investing. A career barbell is 'reliable salary + side bets'. A health barbell is 'extreme stability in sleep and food, occasional intense exertion'. A research barbell is 'bread-and-butter incremental work + a small tail of bold long-shot ideas'. The principle is the same: avoid the middle, where you collect ordinary returns while staying exposed to fat-tail downside. This connects naturally to the Kelly Criterion - Taleb's barbell is roughly what you'd build if you took Kelly seriously and accepted that real-world distributions are fatter-tailed than your model assumes.
## Why does Taleb insist on skin in the game?The fifth Incerto book is a moral argument. Taleb's central claim: people who advise, predict, regulate or rule on behalf of others should bear a meaningful share of the downside if they're wrong. The forecaster who confidently predicts a recession that doesn't happen pays no price. The bank executive who runs the firm into bailout territory keeps a bonus. The politician who supports a war from behind a desk doesn't fight it. Each is making a fragile system more fragile by removing the feedback loop that would otherwise correct them.
The book extends this in several directions. Asymmetric incentives produce asymmetric behaviour: a manager whose upside is capped but whose downside is bounded by 'getting a new job' will reliably take less risk than the situation actually requires. Minority rules: a small intransigent minority can impose its preferences on a much larger flexible majority because the cost of accommodating them is low (kosher food being served to non-kosher guests, allergy menus). The expert problem: people in fields without skin in the game (academic economists, foreign-policy commentators) consistently underperform their reputations, while people with skin in the game (plumbers, traders, entrepreneurs) become quietly better than their reputations suggest.
Read this if you want Taleb on ethics and institutions rather than on probability.
## What are fat tails and why do they break most statistics?A 'fat tail' (or 'heavy tail') is a probability distribution where extreme values occur far more often than a normal distribution would predict. The classical bell curve assumes that observations more than three or four standard deviations from the mean are extraordinarily rare; fat-tailed distributions assume nothing of the sort.
This matters because most financial, social and natural phenomena are fat-tailed. Stock returns are fat-tailed (the 1987 crash was a 22-standard-deviation event under a Gaussian model - meaning 'shouldn't happen in the lifetime of the universe' - but it happened, and similar moves keep happening). City sizes, book sales, war casualties, internet traffic, cyber-attack severity, pandemic mortality, wealth - all fat-tailed. Heights and IQ scores are not, which is part of why we wrongly imagine the Gaussian model is general.
The practical consequence: in a fat-tailed world, the average tells you almost nothing useful and the variance is unstable. A single observation can shift your sample mean by orders of magnitude. Models that assume thin tails - most of finance, most of risk management, most of regulatory capital calculations - systematically underestimate disaster risk. The 2008 financial crisis was exactly this kind of error at scale: VaR (Value-at-Risk) models priced rare events as basically impossible, and a great deal of leverage was built on top of that assumption.
Taleb's technical companion volume Statistical Consequences of Fat Tails (2020, free PDF online) is the rigorous mathematical version of the same argument.
## What is the Lindy Effect?The Lindy Effect - named after a New York deli where comedians informally observed that long-running shows kept running while new shows didn't - says: for non-perishable things, life expectancy increases with age. A book in print for 50 years is more likely to still be in print in 50 years' time than a book published last year. A technology that has lasted a century is more likely to last another century than a technology from last decade. A friendship of 20 years is more durable than a friendship of two.
The Lindy Effect is the opposite of how perishable things age - humans, dogs, strawberries - for which expected remaining life decreases with age. The asymmetry comes from selection: the things that survive long periods do so because they have already proven robust to many shocks, and that survival is information.
Practical applications: when picking a book to read, prefer Aristotle to last year's Davos panel summary. When choosing technology to depend on, prefer plain text over the latest framework. When evaluating a piece of advice, weight the source by how long it's been considered correct, not how recently it was published. Taleb uses the Lindy Effect as a heuristic for filtering signal from noise in domains where you can't easily run experiments.
## What is The Bed of Procrustes about?The third book in the Incerto is the shortest and the most stylised. It is a collection of aphorisms - short, often cutting one-liners - written in the tradition of Nietzsche, La Rochefoucauld and the Stoics. The title refers to the Greek myth of Procrustes, the innkeeper who stretched or amputated travellers to fit his bed; Taleb's argument is that we constantly do something similar with reality, forcing it to fit our theories rather than updating the theories.
The book is not where to start with Taleb. But it's a useful companion to the other four - most of the Incerto's headline ideas appear, compressed, as a single sentence somewhere in Procrustes, and reading it after the others gives you a kind of executive summary in epigram form.
## Where should you start with Taleb?Taleb says any of his books can be a starting point. In practice, three reasonable orders work:
If you want to improve your own thinking - start with Fooled by Randomness. It's the most personal, the most directly applicable, and the least mathematical. You'll come out of it more careful about confusing luck with skill in your own decisions.
If you want the headline ideas - start with The Black Swan. It's the most quoted, the most cultural, and contains the framing that the rest of the Incerto extends. It's denser than Fooled by Randomness; expect to read it slowly.
If you want practical actionable principles - start with Antifragile. It's the longest and most prescriptive, and it's the one most readers find genuinely changes how they make decisions. It assumes you've absorbed the Black Swan idea but it's not strictly required.
Skip The Bed of Procrustes until you've read at least one of the others, and skip Statistical Consequences of Fat Tails entirely unless you want the rigorous mathematics. Skin in the Game is best read after Antifragile - it's the moral and political continuation of the same argument.
For the wider canon Taleb sits in, see our list of 12 best books on probabilistic thinking and the broader 15 best decision-making books.
## What are the main criticisms of Taleb's ideas?Reading Taleb critically is part of taking him seriously. Three lines of critique recur in the academic and popular literature:
The 'unfalsifiable Black Swan' problem. By definition, Black Swans can only be identified after the fact, which means the framework can be applied to almost any large event in retrospect. This makes it powerful as a way of describing the world but harder to use as a predictive tool. Taleb's response is that prediction is exactly what the framework is rejecting; some critics find this unsatisfying.
The 'fat tails everywhere' problem. Statisticians broadly agree that fat tails are common and that Gaussian assumptions are often wrong. But the strong Talebian claim - that effectively all important domains are fat-tailed and that thin-tailed reasoning is therefore systematically dangerous - overstates a real-but-narrower truth. Many domains (engineering tolerances, controlled industrial processes, biology at the level of organism size) are well-modelled by thin-tailed statistics and would suffer if you treated them as fat-tailed.
The combative style. Taleb's polemical tone - frequent personal attacks on named academics, journalists and economists - distracts some readers from the underlying ideas, and has contributed to his being underrated in some academic communities where the ideas would otherwise have travelled further. Whether this is a feature or a bug depends on your taste.
None of these criticisms reduce the value of the headline ideas; they're calibration on how strongly to apply them. The Black Swan, fat tails, antifragility, the barbell, skin in the game - taken together they form one of the most useful frameworks in the modern non-fiction canon for thinking about risk and decision-making, even if individual chapters provoke disagreement.
## Frequently asked questions **Q: What does 'antifragile' actually mean?** Antifragile means gaining from disorder, not just surviving it. A muscle that grows stronger when stressed, a financial portfolio whose biggest gains come during volatile years, a startup that turns customer complaints into product improvements - all antifragile. Taleb introduced the term because English had words for fragile (breaks under volatility) and robust (survives volatility unchanged) but no word for the third category. **Q: What is a Black Swan event?** A Black Swan is a rare, high-impact event that wasn't foreseen and that gets rationalised aft …[truncated for length] --- # Thinking in Bets: Decisions Like a Poker Pro Source: https://expectedvalue.co.uk/blog/thinking-in-bets-annie-duke/ # Thinking in Bets: Decisions Like a Poker Pro Annie Duke's Thinking in Bets reframes every decision as a bet under uncertainty. The ideas of resulting, calibration, and decision groups.This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
Annie Duke's Thinking in Bets argues that every decision is a bet on an uncertain future, and that the way most people judge their decisions - by the outcomes - is the single biggest reason they don't get better at making them. Duke spent twenty years as a competitive poker player; her central claim is that poker teaches a discipline most other careers never force on you, because in poker you have to settle the same uncomfortable question on every hand: did I make a good bet, or did I just get lucky?
This guide walks through the book's main ideas - resulting, decision quality versus outcome quality, decision groups, and how to separate luck from skill - and shows how to apply them to careers, money, and the everyday choices where the feedback loop is too slow or too noisy for instinct to do the work.
Where to buy: Thinking in Bets on Amazon.co.uk →
**TL;DR** Every decision is a bet under uncertainty. Judging a decision by its outcome is called 'resulting' and it's a bug, not a feature. Calibrate beliefs in probabilities, separate the quality of your reasoning from the quality of your luck, and find a small group of honest peers to challenge your thinking. The book is most useful for people whose work involves decisions whose outcomes won't be known for months or years. ## What is the core claim of Thinking in Bets?The book's foundational move is asking you to drop the word 'decision' and replace it with 'bet'. A bet has three properties a decision often hides: there's an explicit cost, an explicit probability of payoff, and an explicit alternative use of the same resource. When you call a job change 'a decision', it's easy to skip those three things. When you call it 'a bet I'm making with two years of my life', they become impossible to ignore.
Duke's framing forces probabilistic language on choices people normally treat as binary. 'Should I quit my job?' is a binary question with a yes or no answer. 'What is the probability that quitting my job in the next twelve months produces a better outcome than staying?' is a bet - and once you've estimated that probability, you can ask the more useful question: 'Is the bet worth it given the cost?'
The shift sounds like semantics. It is not. Treating a decision as a bet imports the whole machinery of expected value calculation, calibration, and bankroll thinking. It also imports poker's relationship with information: a bet always specifies what you know, what you don't, and what you're paying to find out.
## Resulting: Why Judging Decisions by Outcomes Lies'Resulting' is Duke's term for the cognitive shortcut of evaluating a decision by what happened afterwards rather than what was reasonable to think at the time. It's the most damaging mental habit poker amateurs share with everyone else.
The textbook example is from her own career. A player makes a perfect read, calls a bet with the best hand, and loses to a one-card draw their opponent shouldn't have called with in the first place. The amateur who watched the hand says: 'You shouldn't have called.' The professional says: 'You called correctly. They played badly. The next hundred times you face that exact spot, calling makes you money.' The outcome was bad. The decision was right.
Resulting feels obvious to spot in someone else and almost impossible to spot in yourself. Three signs you're doing it:
The fix isn't to ignore outcomes - outcomes are data. The fix is to weigh outcomes against the prior probability that they would happen, which means you need to estimate that probability before the outcome arrives. This is why decision journals matter, and why decision journals appear in almost every book Duke recommends.
## Decision Quality vs Outcome QualityThe book formalises the distinction with a 2x2 matrix that's worth memorising:
Duke argues that 'bad decision, good outcome' is the most dangerous quadrant in the matrix. A bad decision that pays off teaches the brain that the bad decision was actually good - and you'll repeat it. The amateur poker player who makes a wild bluff that wins, then keeps bluffing into the same opponents until the bluffs stop working, is in this quadrant. So is the entrepreneur whose first business succeeded for reasons unrelated to their strategy, then runs the same strategy into the ground at the second.
The diagonal - good decision/bad outcome and bad decision/good outcome - is where outcome-based feedback fails entirely. Skill development requires the ability to identify which quadrant you're in, and that requires forming an opinion on the decision before the outcome lands.
## Calibration: Talking in ProbabilitiesDuke spends a chapter on the language of belief. People say things like 'I'm sure' or 'definitely' when they mean 'I think it's about 70% likely'. The cost is that overstated certainty is impossible to be wrong about - and impossible to learn from.
The fix is to attach a number to every claim you make about the future. Not a precise number - a rough one. 'I'm 80% confident this contract closes by month-end.' 'I'd put about 40% on rates being lower in twelve months.' 'There's maybe a 15% chance this product launch hits the revenue target.'
Two things happen when you do this consistently. First, you discover you've been operating with implicit confidence levels that don't survive being written down. The 'definitely' that turns into 70% on paper is information about how much certainty you actually have. Second, you build the skill of probability calibration - the trained ability to say '70%' and have your 70% predictions actually come true 70% of the time. Calibration is a learnable skill, and it's the foundation everything else in the book sits on.
## Decision Groups: Outsourcing HonestyThe book's most practical chapter argues that solo decision-making, no matter how disciplined, has hard limits. Three or four people who agree to challenge each other's reasoning will outperform any one of them on average - provided they follow specific rules.
Duke calls these groups 'truthseeking groups' and lists what they need to work:
For most readers without access to a poker training group, the practical suggestion is to find one or two people whose judgement you respect, agree on a 'meta-rule' that disagreement is welcome, and run the bigger decisions past them with the explicit ask: 'Tell me what's wrong with my reasoning, not whether you'd do it.'
## Tools the Book Hands You - **Decision journals** - Write the decision, the alternatives, your probability estimates, and the reasoning before you act. Re-read after the outcome lands. - **10-10-10 rule** - Borrowed from Suzy Welch: how will you feel about this decision in 10 minutes, 10 months, and 10 years? Surfaces decisions where short-term emotion is hijacking long-term reasoning. - **Pre-mortems** - Before committing, imagine the project failed and write the post-mortem. Forces you to surface failure modes you'd otherwise ignore. See our guide to /blog/pre-mortem-decision-making. - **Backcasting** - Pick a successful future state and work backwards to identify the choices that have to go right. Complement to the pre-mortem. - **'Wanna bet?' as a tell** - When someone makes a strong claim, mentally ask: 'Would they bet $100 on this at the odds they're implying?' If not, the claim is weaker than it sounds. ## Where Thinking in Bets Is Most UsefulThe book is strongest where the gap between decision and outcome is months or years - investing, hiring, career changes, business strategy, relationships. In those domains, instinct gets calibrated by feedback that arrives too slowly to learn from, and the techniques in the book substitute for feedback you can't get directly.
It's less useful where outcomes are fast and explicit. A surgeon doesn't need this framing for technical decisions during an operation; they get fast, clear feedback. A trader running a thirty-second-holding-period strategy gets enough data to calibrate without journals. The book is for slow-feedback decisions, which is most of the consequential ones in a non-specialist career.
The other limit is that the book is short on what to do when you're wrong. Updating beliefs in light of new evidence is treated lightly - the better book on that specific topic is Philip Tetlock's Superforecasting, which Duke draws on heavily and which we cover in our decision-making books guide.
## Should You Read It?Thinking in Bets is roughly 240 pages and a quick read - most people finish in two or three sittings. The arguments are repeated several times across the book, which makes the core ideas stick but means there's some skim-readable overlap.
It's the right book for you if: you make consequential decisions whose outcomes you won't see for a year or more; you've noticed yourself oscillating between strategies after single bad outcomes; you suspect the people around you are giving you outcome-based feedback rather than process-based feedback; or you want to start a small decision group and need a shared vocabulary.
It's not the right book if you're looking for a deep treatment of probability theory, decision analysis, or behavioural economics - for those, see our broader probabilistic thinking reading list. Thinking in Bets is a vocabulary book, and its job is to install the words 'bet', 'resulting', and 'calibration' deeply enough that they change how you think.
**Pair this with a decision journal** The single highest-leverage habit the book recommends is keeping a decision journal. Without one, the ideas stay theoretical. With one, you build a feedback loop on your own reasoning that compounds over years. Even a plain notebook works - what matters is writing your probability estimate before the outcome lands. ## Frequently Asked Questions **Q: Do I need to know poker to read Thinking in Bets?** No. Duke uses poker examples but explains every term. The book is aimed at general readers - there's no required mathematical background, and the poker hands are illustrative rather than technical. **Q: Is this just behavioural economics rebranded?** Partly. Duke draws heavily on Kahneman, Thaler, and Tetlock. Where the book adds value is in framing decisions as bets and importing the cultural norms from poker training (calibration, decision groups, outcome-blind feedback) that don't appear in the academic literature. **Q: What's the difference between 'resulting' and 'hindsight bias'?** Hindsight bias is believing you 'knew it all along' once you see an outcome. Resulting is judging the decision quality by the outcome quality. They reinforce each other but they're different errors. See our guide to /blog/hindsight-bias for the cognitive-bias side. **Q: How does this compare to Annie Duke's later book How to Decide?** How to Decide (2020) is more practical and workbook-style - it gives you the templates and exercises. Thinking in Bets is the conceptual foundation. Read Thinking in Bets first; pick up How to Decide if you want to operationalise the ideas. **Q: Where does Kelly Criterion fit in?** Duke mentions bet sizing but doesn't develop it. For position sizing under uncertainty, see our guide to /blog/kelly-criterion-optimal-bet-sizing - Kelly is the natural extension of 'every decision is a bet' to the question 'how much should I bet?'. **Q: Is this useful for investing decisions?** Yes - investing is the canonical slow-feedback decision domain. The book pairs well with Howard Marks' memos and Charles Ellis's work. A working investor will get more out of it than a reader looking for stock picks. --- # Expected Value Calculator: How to Calculate EV Step-by-Step Source: https://expectedvalue.co.uk/blog/expected-value-calculator/ # Expected Value Calculator: How to Calculate EV Step-by-Step Learn how to calculate expected value with a clear formula, four worked examples (betting, investing, insurance, career) and the common mistakes to avoid. **Answer**To calculate expected value, multiply each possible outcome by its probability and sum the results. For example, a coin flip paying £10 on heads and -£5 on tails has an EV of (0.5 x £10) + (0.5 x -£5) = £2.50 per flip. Use the calculator above to plug in your own probabilities and payouts.
Expected value (EV) is the average outcome of a decision if you could repeat it many times. To calculate it, you multiply each possible outcome by its probability and add the results. That single sentence is the entire formula - but applying it well is the difference between making good decisions under uncertainty and just hoping for the best.
This guide walks through the EV formula step by step, then works through four real examples - betting, investing, insurance, and a career decision - so you can apply it to any choice you face. If you want the conceptual background first, our expected value explained guide covers what EV means and why it matters before you start calculating.
## What expected value actually isExpected value is the probability-weighted average of all possible outcomes of a decision. If you flip a fair coin and win £10 for heads or lose £8 for tails, the expected value of one flip is £1 - even though that exact outcome never actually happens on a single flip.
What EV tells you is the average result if you could play the same scenario forever. It does not tell you what will happen on any individual decision. This distinction matters because most real decisions only happen once. A choice with positive EV can still go badly the one time you make it, and a choice with negative EV can still pay off occasionally. Over many decisions, however, EV is what you converge to - and what you should usually optimise for.
The point of calculating EV is to strip the emotional content out of a decision and look at the underlying maths. Once you can see that an insurance policy has a negative EV of -£40 per year, or that a job offer has a positive EV of +£12,000, you have a clear quantitative anchor for the decision rather than gut feeling alone.
## The expected value formulaThe formula for expected value is:
EV = (P₁ × V₁) + (P₂ × V₂) + ... + (Pₙ × Vₙ)
Where:
The formula scales to any number of outcomes. For a coin flip there are two terms. For a six-sided die there are six. For a complex investment decision there might be dozens, but the structure is identical: list every possible outcome, assign each one a probability, multiply each value by its probability, and add them together.
If you find yourself listing only the upside outcomes, you are not calculating EV - you are calculating the best case. The hardest part of EV calculation is being honest about the downside cases and their probabilities, especially the ones you would prefer not to think about.
## How to calculate EV in four steps 1. **List every possible outcome** - Write down each distinct way the decision could resolve. Be exhaustive - if your list is shorter than three outcomes, you are probably collapsing real possibilities. For an investment, this might include 'doubles in value', 'breaks even', 'loses 30%', 'goes to zero'. 2. **Assign a probability to each outcome** - Estimate how likely each outcome is. Use base rates where possible (historical data, published statistics) and your honest best estimate where data is thin. Probabilities must sum to 1.0 - if they do not, you have either missed an outcome or double-counted. 3. **Quantify the value of each outcome** - Put a number on what each outcome is worth to you. For financial decisions this is straightforward (gain or loss in pounds). For non-financial decisions you need a unit of value: years of life, hours of free time, customer accounts won, or a 1–10 satisfaction score. 4. **Multiply, add, interpret** - Multiply each outcome's value by its probability, sum the results, and you have the expected value. A positive EV means the decision is on average favourable; a negative EV means it is on average unfavourable. The magnitude tells you how favourable or unfavourable. ## Worked example 1: A simple betting calculationSuppose a friend offers you a bet on a single roll of a fair six-sided die. If it lands on a 6, you win £30. If it lands on anything else, you lose £5. Should you take the bet?
Step 1: List outcomes. There are two: 'rolls a 6' and 'rolls 1-5'.
Step 2: Assign probabilities. P(6) = 1/6 ≈ 0.167. P(1-5) = 5/6 ≈ 0.833. They sum to 1.
Step 3: Quantify values. V(6) = +£30. V(1-5) = -£5.
Step 4: Calculate. EV = (0.167 × £30) + (0.833 × -£5) = £5.00 - £4.17 = +£0.83.
The expected value of one roll is positive - about 83 pence. The bet is favourable in the long run. If you played it 100 times, you would expect to be roughly £83 ahead, on average. On any single roll you will almost certainly either win £30 or lose £5 - never exactly £0.83 - but the EV tells you the underlying maths is on your side.
This worked example shows why EV is the dominant framework in poker, sports betting and any context where the same decision recurs many times. For a deeper application of this idea, our expected value in poker guide walks through pot odds, equity and EV in real game situations.
## Worked example 2: An investment decisionYou are considering investing £10,000 in a small startup. Based on industry base rates and your own assessment, you estimate four possible outcomes over five years.
Probabilities sum to 100%. Calculation:
EV = (0.10 × £90,000) + (0.20 × £15,000) + (0.30 × £0) + (0.40 × -£10,000)
EV = £9,000 + £3,000 + £0 - £4,000 = +£8,000
The expected value over five years is +£8,000. That sounds attractive - but notice how the calculation depends entirely on those probability estimates. If the failure rate is actually 60% rather than 40%, the EV drops to -£2,000 and the decision flips. Sensitivity analysis (recalculating with different probability assumptions) is essential for real investment EV - see our risk vs uncertainty guide for why this matters when probabilities themselves are uncertain.
## Worked example 3: Should you buy this insurance policy?An extended-warranty company offers a £80 three-year warranty on a £400 laptop. Based on published failure rates, you estimate there is roughly a 12% chance of a covered failure in three years, and the average claim value would be £200.
EV = (0.12 × £120) + (0.88 × -£80) = £14.40 - £70.40 = -£56
The expected value is negative £56. On average, you lose money buying this warranty - which is why insurance companies sell warranties profitably. This does not automatically mean you should refuse the warranty. EV ignores variance, and there is a meaningful psychological case for paying a small premium to eliminate the small chance of a £200 surprise. But the calculation puts a price on that comfort: roughly £56 over three years, or £19 per year.
The same logic applies to most insurance products - health, travel, gadget, pet. The product is sold at a negative EV by design (otherwise the insurer loses money). The question is not 'is the EV positive?' but 'is the EV negative enough that I'd rather take the risk myself, or close enough to zero that the peace of mind is worth it?'
## Worked example 4: A career decisionYou have two job offers. Job A pays £55,000 and is at a stable, established company. Job B pays £45,000 base plus equity in an early-stage startup. You want to compare them on three-year EV.
Job A (stable):
EV(A) = (0.95 × £165,000) + (0.05 × £82,500) = £156,750 + £4,125 = £160,875
Job B (startup with equity):
EV(B) = (0.10 × £335,000) + (0.30 × £165,000) + (0.40 × £135,000) + (0.20 × £67,500) = £33,500 + £49,500 + £54,000 + £13,500 = £150,500
On pure EV, Job A wins by about £10,000 over three years. But the EV calculation hides important non-financial factors - career capital from startup experience, learning rate, network value, optionality. EV is the financial anchor, not the whole answer. For more on integrating non-financial factors with EV, see our probabilistic framework for career decisions.
## Common mistakes when calculating expected value1. Missing outcomes. If your probabilities don't sum to 1, you've forgotten a case. The most commonly missed outcome is the boring middle case - 'nothing much happens' - because it's not interesting to imagine. This systematically biases EV upward.
2. Anchoring probabilities to the first number you saw. If you read that 'most startups fail', you might assign a 90% failure rate without checking. Real base rates vary enormously by sector, stage and team experience - see our anchoring bias guide for why first numbers stick.
3. Conflating one decision with many. EV is the average of many trials. If you only get one shot at the decision, the variance matters as much as the mean. A +£8,000 EV with a 40% chance of losing £10,000 is not the same as a guaranteed +£8,000 - the latter is much better even though the EV is identical.
4. Ignoring ergodicity. Some negative outcomes can take you out of the game permanently (a bet that requires you to risk all your savings). In those cases, even a positive EV can be the wrong choice - a topic our ergodicity explained guide goes into depth on.
5. Treating EV as a forecast. EV is a long-run average, not a prediction of what will happen this time. The single biggest misuse of EV is presenting it as 'what will probably happen' rather than 'what would happen on average if this decision were repeated many times'.
**When EV calculations stop being useful** Expected value is the right tool when probabilities can be reasonably estimated and outcomes can be quantified. It is the wrong tool when probabilities are deeply uncertain (true Knightian uncertainty), when outcomes are non-monotonic (e.g. health interventions where 'more' isn't always better), or when one possible outcome is catastrophic and irreversible. In those cases, EV calculation can produce false precision that crowds out better reasoning. ## How to apply EV in everyday decisionsRule 1: Use EV for decisions you'll face many times. Subscription cancellations, recurring purchases, betting decisions, repeated investments - anywhere the same decision pattern recurs, EV is the highest-leverage framework. The more repetitions, the more EV converges to your actual long-run result.
Rule 2: For one-off decisions, calculate EV but also think about variance. A positive EV with a 50% chance of catastrophic loss is rarely the right call for a one-shot decision. Combine EV with the Kelly criterion for sizing decisions where you have an edge but limited bankroll.
Rule 3: When you can't calculate EV precisely, calculate it roughly. A back-of-envelope EV with rough probabilities and round numbers is often enough to flip a decision from 'feels wrong' to 'is clearly wrong' - or vice versa. Don't let the search for precise probabilities prevent you from doing the calculation at all. A rough EV beats no EV every time.
## Frequently asked questions **Q: What does a positive expected value mean?** A positive expected value means that, on average, the decision is favourable - if you could repeat it many times, you would gain rather than lose. It does not guarantee that any individual instance will be favourable. A bet with +£0.50 EV can still lose money on a single roll; the EV is the average across many rolls. **Q: Can expected value be negative?** Yes. A negative expected value means the decision is on average unfavourable - you would lose money or value over many repetitions. Most insurance products, lottery tickets and casino games have negative EV by design. That doesn't always make them wrong - variance and peace of mind can justify a small negative EV - but it does mean you're paying for that benefit. **Q: How precise do my probabilities need to be?** Less precise than you might think. EV is most useful as a directional signal. Whether the failure probability is 35% or 45% rarely flips the sign of the EV - you usually only need probabilities accurate to within ±10 percentage points to make a confident decision. The exception is when EV is close to zero, where small probability changes can flip the conclusion. **Q: What's the difference between expected value and average?** An average is calculated from outcomes that have already occurred (historical data). Expected value is calculated from outcomes that haven't occurred yet, weighted by their forecast probabilities. The two converge as you accumulate enough trials - the long-run average of many independent decisions equals the expected value of each decision. **Q: Why don't most people calculate EV in real life?** Three reasons. First, real outcomes are often hard to quantify in a single unit. Second, real probabilities are often unknown - true uncertainty rather than risk. Third, intuitive decision-making is fast, and explicit EV calculations are slow. The trick is to use EV calculation for the decisions that matter most (large stakes, recurring patterns) and rely on heuristics for low-stakes everyday choices. --- # Monte Carlo Thinking: How to Stress-Test Decisions Source: https://expectedvalue.co.uk/blog/monte-carlo-thinking/ # Monte Carlo Thinking: How to Stress-Test Decisions Monte Carlo simulation lets you stress-test decisions across thousands of scenarios. A practical guide to using it for retirement, projects and investing. **Answer**Monte Carlo thinking means stress-testing a decision by mentally running it thousands of times with different plausible inputs to see the range of outcomes. Instead of asking 'what is the most likely result?' you ask 'how often does this go badly and how badly?'. Most spreadsheet apps now ship Monte Carlo add-ons that automate the simulation in minutes.
Monte Carlo thinking is the practice of stress-testing a decision by mentally - or computationally - running it across thousands of possible futures, then looking at the distribution of outcomes instead of a single point estimate. It's the most reliable counter to one of the most common mistakes in everyday reasoning: planning around the average and being blind-sided by the tail.
The technique was named after the casino in Monaco, but it isn't really a gambling tool. It's a thinking tool. Once you know what it does, you start using it informally - in retirement planning, project estimation, portfolio construction, even career choices - without needing software. This guide covers what Monte Carlo simulation actually is, when it changes a decision, and the practical mental version you can use this week.
**The core idea** Most plans assume the average outcome. Monte Carlo flips that: instead of asking 'what's the most likely result?', it asks 'across thousands of plausible futures, what is the distribution of results, and how often does this plan fail?'. The probability of failure is the number you want, not the average. ## What Monte Carlo simulation actually doesA Monte Carlo simulation takes a model of a decision - any model, simple or complex - and runs it many times, each time drawing the uncertain inputs randomly from realistic probability distributions. The output is not a single number but a distribution of numbers, from which you can read off the probability of any particular outcome.
A trivial example. You want to know whether £500,000 will last 30 years in retirement at a 4% withdrawal rate. The standard way to answer this is to assume an average annual return - say 6% - and project forward in a spreadsheet. The plan looks fine. But this misses the point: the real risk in retirement is not the average return, it's the order of returns. A bad first decade can sink a portfolio that would have survived if those same bad years had landed at the end. This is sequence-of-returns risk and it's invisible to a single-line projection.
A Monte Carlo simulation handles it directly. Instead of one 6% return per year, the simulation draws each year's return from a realistic distribution (say, a mean of 6% with a standard deviation of 16% based on historical equity data), runs the 30-year projection, and records whether the portfolio survived. Then it does it 10,000 times. The output is a single, decision-useful number: 'this plan survives 88% of the time.' Whether 88% is acceptable to you is a personal call - but at least you're now arguing about the right number.
## Where it changes a decisionThree areas where switching from average-based thinking to Monte Carlo thinking routinely changes the decision:
The classic case. A standard '4% rule' calculation assumes a smooth average return; the historical record shows that retiring into a deep bear market (1966, 2000, 2008) drops survival rates significantly even though the long-term average is unchanged. Monte Carlo or historical-sequence backtesting routinely shows that withdrawal rates which look 'safe' on paper fail in 15-25% of simulated futures. Knowing the failure rate lets you build in flexibility - guard rails, variable spending, a part-time income hedge - rather than being blind-sided by an unlucky retirement year.
Project plans almost always estimate using a single 'most likely' duration per task, then sum them. This systematically underestimates total project duration because task overruns compound and cancel less than task savings (the planning fallacy in action). A Monte Carlo run - estimate each task as a triangular distribution with optimistic / most-likely / pessimistic durations, then simulate 10,000 project rollouts - typically shows the 'most likely' total is roughly the 30th percentile of the distribution, not the 50th. The 80% confidence completion date is often weeks or months later than the simple plan suggests.
An average-return projection of a portfolio tells you what happens in the average case. Monte Carlo - or, more often in practice, historical scenario analysis - tells you what happens in the bottom 5% of cases. For long-term investors this is mostly noise; for someone with a defined liability (a house deposit in three years, university fees in seven), the tail risk is the entire point. Knowing the 5th-percentile outcome shapes how much risk you can afford to run with that money.
**The garbage-in problem** Monte Carlo simulations inherit all the assumptions of their input distributions. Use historical equity data assuming returns are normally distributed and you'll under-state tail risk - real markets have fat tails. Use a triangular distribution for project tasks with no real evidence and you're just adding numerical decoration to a guess. The discipline is in calibrating the inputs honestly; the simulation itself is the easy part. ## A worked example: the £500k, 30-year retirementTo make the technique concrete, here is what a real Monte Carlo retirement run looks like end-to-end. The set-up: £500,000 portfolio, 30-year horizon, 60/40 equity/bond split, withdrawing £20,000 a year (4% rule) increasing with 2.5% inflation. The single-line projection - a flat 6% real return - says you finish with £350,000 left over. Comfortable.
Now run it as a simulation. Each year, draw the equity return from a normal distribution with mean 7% and standard deviation 16%; bond return from a distribution with mean 2% and standard deviation 5%; inflation from a distribution with mean 2.5% and standard deviation 1.5%. Rebalance to 60/40 each year. Repeat 10,000 times. The output is not £350,000 - it's a histogram.
A typical result: the median final balance is around £700,000. The 5th percentile is zero (the portfolio runs out before year 30) - this happens in roughly 12% of simulations. The 95th percentile is £2.3m. The 'comfortable £350,000 left over' figure was the average; it tells you almost nothing about whether the plan is robust. The interesting numbers are: how often does the plan fail completely, and what does the failure look like? Once you know the failure rate is 12%, you can decide whether to accept it, reduce withdrawals, hold a cash buffer for guard-rail spending, or work an extra year.
This is also where the calibration matters. Use historical UK equity returns (mean ~5%, std ~20%) instead of US (mean ~7%, std ~16%) and the failure rate climbs to 20%+. Add fat-tailed return distributions and it climbs further. The simulation is a thinking aid for arguing about the right inputs - not a final answer.
## Mental Monte Carlo: the version that needs no softwareMost decisions don't justify a spreadsheet, let alone Python. The valuable habit is mental Monte Carlo - thinking explicitly about the distribution of outcomes rather than the most likely one. The four-step version that works for nearly any decision:
1. **Sketch the model** - What's the decision? What are the uncertain inputs? Don't be precise yet - just list them. For a job change: salary trajectory, role fit, the company's survival, your own opportunity cost of staying. 2. **Imagine three concrete futures** - A clearly bad version, a clearly good version, and the most likely. Be specific - 'the company gets acquired in 18 months and my role is restructured' is more useful than 'something bad happens'. The point is to make the tail real, not just label it. 3. **Estimate rough probabilities** - Don't try for precision. Even crude probabilities (10% / 70% / 20%) are far better than ignoring the tails. The goal is to make sure you're paying attention to the bad case, not to compute it exactly. 4. **Decide on the distribution, not the mean** - Would you take this decision if you knew with certainty the bad case would land? If yes, proceed. If no, the question becomes whether the upside compensates for the downside often enough - which is now a calibrated bet rather than a hope.Mental Monte Carlo is closer to the cognitive habit of pre-mortem thinking than to the formal simulation, but they share the same underlying logic: take the distribution seriously, not the average. It also pairs well with decision journals - writing down your imagined scenarios is the easiest way to stop your brain from quietly deleting the bad ones once you've made the call.
## When Monte Carlo isn't worth the effortThree situations where the formal technique adds nothing - or actively misleads:
If the underlying probability distributions are themselves wild guesses, the simulation will produce confident-looking outputs that are no better than the inputs. The veneer of statistical precision can make speculative numbers feel rigorous. This is especially common in startup financial projections and 10-year strategic plans.
If 90% of the variance comes from one input (will the regulator approve the deal? yes or no?), the simulation just rephrases the question. Decision-tree analysis is usually more honest in those cases - you're explicitly conditioning on the dominant binary outcome.
Monte Carlo gives you the average case across many futures, weighted by probability. But for genuinely catastrophic tail risks - Taleb's domain - the right question isn't 'what's the probability-weighted outcome?' but 'how do I avoid the catastrophic tail entirely, regardless of how unlikely it is?'. Probability and expected value calculations stop being decision-relevant once a possible outcome is ruinous.
## Tools and where to startIf you want to try a real simulation, the bar is lower than you'd think. Three options ordered by how quickly you can be running:
RAND() function and the NORM.INV() function will get you a basic Monte Carlo retirement simulation in under an hour. Acceptable for one-off analysis; impractically slow if you want 10,000 runs.Whichever tool you pick, the value is in looking at the distribution - the histogram, the percentile bands, the failure rate - not just the average. If your output is a single number, you've thrown away the point of the technique.
## Frequently asked questions **Q: What is Monte Carlo simulation in plain English?** It's a way of testing a plan by running it many times with the uncertain inputs varied randomly each time, then looking at the spread of results. Instead of asking 'what's the average outcome?', it asks 'how often does the plan succeed across thousands of plausible futures?'. **Q: Why is it called Monte Carlo?** The technique was developed by physicists working on nuclear weapons at Los Alamos in the 1940s. They named it after the famous casino in Monaco - the connection being that both rely on repeated random sampling. The mathematician Stanislaw Ulam came up with the approach while playing solitaire and wondering about the probability of winning. **Q: How is it different from a sensitivity analysis?** A sensitivity analysis varies one input at a time, holding the others constant. Monte Carlo varies them all simultaneously, with each input drawn from its own probability distribution. This captures interactions and joint extreme outcomes that a one-at-a-time analysis misses entirely. **Q: How many simulations do I need to run?** For most personal-finance and project-estimation use cases, 10,000 runs is enough - the percentile estimates are usually stable to within a percentage point or two. For tail-risk estimation (the 1st or 99th percentile), 100,000+ runs may be needed to get a stable estimate. **Q: Is Monte Carlo simulation reliable for retirement planning?** It's reliable as a directional check on whether your plan has serious sequence-of-returns risk - which a simple average-return spreadsheet completely hides. It's not reliable as a precise probability of success, because the future return distribution is unknown and the simulation outputs a confidence-looking number that depends entirely on your assumed inputs. Use it to size up the tail risk, not as a guarantee. --- # Expected Value vs Expected Utility: When EV Isn't Enough Source: https://expectedvalue.co.uk/blog/expected-value-vs-expected-utility/ # Expected Value vs Expected Utility: When EV Isn't Enough Pure expected value can lead to ruinous decisions. Here's why expected utility, risk aversion, and the St Petersburg paradox matter for real-life choices. **Answer**Expected value is the average monetary outcome of a decision, weighting each possible result by its probability. Expected utility adjusts for the fact that each pound of additional gain is worth slightly less than the previous one - you would not bet your whole pension on a fair coin flip even though the expected value is zero. Use EV for small repeatable decisions, expected utility for big irreversible ones.
Expected value is the workhorse of probabilistic thinking - multiply outcomes by their probabilities, sum them up, pick the highest number. It's a brilliant rule. It's also, sometimes, a terrible one. Take the highest-EV bet often enough and you can go broke before the long run shows up. That gap between "mathematically correct" and "actually wise" is where expected utility lives.
This post is a practical walk through when raw expected value works, when it falls apart, and how the concept of utility - the idea that a pound is worth more to you when you have ten of them than when you have a million - fixes it. We'll cover risk aversion, the St Petersburg paradox, and the Kelly criterion, with the maths kept light and the examples kept real.
**Quick recap before we start** If you're not already comfortable with expected value as a concept, read our explainer at /blog/expected-value-explained/ first. The rest of this post assumes you can compute an EV and understand why it's the right number for repeated identical bets. ## Where does pure expected value actually work?Expected value works perfectly when three conditions are all true:
A poker player computing pot odds inside a single hand has all three. A sports trader placing one of a hundred small positions a week has all three. An insurance company writing one of a million tiny policies has all three. In those settings, EV is the right tool and it's not even close.
It's the corner cases that break the model - and unfortunately for ordinary humans, our biggest decisions are the corner cases.
## What is the St Petersburg paradox?Daniel Bernoulli, 1738. Imagine a fair coin flipped until it comes up tails. If tails appears on flip 1 you win £2; on flip 2, £4; on flip 3, £8; on flip 4, £16; and so on. The pot doubles every time the game continues. How much would you pay to play?
The expected value is:
EV = ½ × £2 + ¼ × £4 + ⅛ × £8 + 1/16 × £16 + …
= £1 + £1 + £1 + £1 + … = ∞
The expected value is infinite. By pure EV reasoning, you should pay any finite amount - £100, £10,000, your house - for one go. And yet ask any normal human and they'd offer about £10, maybe £20 if they're feeling cheerful. That gap between "infinite" and "twenty quid" is the puzzle.
Bernoulli's solution was to introduce utility - the idea that the subjective value of money diminishes as you have more of it. Winning £1,048,576 isn't literally a million times more useful to you than winning £1, even if a calculator says it is. Once you redo the maths using log-utility instead of raw money, the expected utility of the game becomes finite and small, which matches what real people are willing to pay.
## What is diminishing marginal utility, in plain English?Strip the academic phrase away and the idea is intuitive. Imagine you have £0:
Each additional pound is worth less than the one before it because you've already spent your previous pounds on whatever was most valuable. That's diminishing marginal utility. The economic curve looks like log(x) or square-root(x) - fast at first, flattening out.
The flip side, which matters even more, is the asymmetry of losses:
The curve isn't just sub-linear above zero - it's brutally steep below the survival threshold. This is where pure EV reasoning kills people: it can't see the cliff.
## Is risk aversion irrational?You're offered two choices:
Both have the same expected value: £50,000. Most people pick A. By pure EV, they're "leaving money on the table" because A and B are equivalent. By utility theory, they're being entirely rational - the certainty of a life-changing sum is worth far more than the same EV smeared over a coin flip.
This is why insurance exists. The expected value of buying insurance is negative - that's how the insurance company stays in business. But for the customer, paying a small certain loss to avoid a small chance of a catastrophic loss is a positive-utility trade. It's not stupidity, it's correct play. The pure-EV person who self-insures their house against fire is the one being irrational, not the homeowner with a policy.
The general principle: variance is a cost when the downside is catastrophic relative to your bankroll, and the cost shows up in utility, not EV.
## How does the Kelly criterion help you survive?If you've ever heard a poker player or a trader talk about "sizing" bets, they probably weren't using raw EV. They were using something close to the Kelly criterion - a 1956 result by John Kelly that gives the bet size that maximises the long-run growth rate of your bankroll, assuming you can keep betting.
The formula in its simplest form:
f* = (bp − q) / b
Where f* is the fraction of your bankroll to bet, p is your probability of winning, q is the probability of losing (1 − p), and b is the net odds you receive. The point isn't the formula - it's the punchline:
Kelly is the practical bridge between EV and EU: it's a sizing rule that respects the geometric reality of compounding wealth, not the arithmetic illusion that linear EV implies.
## When should you use EV vs expected utility?A practical decision tree for everyday use:
EV is negative - you pay £400/year and the actuarial expectation of payouts is maybe £350. Utility is positive - you're trading £400 of certain comfort for protection against a £200,000 wipeout. Buy the insurance. Don't fall for the "insurance is a scam because EV is negative" argument; it's a scam-shaped solution to a real problem.
EV is wildly positive. Utility is awful. A 30% chance of total ruin is not paid for by a 70% chance of doubling your money - because doubling £50,000 doesn't change your life as much as losing £50,000 ruins it. Don't do this even if the maths "works."
EV (lifetime expected income) might favour the higher-paying-but-fragile startup over the stable corporate role. Utility might not - if the startup tanking means a year of unemployment with a mortgage, the asymmetry of outcomes matters. Our probabilistic framework for career decisions goes deeper on this.
If the most you can lose is £2,000 and a year of evenings, and the upside is real, Kelly sizing says: size it small relative to your wealth, but go. The bet is bounded. Asymmetric upside, capped downside, repeatable in form (you can make many similar bets over a career).
## What are the most common mistakes mixing EV and utility?Three traps that show up constantly:
Expected value is the right tool when bets are small, repeatable, and recoverable. Expected utility is the right tool when any of those conditions break - which is the case for most of the decisions that actually matter in your life.
The single most useful upgrade to your decision-making isn't more sophisticated maths; it's noticing which regime you're in. "Is this the kind of decision where I'll get to converge to the long run, or the kind where one bad outcome ends the game?" That question, asked early, prevents almost every catastrophic-but-mathematically-justifiable mistake.
For more on related ideas, see our explainers on expected value, risk vs uncertainty, and why your brain is bad at risk.
## Frequently asked questions **Q: What's the difference between expected value and expected utility?** Expected value treats every pound as equally valuable - it's the probability-weighted sum of outcomes in pounds. Expected utility weights outcomes by their subjective value (utility), recognising that a pound is worth more when you have less, and that catastrophic losses hurt disproportionately more than equivalent gains help. EV is right for small repeatable bets; EU is right for large or one-off ones. **Q: Why is expected utility theory better than just using EV?** Because real money has diminishing marginal utility - the millionth pound matters less than the first - and real losses can be uncrossable (bankruptcy, eviction). EV calculations are blind to those asymmetries; expected utility builds them in. The St Petersburg paradox is the classic demonstration: pure EV says pay any amount to play, EU says pay roughly £20. **Q: Is risk aversion irrational?** No - it's correct decision-making once you account for diminishing utility and survival constraints. Refusing a coin-flip for £100k vs £50k guaranteed isn't leaving money on the table; it's correctly recognising that £50k of certainty is worth more in utility than £50k of expected variance. **Q: What is the Kelly criterion in simple terms?** Kelly is a formula for sizing bets to maximise long-run growth of your bankroll, assuming you can keep betting. It tells you what fraction of your money to put on a +EV bet so that variance doesn't ruin you before the EV materialises. Most professionals use a fraction of full Kelly (quarter or half) as a safety margin. **Q: When should I ignore expected value?** When the bet is one-off, when the downside includes ruin or any irreversible outcome, when the stakes are large relative to your wealth, or when you can't realistically estimate the probabilities. In any of those cases, raw EV gives you the wrong answer; expected utility gives the right one. **Q: Is buying insurance irrational because EV is negative?** No. The insurance company has positive EV (that's how they stay solvent), but you're not playing the same game. You're trading a small certain cost against a small probability of a catastrophic loss. In utility terms, that's a positive trade - protecting yourself from cliffs is worth paying a premium for. --- # Hindsight Bias: Why Everything Looks Obvious After the Fact Source: https://expectedvalue.co.uk/blog/hindsight-bias/ # Hindsight Bias: Why Everything Looks Obvious After the Fact Hindsight bias makes the past feel inevitable. Here's why it distorts post-mortems, juries and investing - and how to fight it. **Answer** Hindsight bias is the tendency to see past events as having been more predictable than they were. Once an outcome is known, we reconstruct our memory of pre-outcome beliefs to make the outcome feel inevitable. The corrective is to write down probability estimates in advance and check them against actual outcomes - the only reliable way to detect this bias in yourself.On 14 September 2008, the Sunday papers were full of speculation about Lehman Brothers. By Monday morning the bank was bankrupt and global markets were in freefall. Within weeks, the financial press had constructed an elegant narrative explaining exactly why this was always going to happen: subprime mortgages, opaque CDOs, leverage ratios that any sensible person should have spotted, a regulatory regime asleep at the wheel.
It was, the commentariat agreed, obvious in retrospect.
Which is interesting, because the same commentariat had spent the previous three years not finding it obvious at all. In 2007, the IMF described global financial stability risks as 'manageable'. The Federal Reserve was raising rates in mid-2006 because it was worried about inflation, not insolvency. Of the thousands of paid analysts covering US banks, you can count on one hand the ones who publicly forecast a Lehman-style collapse.
This is hindsight bias - the systematic tendency to see past events as more predictable than they actually were. It is one of the most pervasive distortions in human cognition, and it quietly poisons everything from investment post-mortems to jury verdicts to your own performance reviews. Once you learn to spot it, you start seeing it everywhere - including, embarrassingly, in your own thinking.
If you've read Thinking in Probabilities, you already know that humans are terrible at calibrating uncertainty. Hindsight bias is the bias that retroactively destroys the evidence of our miscalibration. It is, in a sense, the bias that hides the other biases.
## What is the 'I knew it all along' effect?Hindsight bias was first formalised by Baruch Fischhoff in a now-classic 1975 paper, 'Hindsight is not equal to foresight.' The setup was elegant. Fischhoff gave participants descriptions of historical events whose outcomes were genuinely uncertain at the time - including a then-obscure 19th-century war between the British and the Gurkhas of Nepal. He asked one group to estimate the probability of various outcomes (British victory, Gurkha victory, military stalemate, political settlement). A second group was told which outcome had actually occurred and asked to estimate the probability they would have assigned beforehand.
The second group consistently inflated their pre-event probability estimates for whatever outcome actually happened. Once they knew the British had won, the British victory suddenly seemed like the predictable result. They were not lying - they genuinely believed they would have predicted it. Their memory of their own uncertainty had been quietly rewritten.
Fischhoff identified three distinct components of the bias:
A closely related concept, creeping determinism, captures the third effect particularly well: as soon as we learn how a story ended, our brains start working backwards, weaving causal chains that make the ending feel inevitable. The financial crisis didn't just happen - it had to happen, given subprime lending and Greenspan and the housing bubble and... and... and...
The trouble is, this same machinery would have constructed an equally elegant narrative if Lehman had been bailed out and the crisis averted. 'Of course they intervened - the systemic risk was too obvious.' Hindsight bias is a story-telling engine that fits the facts to the outcome, not the other way around.
**Fischhoff's Razor**Fischhoff's central insight: knowledge of an outcome cannot be unlearned. Once you know what happened, you cannot honestly estimate what you would have predicted, because the prediction has already been contaminated by the answer. This makes most retrospective self-assessment fundamentally unreliable.
## Why are most investment post-mortems just theatre?Open any post-2008 retrospective and you will find a confident explanation of why the crisis was 'inevitable.' Open any post-2020 retrospective and you will find an equally confident explanation of why the COVID crash and recovery were inevitable too. Open any post-2022 retrospective and the inflation surge is, naturally, also inevitable.
The word 'inevitable' is doing a lot of work here. If everything is inevitable in hindsight, then by symmetry, nothing should be surprising in advance. And yet markets remain capable of generating sustained surprise. The contradiction is hindsight bias, and it has three particularly nasty consequences for investors.
1. Bad decisions get rebranded as bad luck, and good luck gets rebranded as skill.
A fund manager who concentrated heavily into Nvidia in 2022 looks like a genius today. But if the AI thesis hadn't paid off, the same position would look like irresponsible concentration risk. The decision was the same. The outcome was different. Hindsight bias treats these as different decisions, judging the manager on the outcome rather than the quality of the reasoning at the time. As Annie Duke argues in Thinking in Bets, this is resulting - judging decisions by their outcomes - and it is the fastest way to learn the wrong lessons from your investing history.
2. 'Obvious' market crashes were not obvious in advance.
The dot-com bust of 2000 is now treated as a no-brainer call. But in 1999, every reputable investment bank had double-digit price targets on companies with no earnings. Warren Buffett, who did avoid the bubble, was being publicly mocked as a senile dinosaur who 'didn't get it.' Berkshire Hathaway's stock dropped 44% from June 1998 to March 2000 while the Nasdaq doubled. If you'd put Buffett's reasoning to a vote in 1999, he would have lost.
The lesson isn't that crashes are unpredictable. The lesson is that predictability is asymmetric: easy after, hard before. Anyone who tells you they 'called' a crash should be asked to produce time-stamped, public, falsifiable predictions made before the event, with a track record of equally specific predictions that didn't come true. Almost nobody can.
3. We learn the wrong lessons from our own portfolios.
Review your trades from five years ago. Hindsight bias will whisper that the winners were obvious - you knew that one was a bargain - and the losers were aberrations you should have spotted. But the losers were chosen using exactly the same process as the winners. If your process produced a 60% hit rate at the time, that's the rate you should expect going forward, regardless of how each individual decision feels in retrospect.
This is why a good decision journal - written before outcomes are known - is worth more than a hundred post-mortems.
## How does hindsight bias warp courtroom decisions?If hindsight bias only distorted finance, it would still be expensive. But the same bias warps decisions where lives and liberty are at stake.
Medical malpractice trials. Consider a patient who presents with chest pain at 3am. The on-call doctor weighs the symptoms - patient is 34, no family history, mildly elevated heart rate - and concludes a panic attack is more likely than cardiac event. The patient is sent home and dies of a heart attack at 6am.
In court, jurors are shown the autopsy. They are shown the patient's family, devastated. And they are asked: should the doctor have foreseen this?
Research by Kim Kamin and Jeffrey Rachlinski has shown that jurors who know an outcome rate the prior probability of that outcome dramatically higher than jurors who must judge the same case prospectively. In one study, a foreseeability finding rose from 33% in a foresight condition to 57% in a hindsight condition - same facts, different framing. Doctors are not being judged on the quality of their decision at 3am. They are being judged on whether, knowing what happened at 6am, the 3am decision should have led somewhere different.
This is functionally an impossible standard. A doctor who treated every chest pain in a 34-year-old as cardiac would order tens of thousands of unnecessary tests, cause massive anxiety, and miss the genuine cardiac cases hidden among the noise. The base rate matters, as we explore in Base Rate Neglect. But hindsight bias makes base rates feel irrelevant - this case turned out to be the cardiac case, and so retrospectively the base rate was 100%.
Jury decisions on negligence and intent. The same effect distorts product liability cases, accident reconstructions, and corporate negligence claims. Once a jury knows that a particular component failed and caused a death, the prior probability of failure feels obvious. Engineers who made reasonable design tradeoffs against an entire distribution of possible failure modes are judged as if they should have foreseen this specific failure mode.
Historical analysis. Open a popular history of any decade. The narrative runs: cause, cause, cause, event. The narrative never reads: cause, cause, cause, something completely different that almost happened instead. Counterfactuals get edited out of history because they don't fit the determinist arc that hindsight bias demands. The result is a pseudo-deterministic view of the past that makes the future look more predictable than it actually is.
This matters for investors because most market commentary is a kind of pop history. Read enough explanations of why the 1970s inflation was 'inevitable' and you start believing the 2020s inflation must have been equally foreseeable. It wasn't. The fact that it happened doesn't mean someone, somewhere, should have called it.
**The Symmetry Test**The standard correction: when judging a past decision, ask whether you would judge it the same way if the outcome had been different. If you'd let the doctor off had the patient survived, but condemn them because the patient died, you are judging the outcome - not the decision. That's hindsight bias talking.
## What is happening in your brain when hindsight kicks in?Why does hindsight bias happen? Three mechanisms, each documented in dozens of replicated studies.
1. Memory updating. Memory is not a recording. It is a reconstruction, performed afresh each time you remember something, and contaminated by everything you've learned since. When you 'remember' your prior estimate, your brain regenerates a plausible value - and what's plausible now includes the actual outcome. The old estimate is overwritten in much the same way you might overwrite a file.
2. Sense-making. The human brain is a relentless pattern-matching machine. Given any outcome, it will rapidly construct a causal explanation - and once that explanation exists, the outcome feels like its natural conclusion. This is creeping determinism. The story makes the ending feel inevitable; the inevitability makes the story feel correct.
3. Anchoring. Once you know the answer, that answer becomes a numerical anchor. When asked to estimate what you 'would have' predicted, your estimate drifts toward the anchor. The well-documented anchoring effect makes the bias almost impossible to overcome by introspection alone - you cannot wish your way out of it.
The combined effect is robust across cultures, ages, expertise levels, and incentive structures. Telling people about hindsight bias before showing them an outcome reduces the effect modestly. Telling them about it afterwards barely helps. Even researchers who have spent their careers studying the bias exhibit it on themselves. There is no cognitive setting that switches it off.
This is bad news, but also liberating. If you cannot fix hindsight bias from the inside, you must fix it from the outside - by changing your process, not your effort. The same is true of its close cousin, confirmation bias: awareness alone does almost nothing, but a few process changes do most of the work.
## What actually works against hindsight bias?The single most powerful tool for fighting hindsight bias is the decision journal. Before making any significant decision - investing, hiring, strategic - you write down:
This is then sealed (literally or in software) and reviewed only when the outcome is known. Suddenly you have a frozen record of what you actually thought, immune to retrospective revision. Fund managers like Mohnish Pabrai and Guy Spier swear by this technique, and it is widely adopted in superforecasting communities.
The revelation, when you start, is uncomfortable: your prior beliefs were almost always vaguer, more uncertain, and more wrong than you remember. That discomfort is the bias being measured.
Gary Klein's pre-mortem technique is the inverse of a post-mortem. Before a project launches, you imagine that it has failed catastrophically twelve months from now, and you brainstorm the reasons why. By forcing your brain to construct a future-failure narrative before the outcome is known, you disrupt the determinist arc that hindsight bias normally builds.
Amazon, the US military, and many venture capital firms use pre-mortems before major commitments. They consistently surface risks that ordinary risk analysis misses - because ordinary risk analysis tends to weigh the risks that 'feel' relevant, and pre-mortems force you to make every plausible failure mode feel relevant.
You cannot improve at probabilistic prediction without feedback. Calibration training - making numerical predictions ('70% confident this stock outperforms over 12 months') and tracking your hit rate - gives you the feedback loop that ordinary life withholds.
The Good Judgment Project showed that calibration training, combined with structured prediction, can produce superforecasters who outperform intelligence analysts. Brier scores and log-likelihood scoring rules let you track whether your '70% confidence' predictions actually come true 70% of the time. They almost never do at the start. They get closer with practice.
Bayesian thinking is the natural framework for this kind of incremental updating.
When reviewing any past decision, start with the base rate - not the outcome. 'Of all the merger announcements made in 2019, what proportion delivered shareholder value over five years?' That base rate is your prior. The specific case shifts your estimate slightly, but not as much as the outcome alone makes it feel like it should.
This is the lesson of Base Rate Neglect: outcomes are samples from a distribution, not the distribution itself. When you treat them as the distribution, you commit hindsight bias and base rate neglect simultaneously.
A simple but uncomfortable exercise: take any outcome you consider 'inevitable' and write a serious, plausible alternative history in which the opposite happened. Lehman is bailed out and the GFC is averted. The dotcom bubble keeps inflating for another five years. England wins the 2022 World Cup. If you cannot write the counterfactual without feeling stupid, hindsight bias is operating. The counterfactuals were once possible. They are not stupid; they are what was being priced into options markets at the time.
Good historians and good investors share this skill: they hold open the futures that didn't happen, because those were the futures that made the actual future uncertain in the first place.
## What should you read to push back against hindsight bias?If hindsight bias has caught your interest - or made you uncomfortable enough to want to do something about it - these are the indispensable books. UK readers can support independent bookshops via Bookshop.org rather than buying everything from the obvious online giant.
These complement our own reading list of best books on probabilistic thinking, which covers a wider sweep of the literature.
You will not finish these and stop being susceptible to hindsight bias. Nobody does. But you will become quicker to notice when you - or someone else - is constructing a story that makes the past feel inevitable. That noticing is the entire game. The world rewards people who can hold open the futures that haven't yet happened, and resist the urge to retroactively close the futures that already have.
**The Practical Takeaway**The hindsight-resistant investor's checklist. Before judging any past decision: (1) write down the base rate, (2) imagine the s …[truncated for length] --- # Ergodicity Explained: Why Time Averages Matter Most Source: https://expectedvalue.co.uk/blog/ergodicity-explained/ # Ergodicity Explained: Why Time Averages Matter Most Ergodicity separates sensible bets from catastrophic ones. The difference between ensemble averages and time averages - and why ignoring it can ruin you. **Answer**
Ergodicity is a property of a system where the time average (one person's path through many time steps) equals the ensemble average (many people's outcomes at one time step). Coin flips are ergodic: expected value applies to both you over time and to a population at one moment. Wealth is not ergodic: a 50% gain followed by a 50% loss leaves you 25% worse off (1.5 × 0.5 = 0.75). Practical implication: ensemble-average expected value misleads in non-ergodic contexts; long-run individual growth requires geometric mean, log returns, and the Kelly criterion.
Most decisions are framed in terms of expected value: average out the possible outcomes, weighted by their probabilities, and pick the option with the highest number. It's clean. It's mathematical. And for a surprisingly large class of real-world decisions, it's catastrophically wrong.
The hidden assumption inside expected-value reasoning is ergodicity - the idea that the average outcome across many people equals the average outcome for one person across time. When that assumption holds, EV calculations work as advertised. When it doesn't - and in finance, health, careers, and most of the decisions that actually matter, it doesn't - the gap between the two can swallow you whole.
This is the concept that economists Ole Peters and Murray Gell-Mann argue has been quietly broken at the heart of modern economics for sixty years. It's also the most important idea you've probably never heard of. Once you understand it, you start to see ergodicity violations everywhere - and you make different choices.
## What ergodicity actually meansA process is ergodic if the average outcome across many parallel realisations of it (the ensemble average) equals the average outcome for one realisation followed across time (the time average). Coin flips are ergodic: flip 1,000 fair coins once each, or flip one fair coin 1,000 times - both procedures converge on a 50% heads rate.
A process is non-ergodic if those two averages diverge. Most things you actually care about - your wealth over a lifetime, your health, your career, the stock market, the gambler's bankroll - are non-ergodic. The ensemble average tells you nothing useful about what happens to you.
The plain-English version: ergodicity asks whether "average across people" tells you anything about "average across time". Sometimes yes. Often no. The decisions where the answer is "no" are the decisions where most people get hurt.
## The coin-flip game that breaks expected valueHere's a thought experiment from Ole Peters that demonstrates the problem cleanly. Imagine a game where you flip a fair coin repeatedly. The rules:
The ensemble average is straightforward. Across many parallel players, the expected outcome of one round is 0.5 × 1.5 + 0.5 × 0.6 = 1.05. A 5% expected gain per round. Repeat this game forever and the average wealth across all players grows exponentially. By the maths of expected value, this is a brilliant game and you should bet your whole net worth every round.
Now follow one player across time. After a heads-tails sequence: 1.0 × 1.5 × 0.6 = 0.9. A 10% loss every two rounds. The time-averaged growth rate is negative. Almost every player goes broke. The handful who don't drag the ensemble average up because the few that get lucky get astronomically rich, but the typical player - and almost certainly you, if you played - loses everything.
Both calculations are correct. They describe different things. Expected value answers "what's the average outcome across all possible worlds?". Time average answers "what's the typical outcome for one person playing repeatedly?". For repeated decisions in your own life, only the second number matters. The first is a statistical artefact about parallel universes you don't get to live in.
**The intuition pump** If 1,000 people each play this game once, the average wealth across the group ends up higher. If one person plays it 1,000 times, they almost certainly end up with nothing. Ergodicity asks: which of those two scenarios are you actually in? Almost always, it's the second. ## Why your intuition gets this wrongThe intuition that expected value is the right way to think comes from environments where ergodicity actually holds. Most casino games, when played in tiny stakes relative to your bankroll, are roughly ergodic - your bankroll fluctuates around a slow drift. The school maths examples, where probabilities map neatly onto frequency counts, are ergodic. The Bayesian-update problems in undergraduate textbooks are ergodic.
But the moment you scale stakes relative to your bankroll, or introduce path-dependent dynamics, ergodicity breaks. Wealth multiplies rather than adds. Losses can't be recovered by symmetric gains (a 50% loss requires a 100% gain to get back to flat - losses and gains are not commutative under multiplication). One bad outcome forecloses all future outcomes - "ruin" is absorbing. Time matters in ways the ensemble average literally cannot see.
This is why people who reason from textbook expected value are systematically wrong about real-world risk-taking. They take bets that look great on paper, get unlucky once, and discover they've optimised for an average that doesn't apply to their actual life.
## Investing: where ergodicity violations liveThe single most important application of ergodicity is investing. Returns compound multiplicatively, which means past losses constrain future gains. A portfolio that drops 50% needs a 100% gain to break even. Drop 90% and you need a 900% gain. The expected return on a stock can be positive while the median investor experiences a loss - because the asymmetric path crushes most realised trajectories.
Two practical implications:
Diversification isn't optional. The reason index funds work isn't that they have higher expected returns than concentrated portfolios - they often have lower expected returns. They work because they have higher time-averaged returns. Concentrated portfolios have higher variance, and variance corrodes geometric returns. The ensemble average rewards concentration; the time average punishes it.
Risk of ruin trumps expected value. Any positive expected value bet that has a non-negligible probability of wiping you out is a bad bet, no matter how high the EV looks on paper. The Kelly Criterion formalises this: bet to maximise log-wealth, not expected wealth. The log transformation is the trick that converts the non-ergodic geometric process into a quasi-ergodic additive one. Bet less than Kelly says and you grow more slowly; bet more and you go bust.
The financial industry tacitly understands this. "Risk-adjusted returns", "Sharpe ratio", "drawdown control" are all imperfect attempts to capture time-average behaviour while preserving the expected-value vocabulary that the textbooks insist on. Ergodicity gives you a cleaner conceptual handle: ask whether the process is ergodic, and if it isn't, optimise the time average directly.
## Career, health, and life decisionsThe same logic applies to most consequential life decisions. Some examples:
Career risk-taking. Quitting a stable job to launch a startup has a positive ensemble expected value - across all founders, the average outcome is positive once you account for the rare lottery winners. But the time-averaged outcome for any individual founder is much worse, because most founders fail and "having failed once" tends to cascade into reduced future risk-taking, depleted savings, and lost compounding years. The advice "be willing to take asymmetric risks" needs the qualifier: as long as the downside is survivable. A bet that ends your ability to take future bets isn't asymmetric in the right direction.
Health and safety. The same logic explains why we should accept what looks like a poor expected-value trade-off when the downside is irreversible. Wear a seatbelt. Don't free-climb. Get the cancer screening. The expected value of a single skipped seatbelt is positive (small comfort gain, tiny chance of catastrophic loss); the time-averaged value across a lifetime of skipping is sharply negative.
Reputation and trust. Reputation is built linearly and destroyed multiplicatively. A relationship survives many minor disagreements but rarely survives one serious betrayal. Your professional reputation accumulates slowly through dozens of small wins and can be undone by a single public failure. The ensemble math says "on average, the small slip is fine". The time math says "the slip you don't recover from is the only one that matters".
Once you start looking, you find non-ergodic dynamics in nearly every domain where outcomes accumulate over time and bad outcomes are absorbing or hard to reverse.
## Ole Peters and ergodicity economicsOle Peters, a physicist at the London Mathematical Laboratory, has spent over a decade arguing that mainstream economics has built itself on an unexamined ergodicity assumption - and that the assumption is wrong for the situations economics most cares about. His collaboration with Nobel laureate Murray Gell-Mann produced a series of papers reframing classical results in economics through the lens of the time average rather than the ensemble average.
The core claims of ergodicity economics are roughly:
The ideas remain controversial - mainstream economics has not adopted them - but the arguments are technically clean and worth understanding even if you don't accept the full programme. At minimum, ergodicity economics gives you a vocabulary for a problem that the standard frameworks struggle to articulate.
## Practical decision rulesYou don't need to compute log-utility integrals to use ergodicity in everyday decisions. A handful of heuristics capture most of the value:
Ask: "Am I one player playing many times, or many players playing once?" If the latter, expected value works. If the former - almost always, in your own life - you need the time average. Most personal financial decisions are repeated bets by one player.
Test for absorbing states. Could this outcome end my ability to make future decisions? Bankruptcy, severe injury, criminal conviction, broken marriage - these are absorbing. They cap your future trajectory. Avoid them disproportionately to their EV-weighted impact, because the EV calculation systematically understates how much they cost you.
Optimise log-wealth, not wealth. When sizing bets - investments, leverage, career risks - think in log terms. A 10% gain and a 10% loss feel symmetric in log-space (they're not in absolute terms - a 10% gain followed by a 10% loss leaves you below your starting point, a basic geometric mean point). Sizing bets to maximise expected log-growth automatically respects ergodicity.
Diversify aggressively across non-correlated risks. Diversification is the cheapest known method of converting non-ergodic processes into approximately ergodic ones. The more uncorrelated your bets, the more the ensemble average and time average converge, and the more standard EV reasoning starts to apply.
Insure the catastrophes. Insurance has negative expected value (the insurer is profitable, after all). It can have positive time-averaged value, because it removes paths that absorb you. The same logic applies to emergency funds, conservative leverage, and "barbell" portfolio strategies.
**The one-line summary** Maximising expected value is the right strategy if you get to live every possible outcome, weighted by probability. You don't. You get to live one. Optimise for the typical path, not the parallel-universe average - and especially avoid paths that delete the parameter you're trying to optimise. ## Where ergodicity meets other conceptsErgodicity sits at a particular crossroads in the probabilistic toolkit. It's worth seeing the connections:
Risk vs uncertainty: ergodicity is mostly about risk (well-defined probability distributions). Knightian uncertainty layers on top - when you don't even know the distribution, time-average reasoning is even more important than under risk, because the variance of your estimate of the average is itself uncertain.
Second-order thinking: ergodicity is a form of second-order thinking. The first-order question is "what's the EV?". The second-order question is "is the EV the right number to be optimising at all?".
Loss aversion: classical loss aversion is treated as a behavioural bias - humans irrationally weight losses more heavily than gains. Through the ergodicity lens, much of loss aversion looks like correct behaviour: people optimising the time average rather than the ensemble average, which is what they should be doing for repeated personal decisions.
Kelly Criterion sizing: Kelly betting is the operational answer to the question ergodicity raises. "How big should I bet?" becomes "how big a bet maximises my time-averaged growth rate?". Kelly is the answer for the simple binary case; log-utility-optimisation is the answer in general.
## Frequently asked questions **Q: Is ergodicity the same as the law of large numbers?** Not quite. The law of large numbers says that as you take many samples from a distribution, the sample mean converges to the population mean. Ergodicity is a stronger statement: it says the time average of one realisation equals the ensemble average across realisations. For ergodic processes, time-sampling and ensemble-sampling are interchangeable. For non-ergodic ones, they're not - and the law of large numbers only tells you about the ensemble side. **Q: Are stock market returns ergodic?** Not over the timescales most investors care about. Real returns compound multiplicatively, which is non-ergodic. The ensemble average return (across all possible market histories) is higher than the time-averaged return for any individual investor, because variance corrodes geometric growth. This is why diversified, low-cost index investing - which dampens variance - outperforms concentrated stockpicking on average over long horizons. **Q: Does ergodicity disprove expected utility theory?** Ergodicity economics argues that expected-utility theory uses the wrong optimisation target for non-ergodic processes. Mainstream economics disagrees, treating the issue as a question of how to define utility functions. Both sides agree, in practice, that you should optimise log-wealth or something close to it for repeated wealth bets - they just disagree about the philosophical reasoning. For practical decisions, the Kelly / log-wealth answer is robust to which interpretation you prefer. **Q: Where can I read more?** Ole Peters' Ergodicity Economics lecture notes (publicly available online) are the original technical source. For accessible introductions, Nassim Nicholas Taleb's writing on ergodicity in Skin in the Game is widely cited; Luca Dellanna's Ergodicity book offers a concise practical guide. James Stock's macroeconomics textbook has a chapter that introduces the ensemble vs time-average distinction in the context of growth dynamics. **Q: Is everything in life non-ergodic?** No. Genuinely repeatable, well-mixing processes - fair coin flips at small stakes, dice rolls, the orientations of gas molecules in a room - are ergodic and EV reasoning works fine. The trouble is that most of the decisions that actually matter - wealth, health, careers, relationships - happen on path-dependent, multiplicative timelines where the ensemble average diverges sharply from what you actually experience. The skill is recognising which class a decision falls into before applying the wrong tool.If you take one thing from this: when a decision is repeated and the outcomes accumulate over time, optimise the time average, not the ensemble average. When the downside is absorbing, weight it more than its EV-share suggests. And when you can convert a non-ergodic process into a more ergodic one - through diversification, position sizing, or insurance - the conversion is almost always worth more than the apparent expected-value cost of doing so.
The standard expected-value framework is a brilliant tool for decisions that look like coin tosses at the school playground. For the decisions that actually shape your life, it needs the ergodicity correction.
--- # Conditional Probability Explained Simply, With Examples Source: https://expectedvalue.co.uk/blog/conditional-probability-explained/ # Conditional Probability Explained Simply, With Examples Conditional probability is the chance one event happens given another already has. Worked examples in medical testing, weather, and the Monty Hall problem. **Answer**Conditional probability is the probability of event A given that event B has occurred, written P(A|B). It captures how new information updates beliefs. Example: P(disease) = 1%; P(positive_test | disease) = 99%; P(positive_test | no_disease) = 5%. After a positive test, P(disease | positive_test) is not 99% - it's about 17% (by Bayes' theorem), because the rarity of the disease drags the posterior down. This is where most medical-test intuition fails: the base rate matters as much as the test's accuracy.
Conditional probability is the chance that one event happens given that another event has already happened. Written P(A|B) and read "the probability of A given B", it's the single most useful idea in probability for everyday decision-making - and the source of nearly every famous probability puzzle that looks impossible at first glance. This guide builds it up from scratch with worked examples, then shows how it underpins Bayes' theorem, the Monty Hall problem, medical-test interpretation, and weather forecasting.
The phrase conditional probability sounds intimidating but the idea is everyday. Imagine you're told a card has been drawn from a standard 52-card deck. Without any other information, the chance it's the Ace of Spades is 1 in 52. Now suppose someone tells you the card is black. The relevant universe just shrunk from 52 cards to 26, and the chance it's the Ace of Spades is now 1 in 26. The probability changed because new information narrowed what's possible.
That's conditional probability in one sentence: it's the probability of an event in a smaller world, where the smaller world is defined by something you've just learned. Mathematically, P(A|B) = P(A and B) / P(B). In words: out of all the times B happens, how often does A happen too? The denominator throws away every outcome where B didn't happen, leaving you with a fresh, smaller sample space to ask the question in.
Conditional probability is what your brain does informally all day. "How likely am I to be late given I leave at 8.30am?" is conditional. "How likely is it to rain given there are dark clouds?" is conditional. The formal version just lets you assign numbers and combine them without contradicting yourself.
## How do you visualise conditional probability?Draw a rectangle (the whole sample space) and two overlapping circles inside it called A and B. The full rectangle has probability 1. Each circle has its own probability - P(A) and P(B) - and the lens-shaped overlap in the middle is P(A and B), the probability that both happen.
To compute P(A|B), mentally erase everything outside circle B. Circle B is now your entire universe - its probability is rescaled to 1. The lens-shaped overlap, which used to be P(A and B) within the rectangle, now occupies a fraction of circle B. That fraction is exactly P(A and B) / P(B). That's the formula, derived purely by squinting at a diagram.
This is why the formula has a division in it, and why the denominator is P(B) rather than P(A). We're conditioning on B, so B becomes the new total. Swap to P(B|A) and the denominator becomes P(A) instead - same overlap, different rescaling. The two conditional probabilities are usually not equal, and confusing them is the root cause of most probability mistakes (we'll see this in the medical testing section).
You flip a fair coin twice. Someone tells you at least one of the flips came up heads. What's the probability that both came up heads?
Most people answer one-half, because the unknown flip is fifty-fifty. The right answer is one-third. Here's why.
Draw a tree. The first flip branches into Heads and Tails. Each of those branches into Heads and Tails again. You get four equally likely outcomes: HH, HT, TH, TT - each with probability 1/4.
The information at least one heads is event B. It's true for HH, HT, and TH, but not for TT. So P(B) = 3/4. The event both heads is event A, true only for HH, so P(A and B) = P(HH) = 1/4. The conditional probability is:
P(both heads | at least one heads) = P(A and B) / P(B) = (1/4) / (3/4) = 1/3.
The intuition that trips people up is the assumption that one specific flip is now "locked in" as heads. But all you actually know is that at least one of the two flips was heads. Three of the four equally likely scenarios (HH, HT, TH) are still on the table, and only one of them is HH. Conditioning correctly means restricting to the right subset of outcomes - not arbitrarily fixing a coin's value.
**When intuition fails, draw the tree** Whenever you find yourself arguing about a conditional probability with a friend, stop and enumerate every equally-likely outcome on a tree. Cross out the branches that violate the condition. The probability you want is just (favourable surviving branches) / (all surviving branches). It's almost always faster than arguing in words and almost always settles the dispute. ## How does conditional probability apply to medical tests?Suppose a disease affects 1 in 1,000 people in the general population. A test for the disease is 99% accurate, in the sense that it correctly flags 99% of true positives and correctly clears 99% of true negatives. You take the test and it comes back positive. What's the probability you actually have the disease?
The intuitive answer is 99%. The actual answer is closer to 9%.
To see why, imagine running the test on 100,000 people. About 100 of them genuinely have the disease (the 1-in-1,000 base rate). The test correctly catches 99 of those - true positives. The remaining 99,900 don't have the disease, but the test wrongly flags 1% of them - that's 999 false positives. So out of 99 + 999 = 1,098 total positive results, only 99 are real. Conditional on a positive test, the chance you have the disease is 99 / 1,098 ≈ 9%.
This is conditional probability done correctly. The 99% accuracy figure is P(positive test | disease). The number you actually want is P(disease | positive test). They're not the same number - and when the underlying disease is rare, they can differ by an order of magnitude. This is a real problem in medicine. Studies have repeatedly shown that even experienced doctors often invert these probabilities, advising patients much more pessimistically than the data warrants.
The technical name for this confusion is the base-rate fallacy. The fix is mechanical: always ask whether the headline number is the chance of the symptom given the condition, or the chance of the condition given the symptom. They are conditional probabilities running in opposite directions, and switching between them is exactly what Bayes' theorem is for.
## How does Bayes' theorem flip a conditional probability?Bayes' theorem is just a rearrangement of the conditional probability formula. Starting from P(A|B) = P(A and B) / P(B) and P(B|A) = P(A and B) / P(A), you can solve both for P(A and B), set them equal, and rearrange to:
P(A|B) = P(B|A) × P(A) / P(B)
That's the whole theorem. In the medical-testing example, A is having the disease and B is testing positive. We knew P(B|A) = 0.99 (the test's sensitivity), P(A) = 0.001 (the base rate), and we can compute P(B) ≈ 0.011 (the share of the population who'd test positive at all). Plugging in: P(A|B) = 0.99 × 0.001 / 0.011 ≈ 0.09. Same 9% answer, derived directly.
The reason Bayes' theorem matters is that the world rarely hands you the conditional probability you want. It hands you the one running the other way. A weather forecast tells you P(this morning's clouds | rain later) implicitly through years of meteorology data, but you want P(rain later | this morning's clouds). A spam filter knows P(this word appears | spam), but you want P(spam | this word appears). Bayes' theorem is the bridge.
For a deeper walk through Bayesian reasoning in everyday decisions - including how to update your beliefs as evidence accumulates - see our guide to Bayesian thinking in everyday decisions.
## Why does the Monty Hall problem fool everyone?You're on a game show. Three doors hide a car and two goats. You pick door 1. The host, who knows where the car is, opens door 3 to reveal a goat. He then offers you the chance to switch to door 2. Should you?
Most people say it doesn't matter - two doors left, one car, fifty-fifty. The right answer is to switch. Switching wins the car two-thirds of the time; staying wins one-third.
The correct way to see this is via conditional probability. Before the host opens any door, the probability the car is behind your chosen door 1 is 1/3, and the probability it's behind one of the other two doors (door 2 or door 3) is 2/3. The host's reveal of door 3 doesn't change either of those facts about your initial choice - it just tells you which of the other two doors definitely doesn't have the car. The 2/3 probability that was distributed across doors 2 and 3 is now concentrated entirely behind door 2.
The reason this trips people up is that they treat the host's action as if it were random, when it isn't. The host knows where the car is and never opens the door with the car. That structural asymmetry is the source of the conditional information. If the host opened doors at random and happened to reveal a goat, switching really would be a fifty-fifty proposition. The puzzle hinges entirely on what you condition on - the host's deliberate, informed reveal - not on the bare fact that one door has been eliminated.
## How does conditional probability work in weather forecasting?A 30% chance of rain in tomorrow's forecast is a conditional probability statement. Specifically, it's P(rain in your area tomorrow | the atmospheric conditions the forecaster has observed today). The conditioning is invisible because we don't usually write it out, but every weather forecast you've ever read is implicitly conditional on whatever the forecaster currently knows.
This matters when the conditioning changes. "30% chance of rain tomorrow" said on Monday morning is a different number from "30% chance of rain tomorrow" said on Tuesday morning. Both reference the same underlying weather, but they're conditioning on different information sets - Tuesday's forecast incorporates everything that happened in the intervening day. Forecast probabilities aren't fixed properties of the future; they're updated estimates that move as evidence comes in. That's exactly the Bayesian view of probability we touched on above.
It also explains why aggregate forecast accuracy is judged by calibration, not by individual prediction outcomes. A forecaster who says 30% on a hundred days and sees rain on roughly 30 of them is well-calibrated - even though they were "wrong" 70 times. Conditioning on the same 30%-rain prediction repeatedly tells you what 30% really means in their hands. We've covered the calibration concept in more depth in the guide to probability calibration training.
## When does conditioning change nothing?Two events are independent if knowing one happened tells you nothing about the other. Formally, A and B are independent when P(A|B) = P(A). The conditional probability collapses to the unconditional one - knowing B doesn't move the needle on A.
Coin flips are the classic example. The first flip and the second flip are independent: knowing the first came up heads tells you nothing about the second. So P(second is heads | first is heads) = P(second is heads) = 1/2. This is also why the gambler's fallacy is wrong - a roulette wheel that has just landed on red ten times in a row still has the same probability of red on the next spin, because consecutive spins are independent.
Lots of real-world events aren't independent, and treating them as if they were is one of the most expensive mistakes in probability. House prices in two neighbouring cities aren't independent - they share macro drivers. Test scores from twins aren't independent - they share genes and environment. Stock returns across companies in the same sector aren't independent - they share industry conditions. Conditional probability is what you reach for when independence breaks down.
## What are the most common conditional-probability mistakes?1. Inverting the direction. Confusing P(A|B) with P(B|A) is the single most common conditional-probability mistake. It's the medical-testing trap, the prosecutor's fallacy, and a hundred everyday misreadings of statistics. Always check which way round the conditioning runs in any number you're quoted.
2. Ignoring the base rate. Even when the conditional direction is correct, the answer changes drastically with the underlying frequency. A 99%-accurate test means very different things for a 1-in-1,000 disease and a 1-in-3 disease. We've written a deeper piece on this trap in base-rate neglect.
3. Assuming independence that isn't there. Multiplying probabilities together - P(A and B) = P(A) × P(B) - only works when A and B are genuinely independent. When they aren't, the right formula is P(A and B) = P(A) × P(B|A), which can be much smaller or much larger than the naive product depending on the correlation.
4. Conditioning on the wrong thing. Sometimes the conditioning event you're given isn't the one you should be using. The Monty Hall problem is the textbook case: people condition on "door 3 is open" when the correct conditioning event is "the host, knowing where the car is, deliberately opened a door with no car behind it". Different conditioning, different answer.
## Frequently asked questions **Q: What's the difference between conditional probability and joint probability?** Joint probability is P(A and B), the chance both events happen at all. Conditional probability is P(A|B), the chance A happens given that B has already happened. The relationship is P(A|B) = P(A and B) / P(B). Joint probability looks at the overlap; conditional probability rescales the overlap into the world where B is known to have occurred. **Q: How is conditional probability related to Bayes' theorem?** Bayes' theorem is a rearrangement of the conditional probability formula that lets you flip the direction of conditioning. If you know P(B|A) but want P(A|B), Bayes gives you P(A|B) = P(B|A) × P(A) / P(B). That's all it is. The reason it's so famous is that real-world data usually arrives in the wrong direction - you have the test's accuracy but want the patient's prognosis - and Bayes is the bridge. **Q: Are conditional probability and dependent events the same thing?** Not quite. Two events are dependent when knowing one happened changes the probability of the other - i.e. when P(A|B) ≠ P(A). Conditional probability is the language we use to express that dependence with numbers. Independent events still have conditional probabilities, but they collapse to the unconditional ones. **Q: When should I use conditional probability instead of joint probability?** Use joint probability when you're asking how often both events occur in absolute terms over the whole sample space. Use conditional probability when you've already learned that one event has occurred and want to know the chance of the other. Most real-world questions - medical tests, weather forecasts, fraud detection - are conditional, because you start with some information and want to update. **Q: Why is the answer to the two-coin-flip problem 1/3 instead of 1/2?** Because the information given is "at least one of the two flips came up heads", which is true for three of the four equally likely outcomes (HH, HT, TH). Of those three, only one (HH) is two heads. So the conditional probability of two heads given at least one heads is 1/3, not 1/2. The 1/2 answer comes from imagining you've been told a specific coin (say the first) is heads, which is a different conditioning event. **Q: What does P(A|B) = 0 mean?** It means A is impossible given B has happened - the two events are mutually exclusive in a strong sense. For example, P(card is red | card is the ace of spades) = 0, because the ace of spades is black. Note this is different from P(A) = 0; an event can have zero probability overall but non-zero probability conditional on something specific, and vice versa.Conditional probability is the workhorse of careful thinking under uncertainty. Once you can read a probability statement and immediately ask "conditional on what?", you'll spot the base-rate trap, you'll switch the direction of Bayes' theorem in your head, and you'll find that the famous "impossible" puzzles - Monty Hall, the two-child problem, the Tuesday-boy problem - all reduce to the same mechanical question of which subset of outcomes you're really restricting to.
--- # The Prosecutor's Fallacy: How Courts Get Statistics Wrong Source: https://expectedvalue.co.uk/blog/prosecutors-fallacy/ # The Prosecutor's Fallacy: How Courts Get Statistics Wrong Confusing P(evidence | innocent) with P(innocent | evidence) sends innocent people to prison. The Sally Clark case and how to spot the trap. **Answer**The prosecutor's fallacy is the mistake of confusing P(evidence | innocent) with P(innocent | evidence). Example used in the Sally Clark UK miscarriage of justice case (1999): a paediatrician testified that two cot-deaths in one family had a 1-in-73-million probability under the assumption of independence and innocence. The real question was P(innocent | two cot-deaths), which depends on the base rate of double-murder versus double-cot-death and is much closer to 50/50. Bayes' theorem is the fix; intuitive probability reasoning is the failure mode.
The prosecutor's fallacy is the most expensive arithmetic error in the history of courtroom statistics. It happens when an expert witness - or a prosecutor reading from one - quotes the chance of seeing some piece of evidence given that the defendant is innocent, then lets the jury hear it as the chance the defendant is innocent given the evidence. Those two numbers can differ by a factor of millions. Innocent people have gone to prison because of it.
## What is the prosecutor's fallacy?The prosecutor's fallacy is the mistake of confusing P(E | I) with P(I | E), where E is the evidence and I is the proposition that the defendant is innocent. They are not the same number, they are usually wildly different, and confusing them is the same arithmetic error you'd make if you confused 'the chance of a positive test given you have the disease' with 'the chance you have the disease given a positive test'. We covered this inversion in detail in our conditional probability explainer, but the courtroom version of the mistake has its own name because the stakes are uniquely high.
A typical setup looks like this. A forensic scientist testifies that the probability of finding this particular DNA profile, or this combination of fibres, or this rare blood-type pattern in a randomly chosen person from the population is one in a million. The prosecutor then summarises: 'so there is only a one-in-a-million chance the defendant is innocent.' That summary is the fallacy. The one-in-a-million number describes how often the evidence appears among innocent people, not how often a defendant who shows the evidence is innocent.
**The error in one line** Stating P(evidence | innocent) and treating it as P(innocent | evidence). To convert one to the other you need a base rate and Bayes' theorem - never just a swap of the conditioning. ## The trap in plain EnglishImagine a city of one million people. A crime is committed and the police find a partial fingerprint at the scene. A forensic expert testifies that the fingerprint pattern occurs in roughly one person in ten thousand. The defendant matches.
The prosecutor's framing: 'There is a one-in-ten-thousand chance an innocent person would match. Therefore the chance this defendant is innocent is one in ten thousand.' That sounds compelling.
But run the numbers. In a population of one million, a one-in-ten-thousand pattern matches roughly 100 people. The defendant is one of them. Without any additional evidence pointing to this specific person, the chance they are the one true culprit is closer to one in a hundred - not one in ten thousand. The prosecutor's number was off by a factor of a hundred. In real cases, the factor is often a factor of millions.
The correct number depends on the base rate: how plausible is it, before this evidence, that any given person in the suspect pool is the perpetrator? If the prior is one-in-a-million, even a one-in-ten-thousand match leaves the posterior probability of guilt around one percent. Base-rate neglect is the deeper bias under the prosecutor's fallacy - the courtroom is just the highest-stakes venue.
## Sally Clark and the cot-deaths caseThe most cited example of the prosecutor's fallacy in modern legal history is the case of Sally Clark, a Cheshire solicitor convicted in 1999 of murdering her two infant sons. Both children had died suddenly in their cots. Clark maintained both deaths were Sudden Infant Death Syndrome (SIDS).
At trial, the paediatrician Sir Roy Meadow testified that the chance of two SIDS deaths occurring in a single affluent, non-smoking family was one in 73 million. He arrived at the number by squaring an estimate that the chance of one SIDS death in such a family was about one in 8,543. The squaring assumed the two deaths were independent events.
That assumption was wrong - siblings share genes and home environment, so a second cot death is more likely after a first, not the same probability as a fresh draw. But the squaring was only the smaller error. The larger error was the prosecutor's fallacy: even if the chance of two unrelated SIDS deaths really were one in 73 million, that is the probability of seeing the evidence given innocence. It is not the probability of innocence given the evidence. To compute the second number you would need to compare it against the prior probability of two infanticides, which is also extremely rare. When statisticians did that calculation properly, the posterior probability that Clark was innocent was substantially higher than the prosecution's framing suggested.
The Royal Statistical Society publicly criticised the testimony in 2001, noting it had 'no statistical basis'. Clark's conviction was overturned on appeal in 2003 after she had served more than three years in prison. She died in 2007. Two further wrongful convictions of mothers (Angela Cannings, Donna Anthony) were quashed on similar grounds.
## People v. Collins (1968)Long before Sally Clark, the California Supreme Court reversed a robbery conviction in People v. Collins (1968) for the same arithmetic mistake. A witness in the original trial described the assailants as a Black man with a beard and moustache and a white woman with a blonde ponytail, leaving the scene in a yellow car. A prosecution expert assigned probabilities to each of these features in the general population - yellow car: 1 in 10; man with moustache: 1 in 4; girl with ponytail: 1 in 10; girl with blonde hair: 1 in 3; Black man with beard: 1 in 10; interracial couple in car: 1 in 1,000 - and multiplied them together to get one in twelve million.
Two errors stacked on top of each other. The features were not independent (a man with a beard is much more likely to also have a moustache), so the multiplication was invalid. And the resulting one-in-twelve-million figure described how rarely a random couple would match the description, not how unlikely it was that an innocent couple matching the description had been arrested. In a population of millions of couples, even a one-in-twelve-million combination would on average appear in several couples. The Supreme Court of California reversed the conviction, calling the testimony 'a veritable sorcerer' in the courtroom and warning future prosecutors against using probability without grounding it in base rates.
## The DNA random-match-probability trapModern DNA evidence has not made the fallacy go away - if anything, the larger numbers make it more dangerous, because juries are more likely to take a one-in-a-billion figure as case-closed.
The number forensic labs report is the random match probability (RMP): given a particular DNA profile, what is the chance a randomly chosen unrelated person would match. A typical autosomal STR profile RMP in casework is around one in a billion. The prosecutor's fallacy is to translate that as 'one in a billion chance the defendant is innocent.'
Two things make that translation wrong. First, the suspect pool is rarely a random sample of the global population - close relatives have far higher match probabilities, and forensic databases of millions of profiles produce expected matches purely by chance (this is sometimes called the database search fallacy). Second, RMPs say nothing about laboratory error, contamination, or sample mix-ups, which on a per-case basis can be far higher than one in a billion. The 1990s case of Andrew Deen in the UK was overturned partly because the Court of Appeal accepted the trial judge had effectively encouraged the jury into the prosecutor's fallacy on a one-in-three-million RMP.
The proper way to present DNA evidence - endorsed by the Royal Statistical Society and the National Research Council in the US - is as a likelihood ratio in a Bayesian framework, never as a direct posterior probability of guilt.
Beware too of the availability heuristic running in parallel: a number that sounds vivid and memorable - ‘one in a billion’ - feels more decisive than the same evidence reframed as a likelihood ratio, even when the underlying information is identical.
**The fallacy in one sentence** If a witness gives you a probability that conditions on innocence (or randomness, or chance), you cannot convert it into a probability about guilt without a prior. Anyone who tries is making the prosecutor's fallacy. ## The mirror image: the defendant's fallacyThe prosecutor's fallacy has a mirror twin known as the defendant's fallacy (or the defence attorney's fallacy). It runs the inversion the other way: the rare evidence narrows the field to a small group, but it doesn't pinpoint the defendant, so it shouldn't increase our belief in guilt at all.
The most cited example is from O.J. Simpson's 1995 murder trial. Simpson had a documented history of domestic violence against Nicole Brown Simpson. The defence pointed out that of all men who batter their wives, only a tiny fraction (something like 1 in 2,500) go on to murder them. Therefore, the argument went, the history of abuse was almost no evidence of guilt at all.
That reasoning ignores the conditioning. The relevant question is not 'given that he was a batterer, what is the chance he was a murderer?', it is 'given that he was a batterer and his wife was murdered, what is the chance he was the murderer?'. The statistician I.J. Good showed in Nature that conditioning on the murder having happened, the prior history of domestic violence raises the probability the husband is the killer to roughly 80 percent - a strong piece of evidence, not a negligible one. Both sides can mishandle conditional probability; the courtroom rewards whichever framing the jury hears more clearly.
This family of conditional-probability mistakes extends well beyond the courtroom. The gambler’s fallacy is the same shape of error in the opposite direction: treating independent events as if a past streak changes the next probability. Whenever someone tells you what something means without naming what the probability is conditioned on, ask for the missing piece.
## How to compute the right answerThe correct conversion uses Bayes' theorem. We covered the full mechanics in Bayesian thinking for everyday decisions, but the courtroom-specific version is short.
Let G be the proposition the defendant is guilty and E the evidence. Bayes says:
P(G | E) = P(E | G) × P(G) / P(E)
The key term the jury is rarely given is P(G), the prior probability of guilt before the new evidence. In a city of a million people with no other information about who committed the crime, P(G) for any single person is around one in a million. P(E) can be expanded as P(E | G) × P(G) + P(E | not G) × P(not G).
For a one-in-ten-thousand match in a city of a million: P(E | G) is roughly 1 (the perpetrator obviously matches their own evidence), P(G) is 1 in a million, P(E | not G) is 1 in 10,000, and P(not G) is essentially 1. Working it through gives P(G | E) ≈ 0.01 - about one percent - not the prosecutor's framing of 99.99% guilty.
The same calculation done with the Sally Clark numbers, conservatively assigning a prior probability of double infanticide and a prior probability of double SIDS, yields a posterior probability of innocence many orders of magnitude higher than the one-in-73-million figure the jury heard.
## How to spot the prosecutor's fallacyYou don't need to be a statistician to catch the fallacy in the wild. A small checklist suffices.
P(evidence | something), ask what the something is. If it's 'innocence' or 'random chance', the speaker has not yet given you a probability of guilt.The prosecutor's fallacy is just the highest-profile version of a much wider failure mode. Whenever you read that 'a 99% accurate test for a one-in-a-thousand condition came back positive,' the headline-grabbing number is P(positive | sick), not P(sick | positive). The two are connected by base rates and Bayes. We've worked through the medical-test version of this trap in the false positive paradox.
The same shape appears in airport security alarms, fraud detection, polygraph testing, performance reviews, school admissions tests, hiring screens, and any system that flags rare events with imperfect signals. The pattern is always: an impressive sounding probability conditioned on the wrong thing, treated as if it were conditioned on the thing the audience cares about. Once you see the pattern you cannot un-see it.
## Frequently asked questions **Q: Is the prosecutor's fallacy the same as base-rate neglect?** They are closely related. Base-rate neglect is the broader failure to factor in prior probabilities; the prosecutor's fallacy is the specific courtroom manifestation in which a conditional probability is presented as if the conditioning could be silently inverted. You cannot commit the prosecutor's fallacy without also neglecting the base rate, but you can neglect base rates in plenty of contexts that have nothing to do with a courtroom. **Q: Can DNA evidence ever be presented without falling into the trap?** Yes. The Royal Statistical Society and the US National Research Council recommend presenting DNA matches as likelihood ratios - the ratio of P(evidence | guilt) to P(evidence | innocence) - and asking the jury to combine that ratio with the other case evidence. This frames the DNA result as one input into a Bayesian update rather than a standalone verdict. **Q: Does the prosecutor's fallacy only apply to expert testimony?** No. It can be made by judges, journalists, witnesses, or the lawyers themselves. The Sally Clark case is the most famous because of an expert witness, but the Collins case turned on a prosecutor's closing argument and several appellate opinions have flagged judicial directions that effectively invited the jury into the same error. **Q: What is the database search fallacy?** When a suspect is identified by trawling a large database of DNA profiles for a match, the probability of finding at least one matching profile by chance is much higher than the per-person random match probability. Treating the per-person RMP as the relevant figure in this scenario is sometimes called the database search fallacy and is essentially a special case of the prosecutor's fallacy. **Q: Has the prosecutor's fallacy been formally banned in any jurisdiction?** No jurisdiction has banned it by name, but several appellate decisions have set strong precedents. In the UK, R v. Doheny and Adams (1996) gave detailed guidance on how DNA evidence should be presented to juries. In the US, People v. Collins remains a leading citation. Both rulings encourage Bayesian framing rather than the bare quotation of conditional probabilities. ## Bottom lineThe prosecutor's fallacy is one swap of two letters - the difference between P(E | I) and P(I | E) - but the consequences have ranged from overturned convictions to years of wrongful imprisonment. The fix is not advanced statistics; it is the discipline to ask, every time you hear a courtroom probability, which way is the conditioning running, and what is the prior. Once you have those two numbers, Bayes does the rest. Without them, no impressive-sounding ratio tells you anything about whether the defendant did it.
The Dunning-Kruger effect describes a cognitive pattern where people with low ability in a domain overestimate their competence, while genuine experts often underestimate theirs. The original 1999 Dunning + Kruger paper measured this across humour, grammar, and logic tests at Cornell. The mechanism: the skill to perform well in a domain is the same skill needed to evaluate performance accurately. Practical implication: when a topic feels easy after one hour of reading, you may be at the early-confidence peak rather than at competence.
The Dunning-Kruger effect is the cognitive bias most people have heard of and most people misunderstand. The popular version - usually presented as a chart with 'Mount Stupid' towering on the left and a flat 'plateau of sustainability' on the right - is so widely shared that it's almost a meme. The actual research is narrower, more nuanced, and arguably more useful.
The short version: people who are bad at something tend to overestimate how good they are. People who are very good at something tend to slightly underestimate themselves. The gap between perceived ability and actual ability is where bad decisions get made.
This guide explains what the research shows, what it doesn't show, where the famous chart came from (it's not from the original paper), and - most importantly - how to apply the insight to your own thinking and decisions. If you've ever wondered why the most confident person in the meeting is often the most wrong, this is the bias to understand.
**What this article is about** Calibration: the gap between how confident you are and how often you're right. The Dunning-Kruger effect is one specific case of bad calibration. Throughout the article we'll come back to the practical question: how do you calibrate your own confidence so you make fewer overconfident mistakes? ## What the Research Actually FoundIn 1999, two psychologists at Cornell - David Dunning and Justin Kruger - published a paper with the unwieldy title 'Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments'. The methodology was straightforward. They asked university students to take tests in domains like grammar, logical reasoning, and humour. After the test, students estimated how well they'd done - both as a raw score and as a percentile relative to other students.
The headline finding: students in the bottom quartile (the worst performers) consistently estimated they'd performed in roughly the 60th-70th percentile. They thought they'd done above average when they'd actually done dreadfully. Students in the top quartile, meanwhile, were closer to accurate but slightly underestimated their relative performance - partly because they assumed the questions had been similarly easy for everyone.
Crucially, the bottom-quartile students could not revise their estimates upward when shown others' performance. They didn't have the metacognitive skill - the ability to evaluate their own thinking - to recognise their answers as wrong. The very thing that made them bad at the task (poor reasoning, fuzzy grammar) was the thing they'd need to recognise that they were bad. This is the genuinely interesting bit of the paper: incompetence is in some sense self-concealing.
## The Famous Chart Is Mostly WrongThe chart you've seen - with 'Peak of Mount Stupid' on the left, 'Valley of Despair' in the middle, and 'Plateau of Sustainability' or 'Slope of Enlightenment' on the right - is not from the original paper. It's from the broader pop-psychology world (variously attributed but not properly sourced) and it dramatically overstates what the research showed.
The actual data shows something much milder: a fairly steady upward trend in self-assessed ability with actual ability, with the bottom performers overestimating themselves by roughly 30-50 percentile points. There's no Mount Stupid. There's no Valley of Despair. There's just a gentle disconnect between what bottom-performers think they know and what they actually know.
This matters because the popular narrative - 'beginners are wildly overconfident, then experience humbles them, then real expertise rebuilds confidence' - describes a story arc that fits films and personal-growth books better than it fits actual data. The real effect is smaller, less dramatic, and applies more narrowly than the meme suggests.
**Why the meme spread anyway** The Mount Stupid chart is a great story. It's narratively satisfying, easy to remember, and gives us a way to point at confident people we disagree with and say 'they're on Mount Stupid'. But pointing the chart at others without applying it to yourself is the most Dunning-Kruger thing you can do. ## Why Bottom-Performers Overestimate ThemselvesTo know that an answer is wrong, you usually need to know what 'right' looks like. If you don't understand grammar, you can't tell that your sentence is grammatically wrong. The bias isn't 'I'm a beginner so I should be humble' - it's that beginners don't have the equipment to detect their own errors.
If you've only ever seen your own work, you don't know how it compares to expert work. The novice writer thinks their first draft is decent because they've never read a really good draft of the same kind of thing.
Most people, on most days, default to mild overconfidence. It's not a flaw of bottom-performers specifically - top-performers also rate themselves highly. The Dunning-Kruger gap is the bottom group failing to also recognise the gap between their self-assessment and reality.
The effect is much stronger on tasks that have a clear right answer (logic, grammar) than on subjective tasks. On hard tasks, top-performers also become overconfident - they assume the questions have a 'trick' they've spotted that others have missed.
## Real-World Examples Where This MattersThe retail investor who beat the market in their first year of trading often genuinely believes they have a system. They don't have the reference frame to know that beating the market for 12 months is roughly what you'd expect by chance for half of all participants. The under-confident expert, meanwhile, is the value investor who admits they don't know which stock will go up but does know which businesses are reliably mispriced - and stays calibrated.
This connects directly to expected value: bad decisions don't come from doing the maths wrong. They come from being wrong about your confidence in the inputs.
In a famous Swedish study, 88% of drivers rated themselves as above-average. Mathematically impossible. The bottom-quartile drivers - the ones causing most of the accidents - were the most likely to rate themselves as 'better than average'. The top drivers were close to accurate.
The most confident candidate in an interview is often not the most competent. Real expertise frequently comes paired with the ability to articulate uncertainty, edge cases, and what could go wrong. Hiring on confidence selects for the wrong tail of the Dunning-Kruger distribution.
The dangerous version of this bias is in domains where consequences are real. The patient who has Googled their symptoms and confidently rejects the GP's diagnosis. The amateur who reads one article on tax law and represents themselves in court. Confidence without calibration is genuinely costly here.
Strong opinions on complex topics - economic policy, geopolitics, public health - frequently cluster among people with the least exposure to the actual evidence. Experts in these fields are usually the ones hedging hardest because they know how messy the data is.
## How to Calibrate Your Own Confidence **The single most useful exercise** Before checking the answer to anything, write down a confidence level (e.g., 70%, 90%, 99%). Track this over time. After 100 such predictions, you'll find your 90%-confident predictions were correct in roughly 75% of cases. That gap is your overconfidence. Closing it is what calibration training is.Don't say 'GDP will grow by 2%'. Say 'I'm 80% confident GDP will grow between 1% and 3.5%'. This forces you to think about how much you actually know. People who write tight intervals on questions they don't know much about are the most miscalibrated.
Find out what people who do this for a living disagree about. If even the experts can't agree on something, you almost certainly should not be confident either. The disagreement among experts is your single best calibration signal.
Vague predictions ('the economy is going to do badly soon') can never be wrong. Specific predictions ('I think the FTSE will be below 7000 by year-end with 70% confidence') are checkable. You only learn calibration from making predictions that get tested.
If you can't remember a single thing you've changed your mind about in the last year, you're not learning. Strong opinions held loosely is the right shape - strongly held until evidence comes in, then updated. People who never change their minds are people who don't read the evidence.
Before you commit publicly to a view, ask: 'what would have to be true for me to be wrong about this?' If you can't think of anything, you're not seeing the question clearly. The discipline of articulating what would change your mind is calibration training in itself.
Sounds simple. Most people are bad at it. The number of conversations that would be improved if more people were comfortable saying 'I haven't thought about this enough to have a view' is roughly all of them. There is no social cost to admitting uncertainty among people whose opinions are worth having.
## What the Effect Doesn't MeanThe Dunning-Kruger effect is widely overapplied. A few things it does not show:
The Dunning-Kruger effect is one of about a dozen biases that systematically warp how we estimate probabilities. Knowing they exist isn't enough - most of these biases are unaffected by the knowledge that they exist. Active calibration practice is what closes the gap.
For the underlying probabilistic reasoning, start with thinking in probabilities and expected value explained. For closely related biases, hindsight bias, base rate neglect, and the difference between correlation and causation are the next most important to understand.
The best single-line takeaway: confidence is cheap; calibration is expensive. People who pay the cost of staying calibrated make consistently better decisions than people who don't, across the domains the calibration research literature has studied.
--- # The Law of Large Numbers: Why Casinos Always Win Source: https://expectedvalue.co.uk/blog/law-of-large-numbers/ # The Law of Large Numbers: Why Casinos Always Win The Law of Large Numbers explains why a single bet is wildly volatile but casinos profit reliably - the maths behind insurance and long-run averages. **Answer**The law of large numbers says that the average outcome of many trials converges on the expected value, even though any single trial can be miles off. Casinos rely on this: every individual game might swing for or against the house, but across a million bets the small house edge becomes a near-guaranteed profit. The same logic applies to insurance, sports betting, and long-term investing.
A single roulette spin is one of the most uncertain events in everyday life. Whether the ball lands on red or black is, as far as the gambler is concerned, an unforeseeable coin-flip. And yet a casino's profit at the end of the year is so predictable that it shows up in published financial statements with the consistency of a utility bill. The thing that bridges the wild uncertainty of one spin and the boring reliability of the year-end accounts is the Law of Large Numbers - and once you understand it, a surprising amount of how the modern world works (insurance, opinion polls, fund management, scientific replication) becomes much clearer.
**What the Law of Large Numbers actually says** If you repeat a random experiment many times under the same conditions, the average of the outcomes converges to the expected value. The more trials you run, the closer your observed average gets to the true average - and the gap is governed by mathematics, not luck. ## What does the law of large numbers actually say?The Law of Large Numbers (LLN) is one of those mathematical results that sounds obvious until you try to state it precisely. The casual version is: 'In the long run, things average out.' That's true but it hides three subtle and important details.
The first detail is what 'in the long run' means. The Law guarantees convergence to the expected value as the number of trials goes to infinity, but it makes no promise about how quickly. Ten flips of a fair coin can produce eight heads. A hundred flips can produce 60 heads. A thousand flips will rarely produce more than 540 heads, and a million flips will almost certainly land between 49.9% and 50.1% heads. The convergence is real, but the speed of convergence depends heavily on the experiment.
The second detail is what 'averages out' means. The Law applies to the average outcome (or equivalently, the proportion of one type of outcome). It does not apply to the absolute count or the running difference. After a million coin flips you should expect very close to 500,000 heads, but the absolute gap between heads and tails will tend to grow as you flip more coins, not shrink. The proportion converges; the absolute deviation does not. This is the source of the gambler's fallacy - see our Gambler's Fallacy explainer for why a string of reds doesn't mean black is 'due'.
The third detail is that there are actually two Laws of Large Numbers - the Weak Law and the Strong Law - which differ in how strict their convergence guarantees are. For practical purposes you don't need to care which one you're invoking; both deliver the same intuition.
## A worked example: 10 flips, 100 flips, 10,000 flipsImagine flipping a fair coin and recording the proportion of heads after each flip. Here's roughly what happens in a typical run:
The pattern is precise: the standard deviation of the sample proportion shrinks in proportion to one over the square root of the sample size. Quadruple the sample, halve the noise. This is why a poll of 1,000 people has a margin of error around ±3 percentage points and a poll of 4,000 has a margin around ±1.5 - sample size buys precision, but with diminishing returns.
## Why casinos always winEvery casino game has a built-in mathematical edge for the house. On European roulette, the wheel has 37 slots - numbers 1 to 36 plus a single zero. A bet on red pays even money, but red wins on only 18 of 37 spins. The expected value of a £1 bet on red is therefore (18/37) × £1 + (19/37) × (−£1) = −£0.027. Every spin loses you, on average, 2.7p per pound staked. (American roulette has a double-zero, which lifts the house edge to 5.26%.)
For a single player on a single spin, that 2.7p edge is invisible - the variance is so much larger than the edge that any one outcome is dominated by luck. But the casino isn't playing one spin against one player. It's playing thousands of spins per hour across hundreds of tables, with millions of spins per year company-wide. By the Law of Large Numbers, the casino's actual return per pound staked converges to that 2.7% expected edge with extraordinary precision. The variance per spin is enormous; the variance of the year's gross gambling revenue is tiny in proportional terms.
This is the punchline: a casino isn't beating any individual gambler - at the table level, individual gamblers regularly win. The casino is beating the aggregate. It's running enough trials, at a small but consistent edge, that the Law of Large Numbers converts a tiny edge into a near-deterministic profit stream.
**The same logic applies to your own betting** If you have a small positive edge in any repeated bet (matched betting, advantage poker, certain prediction-market trades), the Law of Large Numbers is on your side too - but only if you have the bankroll to weather the variance over hundreds or thousands of trials. A 5% edge with a £100 bankroll is not the same as a 5% edge with a £10,000 bankroll. The Kelly Criterion (see our Kelly guide) sizes bets to extract the long-run edge without going bust to short-run variance. ## Insurance: the same maths from the other sideInsurance companies use the Law of Large Numbers as the basis of their entire business model - and unlike casinos, they're not extracting a profit from chance, they're transferring risk from individuals (who experience high variance) to a pool (where the average is highly predictable).
Take car insurance. For an individual driver, whether they crash this year is a high-variance event: most years it doesn't happen at all, but when it does the cost can be tens of thousands of pounds. For an insurer with one million policy-holders, the proportion who crash this year is one of the most predictable numbers in the company. The standard deviation of the loss ratio is small enough that the insurer can quote premiums a year in advance and trust the maths.
The premium you pay is roughly the expected cost of your individual risk plus a margin (administration, profit, regulatory capital). You're paying to convert a high-variance personal cost into a low-variance regular payment, and the insurer is in the business of pooling enough customers that the Law of Large Numbers makes their portfolio outcome reliable. It's the same convergence the casino enjoys - used as a service to customers rather than as a hidden tax on them.
For a deeper dive into how insurance pricing actually works, see our insurance and probability guide.
## Sample size in research and pollsThe same maths underlies every opinion poll and almost every published research finding. When a polling firm reports '52% support, with a margin of error of ±3 points', that margin is computed from the standard error of a proportion at the given sample size. Polls aren't approximate because the methodology is sloppy - they're approximate because the Law of Large Numbers gives precise but finite convergence at finite sample sizes.
A practical heuristic: for a binary question (yes/no, candidate A vs candidate B), the margin of error at a 95% confidence level is approximately 1/√n × 100 percentage points. With 1,000 respondents, that's about ±3.2 points. with 100 respondents, it's about ±10 points. Below 100 respondents the margins are wide enough that the result is barely informative, and above 5,000 respondents the marginal precision gain becomes very expensive to buy.
The same intuition explains why we should be skeptical of striking findings in small studies. A medical trial of 30 patients reporting a 'large' effect is much weaker evidence than a trial of 3,000 patients reporting a 'small' effect - the small study's headline result is much more likely to be a fluctuation that won't replicate. The replication crisis in social science and biomedicine is largely a story about people forgetting that small samples have wide error bars and treating point estimates as reliable.
## Common confusions worth flagging ## 1. The Law does not say short-run results balanceThis is the gambler's fallacy. After ten reds in a row at roulette, the next spin is no more likely to come up black than usual - the wheel has no memory. The Law of Large Numbers says the long-run proportion of red converges to 18/37 (about 48.6%), but it makes no claim about any specific upcoming spin. A run of ten reds is rare in advance but irrelevant once it has happened; what matters for the next spin is only the underlying probability.
## 2. The Law says nothing about absolute countsIf you flip a fair coin a million times, the absolute gap between heads and tails will, with high probability, be larger than after a thousand flips. The proportion converges to 50%, but the count of heads minus the count of tails tends to grow (in absolute terms) with sample size. People intuitively expect 'balance' - that running deficits should narrow - and that's wrong. The convergence is in proportion, not in count.
## 3. The Law assumes independent and identically distributed trialsThe classical Law of Large Numbers assumes each trial is drawn from the same distribution and is independent of the others. In real life, trials often violate one or both. Stock returns are not independent across years; sports outcomes are not independent within a tournament. Convergence still typically happens but at slower rates and with bigger occasional deviations than the classical formula predicts. This is why fund managers can have 'great decades' that disappear over the next decade - the LLN is still operating, but the relevant timescale is longer than a career.
## 4. The Law does not equalise everyone's experienceEven at very large sample sizes, individual experience can differ wildly from the long-run average. A million coin-flippers will produce a tight distribution of average outcomes, but a few of them will have unusually long head streaks at some point during their flipping. The aggregate looks neat; the individual stories within the aggregate are noisy. This is closely related to the concept of ergodicity - the question of whether time averages and ensemble averages are the same - and is one of the most-missed subtleties in interpreting probabilistic claims.
## How to think with the Law in everyday decisionsThe Law of Large Numbers is most useful as a check on intuition. Three concrete prompts:
1. **If a small sample is producing a striking result, distrust it** - Anything based on n < 30 in research, or fewer than 30 trades / matches / decisions in your own life, has wide error bars. Wait for more data before drawing conclusions. 2. **If something has 'always worked' across many trials, take it more seriously** - Strategies that have produced positive expected value over hundreds or thousands of trials are much more credible than strategies that 'worked five out of seven times last year'. 3. **If you have a real edge, plan for the variance** - A small edge produces reliable profit only if you can run enough trials. Bankroll, time horizon, and bet sizing all matter - see the Kelly Criterion link above and our position sizing guide for how to think about this concretely. 4. **If you observe something extreme, suspect regression to the mean** - An exceptional sales month, a runaway-good fund manager, a single very-fast race time - these are usually a mix of skill and luck, and the next observation will tend to be closer to the average. See our regression to the mean explainer. ## Frequently asked questions **Q: Does the Law of Large Numbers mean a losing streak is 'due' to end?** No. The Law of Large Numbers says nothing about any specific upcoming trial. The probability of a fair coin landing heads on the next flip is 50% regardless of the previous results. A losing streak is unusual in advance but irrelevant in retrospect - the next outcome depends only on the underlying probability, not on what came before. Believing otherwise is the gambler's fallacy. **Q: What's the difference between the Weak and Strong Law of Large Numbers?** Both say that the sample average converges to the expected value as the sample size grows. The Strong Law makes a stricter mathematical claim - that the sample average converges 'almost surely' (with probability 1) - while the Weak Law makes a slightly weaker claim about probabilities approaching 1. For almost every practical use the distinction doesn't matter; both deliver the same intuition that long-run averages converge. **Q: How does the Law of Large Numbers connect to expected value?** Expected value is the long-run average outcome of a repeated random trial. The Law of Large Numbers is what justifies treating expected value as a meaningful number - without convergence, the expected value would be a theoretical construct with no real-world counterpart. With convergence, it's the number your actual results approach as you run more trials. See our expected value guide for how to compute and use it. **Q: Why do polls have a margin of error?** Because the Law of Large Numbers delivers convergence at a finite rate, not instantaneously. A sample of 1,000 voters gives an estimate of the population average that's typically within ±3 percentage points of the true value, at 95% confidence. Doubling the sample size shrinks the margin of error by a factor of √2 (about 1.4x), so the precision gains are real but with diminishing returns. **Q: Can I use the Law of Large Numbers to predict the next coin flip?** No. The Law applies to long-run averages, not individual trials. Each coin flip is independent of the others - the probability of heads is 50% on every flip, no matter what came before. The Law tells you about the average over many flips, not about the next flip specifically. **Q: Why doesn't the Law of Large Numbers apply to stock-market 'cycles'?** It does, but the underlying distribution is unstable. Classical Law of Large Numbers assumes trials are independent and drawn from the same distribution. Stock returns are correlated across time, and the distribution itself shifts (different decades have different mean returns). Convergence still happens, just slowly enough that a 30-year career can finish before the long-run average becomes obvious. This is why short-run market patterns - even ones that look highly reliable - should be treated with skepticism. **Q: Is the Law of Large Numbers the same as the Central Limit Theorem?** No, but they're closely related. The Law of Large Numbers tells you that the sample average converges to the expected value. The Central Limit Theorem tells you about the shape of the distribution of sample averages - specifically that, for many distributions, the sample average is approximately normally distributed around the expected value, with a standard deviation that shrinks as 1/√n. The Law tells you what to expect on average; the CLT tells you how spread out the deviations are. --- # The Availability Heuristic: Why Vivid Examples Mislead Source: https://expectedvalue.co.uk/blog/availability-heuristic/ # The Availability Heuristic: Why Vivid Examples Mislead Why we judge risk by what comes to mind first - fear of flying, market news, terrorism - and the practical techniques to debias your thinking. **Answer**The availability heuristic is the mental shortcut where we judge how likely something is by how easily examples come to mind. Plane crashes, shark attacks, and lottery wins all feel more probable than they are because we hear about them constantly. The fix is to ground estimates in actual base rates rather than the most vivid story you can recall.
The availability heuristic is the mental shortcut where we judge how likely or common something is by how easily examples come to mind. Vivid, recent, or emotionally charged events feel far more probable than they actually are. The shortcut helped our ancestors react quickly to immediate dangers, but in a world saturated by news, social media, and selectively memorable stories, it now leads us into systematically wrong assessments of risk, frequency, and impact.
Identified by Amos Tversky and Daniel Kahneman in 1973, the availability heuristic explains a wide range of everyday misjudgements - from why people fear flying more than driving despite the contrary statistics, to why investors pile into the stock that just appeared on the front page, to why the latest crime story rewires whole neighbourhoods' perception of safety. This guide covers the mechanism, four canonical examples, and practical techniques to debias your decisions when memory is doing the maths.
## What does the availability heuristic actually do?When you ask yourself how common a kind of event is, you do not run a statistical query in your head. You search memory for examples and judge frequency by how quickly and vividly they appear. If many examples come up easily, the event feels common. If you struggle to recall any, it feels rare.
This works adequately when memory accurately samples reality. It breaks when memory is biased - and memory is biased almost everywhere that matters. Vivid events stick. Repeatedly reported events stick. Recent events stick. Self-relevant events stick. None of these factors correlate well with actual frequency in the wider world. Once memory is unrepresentative, your felt sense of probability follows it into error.
The mechanism feels honest from the inside. You are not making things up; you are remembering real events. The bias is invisible because the brain treats ease of recall as a reliable cue for prevalence - a heuristic that worked when our information environment was small, intimate, and balanced, and that fails when news cycles, social feeds, and entertainment systematically over-represent dramatic events.
**Tversky and Kahneman's letter-frequency demo** The classic 1973 experiment asked people: in English text, are there more words starting with K, or more words with K as the third letter? Most people answer 'starting with K' - they can recall such words easily (kitten, kettle, key). The truth is the opposite: K-as-third-letter words (acknowledge, ask, take) are roughly twice as common. Memory's accessibility advantage for word starts produced a confidently wrong answer. The same mechanism is operating any time you judge frequency by recall. ## Example: why does flying feel scarier than driving?Plane crashes are extraordinarily rare. Per passenger-mile, commercial aviation is roughly 100 times safer than driving in the UK. Yet a non-trivial fraction of adults are afraid of flying, and a significant minority refuse to fly at all. Almost nobody is afraid of getting in a car the same week.
Availability explains the gap. Plane crashes are catastrophic, televised, and saturate news cycles for days. Car crashes - which kill far more people per year - happen one or two at a time and rarely attract national attention. The mind picks up on the easily recalled imagery of the rare event and over-weights it; the routinely fatal event remains background noise.
The same pattern operates with shark attacks (about 10 fatalities a year worldwide, vs ~700,000 deaths from human-on-human violence), terrorism (extraordinarily rare in most Western countries, but hugely available because every incident is newsworthy), and home invasions (statistically rare, but intensely newsworthy because of the fear they evoke). In every case, vivid availability inflates the felt probability far above the statistical one.
## Example: how does recent news mislead investors?An individual investor reads about a hot AI stock that just announced a great earnings report. The headline, the chart, and the bullish CNBC clip are all vivid in mind. The base-rate question 'most stocks underperform the index' does not enter consciousness. Memory has supplied a confident-looking case study; the broader market context - silent, abstract, unmemorable - does not.
This is the same machinery that fuels every individual-stock-picking bubble. Tesla, GameStop, the dotcom darlings, the meme stocks of 2021 - each generated a cascade of vivid stories that made buying feel obvious to people whose mental model was dominated by what they had just read. Whether they made or lost money mostly depended on luck and timing, not the felt confidence the availability cascade provided.
The defence is institutional: index funds, automatic monthly investment, and a rule that no single stock above a small share of net worth - all of which sidestep availability by removing the moment-to-moment decision. We cover the broader probability framework in our guide to risk vs uncertainty and the role of correlation vs causation in interpreting investment news.
**Recency amplifies availability** Availability and recency interact. After a single shark attack hits the news, swimmers across the country avoid the sea for weeks. After a market crash, investors avoid stocks for years - even though the historic best buying opportunities have come immediately after crashes. The bias is strongest when the vivid event is fresh; it fades as memory rebalances. Decisions made in the first 48 hours after a dramatic event are the most likely to misweight base rates. ## Example: how does the media warp your sense of risk?If you watch any country's evening news for a fortnight, you will come away with a visceral sense that violent crime, child abduction, terrorism, and pandemics are dominating modern life. Yet long-run data on most of these has been static or declining for decades in most developed countries. The disconnect comes from how news editors select stories: vivid, alarming, and emotionally engaging items get priority; statistically common but quiet outcomes (most journeys are safe, most weeks are not catastrophic) generate no headlines.
The cumulative effect is that heavy news consumers consistently overestimate violent crime rates, foreign-born terrorism risks, and abduction rates while underestimating mundane causes of harm - heart disease, drug overdoses, suicide, road traffic injuries, and falls in older adults. The availability heuristic, applied to a curated information diet, distorts an entire population's risk perception in the same direction.
This effect was rigorously documented in Hans Rosling's Factfulness, which showed that even highly educated audiences - including journalists and policymakers - score worse than chance on basic global development questions, partly because their availability sample is skewed toward catastrophic stories.
## Example: how does availability bias affect hiring?Annual performance reviews are supposed to assess a year of work. In practice, they often weight the most recent month or two disproportionately - the project that finished last week is vivid; the steady performance of months ago is not. Managers who believe they are weighing the year evenly are typically weighing the recent past at three to five times its proportional weight, simply because of memory accessibility.
The same effect runs in interviews. A candidate who delivered one memorable answer is rated higher than an equally consistent candidate whose answers were strong but unremarkable. A candidate whose CV mentions a famous company is judged more capable than a candidate from an equally good but less recognised firm - the famous company is more available in memory and so signals expertise more strongly.
Both can be debiased. For reviews, monthly or quarterly notes, written at the time, force a more uniform sample of memory. For interviews, structured scoring against fixed criteria - completed independently before discussion - flattens the availability advantage of the memorable candidate. We discuss the broader logic in our piece on probabilistic frameworks for career decisions.
## What actually works against the availability heuristic?Awareness alone does not fix the bias - Tversky and Kahneman themselves emphasised that knowing about availability does not make you immune to it. What helps is changing the inputs to your judgement so memory is less load-bearing.
Availability rarely fires in isolation. It overlaps with several other well-documented biases in ways that compound the effect.
The availability heuristic is the bias that, more than any other, explains why people in modern information environments feel anxious about the wrong things. Combat it not by trying to remember it in the moment - that does not work - but by changing your inputs and pre-committing to base-rate-anchored decisions for the things that matter most. For the broader toolkit, our guides to thinking in probabilities and the Dunning-Kruger effect cover adjacent territory.
--- # Probability Calibration: Predict Like a Superforecaster Source: https://expectedvalue.co.uk/blog/probability-calibration-training/ # Probability Calibration: Predict Like a Superforecaster Learn probability calibration like a superforecaster - Brier scores, drills, and free tools to sharpen your forecasts. Based on Philip Tetlock's research. **Answer**Probability calibration is the skill of matching your confidence to your accuracy: when you say you are 70% sure, you should be right roughly 70% of the time. To train it, predict outcomes with explicit probabilities, log them, and score yourself against reality using a Brier score - the gap between stated confidence and actual outcome - until that gap shrinks.
Most people are confidently wrong. They claim to be 90% sure of things that turn out true only half the time. Probability calibration is the discipline of fixing that gap - making your stated confidence match your actual hit rate. The good news, established by two decades of research from Philip Tetlock and the Good Judgment Project, is that calibration is a trainable skill. Anyone willing to keep score, sit with uncertainty, and revise their views can learn to forecast with the precision of a superforecaster.
This guide explains what calibration is, how Brier scores measure it, what Tetlock's experiments have found about the traits of top forecasters, and the specific drills and free tools that will move your numbers in the right direction.
## What probability calibration actually meansA forecaster is well-calibrated when their stated probabilities track real-world frequencies. If you say a hundred different events are 70% likely, roughly 70 of them should happen. Not 50, not 90 - 70. Calibration is about the honesty of your confidence, not about how often you happen to be on the winning side.
Accuracy and calibration sound similar but capture different things. A weather forecaster who says "40% chance of rain" every single day in a climate where it actually rains 40% of the time is perfectly calibrated, even though they look uncertain. A pundit who screams "100% certain!" and turns out to be right 80% of the time is wildly overconfident - closer to a coin flip than to truth - even though they were technically correct most of the time.
This is the lens probabilistic thinkers use to evaluate their own judgement. Once you treat your beliefs as numerical claims about the world, you can grade them. And once you can grade them, you can improve them.
**Quick definition** Calibration measures the relationship between confidence and reality. A perfectly calibrated forecaster's 60% predictions come true 60% of the time, their 80% predictions come true 80% of the time, and so on across the full probability spectrum. ## The Tetlock research: how we know calibration is trainableIn the 1980s, the political psychologist Philip Tetlock asked a simple question: are pundits any better than chance at predicting world events? He recruited 284 experts - academics, government analysts, journalists - and tracked roughly 28,000 forecasts over 20 years. The headline finding, published in 2005's Expert Political Judgment, was unflattering. The average expert was barely better than random guessing, and famous experts were often worse than less well-known ones, because their confidence outran their evidence.
Crucially, Tetlock identified two cognitive styles. Hedgehogs cling to one big idea and apply it everywhere; foxes draw from many frameworks and update freely. Foxes were systematically better forecasters. That distinction set up the next phase of his work.
From 2011 to 2015, the U.S. intelligence community ran the Aggregative Contingent Estimation tournament, pitting forecasting teams against one another on real geopolitical questions. Tetlock's Good Judgment Project didn't just win - it beat the official intelligence community estimates by such a margin that competitors withdrew. The top 2% of forecasters, dubbed superforecasters, posted Brier scores roughly 30% better than ordinary smart participants. Most importantly, they got better over time, and so did the people they trained.
That is the paper trail that justifies treating calibration as a skill, not a personality trait.
## How a Brier score worksThe Brier score, proposed by meteorologist Glenn Brier in 1950, is the standard metric for grading probabilistic forecasts of binary events. The formula is simple: take the difference between your stated probability and the outcome (1 for happened, 0 for did not happen), square it, and average across many forecasts. Lower is better. A perfect Brier score is 0; a coin-flip forecast on a 50/50 event scores 0.25; predicting 0% or 100% on something that goes the other way scores the maximum of 1.
Two properties make Brier scores useful. First, they reward calibration directly: the score penalises overconfidence and underconfidence symmetrically. Second, they decompose neatly - statisticians can split a Brier score into a calibration term and a resolution term, so you can see whether you are losing points by being miscalibrated or by being too timid (always sitting near 50%).
**Worked example** You forecast three events at 80% likely. Two happen, one doesn't. Your Brier score is ((0.8−1)² + (0.8−1)² + (0.8−0)²) ÷ 3 = (0.04 + 0.04 + 0.64) ÷ 3 ≈ 0.24. Not great - that 80% confidence cost you on the miss. Now imagine you'd said 67% (your true hit rate of 2 in 3): ((0.67−1)² × 2 + (0.67−0)²) ÷ 3 ≈ 0.22. Lower, because the confidence matched reality. ## What separates superforecasters from the restThe 2015 book Superforecasting distilled the traits that consistently distinguished top performers. None of them are rare gifts. They are habits.
- **They start with the base rate** - Before considering specifics, superforecasters look up how often this kind of event happens historically. A 5% base rate is the gravity their forecast has to escape. - **They break problems into smaller pieces** - A messy question ("will Russia invade by year-end?") becomes a chain of sub-questions ("are troops at the border?", "have evacuations begun?", "what's the diplomatic posture?"). - **They update incrementally** - Rather than swing wildly between 20% and 80% on each headline, they nudge their estimate by a few percentage points as evidence accumulates - Bayesian updating in spirit if not in formal math. - **They hold strong opinions weakly** - They commit to a number, but treat it as provisional. Being proven wrong is information, not humiliation. - **They look for disconfirming evidence** - Most people search for reasons their hypothesis is right. Superforecasters deliberately hunt for reasons it might be wrong. - **They take the outside view** - Stepping back, asking what a reasonable observer with no stake in the outcome would conclude - neutralising the gravitational pull of personal narrative. - **They keep score** - Every forecast is logged with a probability, a deadline, and a resolution. Without the score, there is no feedback. Without feedback, there is no learning.Notice what is not on this list: domain expertise. The Good Judgment Project found that subject-matter experts had no systematic edge over generalist superforecasters who applied these habits rigorously. The process beat the credentials.
## Calibration training drills you can do this week 1. **Run a trivia calibration drill** - Find 50 binary trivia questions you don't already know the answer to. Answer each one and assign a confidence between 50% and 100%. Grade yourself: of your 80% answers, how many were right? Of your 90% answers? Most people discover their 90% confidence is more like 70% accuracy. The gap is your overconfidence, made legible. 2. **Forecast 10 events with deadlines** - Pick ten things that will resolve in the next 30–90 days - election results, sports outcomes, product launches, weather. Assign a probability to each, in writing, before the resolution date. After they resolve, calculate your Brier score and compare it against later batches. 3. **Practise the pre-mortem reflex** - Before committing to a forecast, ask: "if this turns out wrong, what was the most likely reason?" Generating two or three failure modes typically pulls overconfident estimates closer to the centre. The pre-mortem is itself a tool of probabilistic thinking - see our pre-mortem guide for the full method. 4. **Argue against your own view in writing** - Once a week, write a 200-word case for the opposite of something you believe. The exercise reliably surfaces hidden assumptions and forces you to engage with the strongest version of the counter-argument, not a strawman. This is the disconfirming-evidence habit in action.The first drill is the most important. Most people have never felt the visceral surprise of seeing their 95% answers proven wrong 30% of the time. That feeling - the calibration gut-punch - is what motivates everything after.
Pair these drills with consistent record-keeping. A spreadsheet with five columns (date, claim, probability, deadline, outcome) is enough. Once a month, review the rows that have resolved and compute a rough Brier score. The trend matters more than the absolute number.
## Free tools for tracking your calibration - **Good Judgment Open** - The public-facing successor to Tetlock's tournament. You forecast on real geopolitical and economic questions and see your Brier score against a benchmark community of strong forecasters. The closest thing to formal training that exists for free. - **Metaculus** - A community forecasting platform with thousands of resolved questions, a robust scoring system, and detailed personal calibration plots that show your hit rate at each confidence level. Ideal for technology, science, and global affairs forecasting. - **Quantified Intuitions** - A lightweight web app from Sage with focused calibration drills (estimation, pastcasting, calibration-in-context). Best for short, deliberate practice sessions of 10–15 minutes. - **Calibrate Your Judgment (Open Philanthropy)** - A free trivia-style calibration trainer used inside Open Philanthropy and adjacent organisations. Excellent for the first drill above - instant feedback on hundreds of questions. - **PredictIt and Manifold** - Real-money and play-money prediction markets. The market price aggregates many forecasters' calibrated views into a single probability - see our explainer on how prediction markets work to understand what those numbers actually mean.Pick one and use it weekly. The platform matters less than the consistency. A hundred forecasts logged over six months will tell you more about your judgement than ten thousand opinions you've voiced and forgotten.
## Books to deepen your understandingAffiliate disclosure: the book links below are affiliate links - if you buy through them, we may earn a small commission at no extra cost to you. The recommendations reflect the books' standing in the field, not the commission rate.
Superforecasting: The Art and Science of Prediction by Philip Tetlock and Dan Gardner is the practitioner-friendly synthesis of the Good Judgment Project. It mixes biography, methodology, and the seven habits at length, and it is the single best entry point into calibration as a discipline.
Expert Political Judgment: How Good Is It? How Can We Know? by Philip Tetlock is the academic precursor - denser, more statistical, and the foundation for everything that came after. Read this if you want to see the original 20-year dataset and the hedgehog-versus-fox distinction in its primary form.
Two complementary classics: Thinking, Fast and Slow by Daniel Kahneman explains the cognitive machinery that produces miscalibration, and Noise by Kahneman, Sibony and Sunstein extends the analysis to the variability between experts on the same case. Together they explain why calibration is hard before showing you how to fix it. Our broader decision-making book list and the probabilistic thinking reading list place these in a wider context.
## Common calibration mistakesOverconfidence on familiar topics. The more you know about a field, the more confident you feel - and the more your confidence outruns the evidence. Domain experts are often worse-calibrated than informed generalists for exactly this reason. The fix is to keep checking your specialist forecasts against base rates the same way you would for an unfamiliar topic. Our piece on the Dunning-Kruger effect covers the cognitive mechanism behind this trap.
Hedging everywhere by sitting near 50%. The opposite failure: refusing to commit to a number out of false humility. A forecaster who calls everything 50/50 is uninformative even if technically well-calibrated, because they have provided zero useful signal. The Brier-score decomposition catches this: low calibration error but also low resolution.
Ignoring base rates. Treating every situation as unique strips away the most useful prior you have. If 5% of startups in this category succeed, your forecast for this specific startup needs a very strong reason to depart from 5%. We cover this in base rate neglect at length.
Anchoring on the first number you saw. Whatever probability you encountered first - in a headline, a friend's guess, a market price - silently pulls your own estimate towards it. Force yourself to write down a forecast before looking at any external estimates.
Forgetting to update. Sticking with a forecast made a month ago, when three relevant news items have since landed, is a classic error. Calibrated forecasters re-resolve their own probabilities at every meaningful new data point - see Bayesian thinking for everyday decisions for the underlying logic.
## Where calibration pays off in practiceCalibration is upstream of every probabilistic decision-making technique. Expected value calculations are only as accurate as the probabilities you feed in - overconfident inputs produce systematically wrong EVs. The Kelly criterion for sizing investments or bets requires a calibrated edge estimate; if you think you have a 60% chance when reality is 52%, Kelly will tell you to bet aggressively into negative expected value. Even simple risk-versus-uncertainty reasoning depends on knowing when your probabilities are real and when they are guesses dressed up as numbers.
The professional contexts where calibration training has caught on most quickly - intelligence analysis, pandemic preparedness, financial trading, AI safety research - are exactly the ones where the cost of confidently wrong predictions is highest. The lesson generalises. Any domain where you are paid for judgement under uncertainty is a domain where calibration is a leveraged skill.
## Frequently asked questions **Q: How long does it take to become well-calibrated?** The Good Judgment Project found measurable improvement in Brier scores after about 100 logged forecasts, and substantial improvement after a year of consistent practice. The first 50 forecasts mostly serve to make you confront how miscalibrated you started out. **Q: Is calibration the same as Bayesian thinking?** They are closely related but not identical. Bayesian thinking is the formal method for updating beliefs in light of evidence; calibration is the property of having those beliefs match reality. You can be a competent Bayesian and still be poorly calibrated if your priors are overconfident, and you can be well-calibrated through experience even without doing any explicit Bayesian arithmetic. **Q: What's a 'good' Brier score?** It depends on the difficulty of the questions. On the Good Judgment Project's geopolitical questions, a Brier score below 0.25 was respectable, below 0.20 was strong, and superforecasters averaged around 0.15. On easier domains the targets are lower. The right benchmark is your own past performance, not an absolute number. **Q: Can you train calibration without keeping score?** Almost certainly not. Without a written record of the probability, the deadline, and the resolution, every forecast disappears into hindsight bias - you remember the ones you got right and forget the ones you got wrong. Keeping score is what makes the feedback loop honest. **Q: Does calibration matter if I'm not making big decisions?** It matters for the small ones too. Calibrated thinking improves how you weigh advice, evaluate news headlines, and notice when a confident voice is bluffing. It is also one of the few intellectual habits that compounds - every honest probability you state is training data for the next one. ## Where to start this weekOpen a spreadsheet. Add five rows: date, claim, probability, deadline, outcome. Make ten forecasts about events that will resolve in the next 90 days - anything from sports results to product launches to news headlines. Sign up for a free Good Judgment Open or Metaculus account and add five forecasts there too. Set a calendar reminder for 90 days from now to grade them and compute a rough Brier score. Then read Superforecasting while you wait.
That is the entire training programme. The rest is repetition, honest scorekeeping, and the willingness to be wrong in writing. Within a year, your stated 80% confidence will start to mean what it says - and every probabilistic decision downstream of your forecasts will get measurably better.
--- # Sequence-of-Returns Risk: Why the Order Matters Source: https://expectedvalue.co.uk/blog/sequence-of-returns-risk/ # Sequence-of-Returns Risk: Why the Order Matters Same average return, wildly different outcomes: why the ORDER of your returns decides how long a pension pot lasts. Sequence-of-returns risk explained. **Answer** Two retirees with the SAME average return can finish tens of thousands of pounds apart if the ORDER of those returns differs. An early crash in retirement is far worse than a late one, because you sell units at the bottom to fund withdrawals. That path-dependence is sequence-of-returns risk, and a plain average hides it entirely.The average annual return on your pension pot is almost useless on its own. Two people can retire on the same day, hold the identical fund, earn the exact same set of yearly returns, and still end up with wildly different amounts of money - purely because those returns arrived in a different order. This is sequence-of-returns risk, and it is one of the few genuine free lunches in retirement planning to understand early, because the fix costs almost nothing if you plan for it in advance.
## What is sequence-of-returns risk?Sequence-of-returns risk is the danger that the timing of good and bad years - not just their average - determines how long your money lasts. It matters whenever you are adding money to a pot or taking money out of it. For a retiree drawing an income, a run of poor returns in the first few years is disproportionately damaging: every withdrawal in a falling market permanently removes units you can never buy back, so there is less capital left to recover when markets rebound.
The key word is path-dependence. Most people are taught to think about investing through the average, or the compound annual growth rate. That single number is a summary that throws away the one thing that matters most when you are living off the pot: the shape of the journey.
## Why does the order of returns matter if the average is the same?Here is the counter-intuitive part. If you never add or remove money, order does not matter at all. Multiplying by 1.2 then 0.8 gives exactly the same result as multiplying by 0.8 then 1.2, because multiplication is commutative. A lump sum left completely untouched ends up in precisely the same place regardless of the sequence.
Cash flows break that symmetry. The moment you withdraw a fixed pound amount each year, a fall early on shrinks the base that every later return compounds from. This is really a lesson about ergodicity - the average outcome across many parallel investors (the ensemble average) is not the same as the outcome you actually live through over time (the time average). The advertised average return is an ensemble statistic; your pension is a single path.
## A worked example: two retirees, same average, £75,760 apartMeet two retirees, Ada and Ben. Both retire with a pot of £500,000 and withdraw £25,000 at the end of every year to live on. Both earn the exact same five annual returns - a nasty crash, a mild fall, and three good years: -30%, -10%, +10%, +20% and +25%. The arithmetic average is identical for both (+3% a year), and so is the compounded average. The only difference is the order.
Ada has bad luck: her crash lands first. Ben has good luck: the same crash lands last. Each year the market return is applied, then the £25,000 withdrawal is taken.
- **Starting pot**: Ada: £500,000 | Ben: £500,000 - **Year 1 return**: Ada: -30% to £325,000 | Ben: +25% to £600,000 - **Year 2 return**: Ada: -10% to £267,500 | Ben: +20% to £695,000 - **Year 3 return**: Ada: +10% to £269,250 | Ben: +10% to £739,500 - **Year 4 return**: Ada: +20% to £298,100 | Ben: -10% to £640,550 - **Year 5 return**: Ada: +25% to £347,625 | Ben: -30% to £423,385 - **Final pot after 5 years**: Ada: £347,625 | Ben: £423,385 - **Gap on identical average return**: £75,760After just five years, Ada and Ben are £75,760 apart despite earning the same average return and taking the same income. Ada took her withdrawals out of a pot that had already been savaged by the year-one crash, so she was selling a larger fraction of a smaller portfolio at exactly the wrong time. Ben spent his early years withdrawing from a pot that had grown, and only met the crash once he had already banked years of gains.
Stretch this over a real 30-year retirement and the effect compounds. The same average return can leave one retiree with a healthy surplus and push an identically-invested neighbour towards running out of money. That worst-case - the pot hitting zero while you still need income - is the risk of ruin, the same downside that drives sensible position sizing in fractional Kelly staking.
## Why is an early crash so much worse in retirement?Withdrawing in a downturn does two things at once. It crystallises the loss - a paper fall becomes permanent the moment you sell units to raise cash - and it removes shares that would otherwise have participated in the recovery. A portfolio down 30% needs to gain about 43% just to get back to where it started, and if you are simultaneously drawing an income, you are handing away the very units that would have delivered that rebound.
This is why sequence risk is front-loaded. Research on retirement withdrawals consistently finds that the returns in roughly the first decade after you start drawing an income matter far more to whether the money lasts than the returns in the final decade. Survive the first ten years without a severe early crash and the danger fades sharply.
## How is this different when you are still saving?For someone still paying into a pension, sequence risk runs in reverse. An early crash is actually good news for an accumulator: your regular contributions buy more units while prices are low, and those cheap units compound as markets recover. The dangerous sequence for a saver is the mirror image - strong early gains followed by a crash right before you retire, when the pot is at its largest and a big percentage fall wipes out the most pounds.
The practical takeaway is that the decade either side of your retirement date - the last few years of saving and the first few of spending - is the danger zone. That is the window where the size of your pot and your exposure to a bad sequence are both at their peak. It is also where the gap between expected value and expected utility is widest: the maths might say stay fully invested for the highest average outcome, but the pain of a ruinous early crash is not symmetric with the joy of an extra few percent of upside.
## What does this mean for the 4% rule and UK drawdown?The famous 4% rule - the idea that you can withdraw 4% of your starting pot, rise it with inflation each year, and rarely run out over 30 years - exists precisely because of sequence risk. William Bengen derived it in 1994 by testing withdrawal rates against the worst historical return sequences, not the average ones. The 4% figure is deliberately conservative because it has to survive a retiree who is unlucky enough to retire straight into a crash.
In the UK, pension freedoms let you take an income directly from an invested pot through flexi-access drawdown. That flexibility is powerful, but it puts sequence risk squarely on your shoulders: a bad first few years can permanently impair the pot. An annuity does the opposite - it hands the sequence risk (and longevity risk) to an insurer in exchange for a guaranteed income, at the cost of flexibility and any upside. Many retirees split the difference, securing essential spending with a guaranteed floor and leaving the rest invested. The government's retirement income guidance and the free Pension Wise service are the sensible starting points before making that call.
**The mechanism in one line** Order only matters when money is moving in or out. No cash flows, no sequence risk. Add withdrawals and an early crash is permanent; add contributions and an early crash is a gift. ## How can you protect against sequence-of-returns risk? - **Hold a cash and bond buffer** - Keeping one to three years of spending in cash and short bonds means you can pause selling equities after a fall and let the pot recover, rather than crystallising losses to fund income. - **Use flexible (guardrail) withdrawals** - Trimming your income in bad years and topping it up in good ones directly attacks the problem - you take out less when the pot can least afford it. - **Build a bond tent** - Temporarily raising your bond allocation in the few years either side of retirement cushions the danger zone, then gliding back towards equities once the highest-risk window has passed. - **Secure a guaranteed floor** - Covering essential spending with a state pension plus a modest annuity means a bad sequence only ever threatens your discretionary spending, never the lights-and-food budget. - **Be realistic about the starting withdrawal rate** - The lower your initial withdrawal rate, the less a bad early sequence can hurt you. Starting nearer 3% than 5% buys a large margin of safety. ## Frequently asked questions **Q: What is sequence-of-returns risk in simple terms?** It is the risk that the order of your investment returns, not just their average, decides how long your money lasts. A run of bad years early in retirement is far more damaging than the same bad years late on, because you are selling units at low prices to fund withdrawals. **Q: Does the order of returns matter if I am not withdrawing money?** No. With no money moving in or out, reordering the same set of returns produces the identical final balance, because multiplication is commutative. Order only starts to matter once you add withdrawals or contributions. **Q: When is sequence risk at its highest?** In the roughly ten years either side of your retirement date - the last few years of saving and the first few of drawing income. That is when your pot is largest and your exposure to a bad sequence is greatest. **Q: How do you reduce sequence-of-returns risk?** Hold a cash and bond buffer so you can avoid selling equities after a fall, use flexible withdrawals that flex down in bad years, build a bond tent around the retirement date, secure essential spending with a guaranteed income floor, and keep the starting withdrawal rate conservative. **Q: Is the 4% rule safe because of sequence risk?** The 4% rule was designed around sequence risk. It was derived by testing withdrawals against the worst historical return sequences, so the figure is deliberately conservative to protect a retiree unlucky enough to retire into a crash. It is a useful anchor, not a guarantee. ## SourcesEvery car or home insurance quote hides a lever most people never touch: the voluntary excess. Slide it up and your premium falls. Slide it down and the insurer absorbs more of your risk - for a price. Most buyers set it by gut feeling, which in practice means loss aversion sets it for them.
There is a better way: treat the excess decision as an expected value calculation. It takes five minutes with the quotes you already have, and for many households it is worth a real saving every year. This guide walks through the exact calculation, a worked example with a realistic quote ladder, and the behavioural trap that pushes people into zero-excess policies that quietly overcharge them.
## How does insurance excess actually work?Your excess is the amount you pay toward any claim before the insurer pays the rest, and on UK policies it comes in two parts. The compulsory excess is set by the insurer based on your risk profile - your age, licence history, car group, postcode - and cannot be changed. The voluntary excess is the part you choose when you buy the policy, and it is the lever this article is about.
The two add together. With a £150 compulsory excess and a £250 voluntary excess, a £2,000 repair bill costs you £400 and the insurer £1,600. Three details matter for the maths later:
Car insurance itself is a legal requirement (the government's vehicle insurance rules set the minimum), but every pound of excess above the compulsory level is a choice. Home insurance works the same way, with the wrinkle that some perils - escape of water is the classic - often carry their own higher compulsory excess.
## What does expected value say about your excess?Raising your excess is selling a slice of insurance back to the insurer. The only question is whether they are paying you a fair price for it. Two numbers decide that:
Raising the excess by ΔE costs you an expected p × ΔE per year (the chance of a claim times the extra amount you would pay in that claim). So the deal is worth taking when:
S > p × ΔE
Rearranged, each step up the excess ladder has a break-even claim probability p* = S ÷ ΔE. If your real claim frequency is below p*, the step is positive expected value. If you claim more often than p*, keep the lower excess.
There is a structural reason the maths so often favours the higher excess: premiums are not pure claim costs. As we cover in how insurance companies use probability, every premium carries a loading for expenses, reinsurance and profit on top of expected claims. When you self-insure a £250 slice of risk you can comfortably afford, you keep that loading. It is the same arithmetic that makes extended warranties a poor deal for most buyers.
**The one-line rule** Divide the annual saving by the extra excess. If the result is bigger than your honest yearly claim probability, take the higher excess. ## A worked example: the £250 questionHere is an illustrative quote ladder for one driver, one insurer, identical cover - only the voluntary excess changes. The numbers are invented for the example, but the shape (each step saves less than the last) is exactly what real quote ladders look like:
| Voluntary excess | Annual premium | Saving vs £0 | Saving for this step |
|---|---|---|---|
| £0 | £560 | - | - |
| £250 | £505 | £55 | £55 |
| £500 | £476 | £84 | £29 |
Suppose this driver's honest record is one fault claim every eight years - an annual claim probability of about 12.5%.
Step one, £0 to £250: the break-even probability is £55 ÷ £250 = 22% a year, or a claim every four and a half years. Our driver claims far less often than that, so the step is worth an expected £55 - (0.125 × £250) = +£23.75 a year. Take it.
Step two, £250 to £500: the break-even drops to £29 ÷ £250 = 11.6% a year. Our driver's 12.5% now sits just above it: the step costs an expected £31.25 against a £29 saving, so it is -£2.25 a year. Decline it.
Same driver, same logic, opposite answers on the two steps - which is why the blanket advice you see everywhere ("raise your excess to save money") is only half right. Each step has to justify itself, and the diminishing savings mean there is usually a level where the ladder stops paying. Your own quote ladder and your own claim history are the real inputs: rerun the two divisions with those and the answer falls out.
## When is a higher excess the wrong choice?Expected value is the right framework, but three situations override a +EV step up the ladder:
One caveat cuts the other way: because the excess applies per claim, a year with two fault claims doubles the pain. If that risk is material for your household, weight p upward rather than using a single-claim frequency.
## Why do we over-insure small risks?If the maths so often favours a higher excess, why do zero-excess policies and excess-waiver add-ons sell so well? Because a certain £55 saving against an uncertain £250 cost feels like a bad trade. Losses loom roughly twice as large as equivalent gains in human judgement - that is loss aversion - and insurers price that feeling, not just the risk. Paying a reliable premium to avoid an affordable, uncertain loss is negative expected value by construction; the product exists because the discomfort is real even when the arithmetic is not on its side. (The distinction between sensible caution about ruin and reflexive flinching at any loss is exactly the difference between risk aversion and loss aversion.)
Here is the part the zero-excess pitch never mentions: you are self-insuring small losses anyway. Claiming £350 on a policy with a £250 excess returns £100 today and typically costs you your no-claims discount (NCD - the premium discount insurers grant for claim-free years) plus higher quotes at renewal for several years. For most drivers that trade is clearly negative, so small claims go unclaimed regardless of the excess level. A low excess buys much less protection than it appears to - you pay every year for cover you would rationally never invoke.
Insurance earns its keep against ruin: the written-off car, the rebuilt kitchen, the liability claim. It is a poor tool for £250 annoyances. Buy the catastrophe cover generously, self-insure the small stuff deliberately, and let the premium saving compound in your own account rather than the insurer's.
## How to choose your excess in four steps 1. **Pull the quote ladder** - Get quotes from the same insurer at three or four voluntary excess levels (£0, £250, £500, £750). Comparison sites let you re-run with only the excess changed. Note the saving for each step, not just the total. 2. **Estimate your honest claim probability** - Count household fault claims over the last ten years and divide. One claim in eight years is about 12.5% a year. Be honest rather than optimistic - the calculation is only as good as this number. 3. **Compute the break-even for each step** - For every step up the ladder, divide the step's annual saving by the extra excess. Take each step where the result exceeds your claim probability; stop at the first step where it does not. 4. **Apply the affordability cap** - Whatever the maths says, never set a combined excess above what you could pay tomorrow from your emergency fund without borrowing. A +EV bet you cannot cover is still a bad bet. ## Frequently asked questions **Q: What is the difference between compulsory and voluntary excess?** The compulsory excess is fixed by the insurer based on your risk profile and cannot be changed. The voluntary excess is chosen by you when you take out the policy. When you claim, the two add together and you pay the combined amount before the insurer pays the rest. **Q: Does increasing the voluntary excess always reduce the premium?** Almost always, but with diminishing returns - each step up the ladder saves less than the one before, and some insurers flatten out entirely at higher levels. Always price the specific steps rather than assuming a bigger excess keeps paying. **Q: Do I pay the excess if the accident was not my fault?** Usually you pay it upfront and your insurer recovers it from the at-fault driver's insurer, refunding you once liability is settled. Some policies waive it immediately when fault is clear. The excess decision therefore mostly concerns claims where you are at fault or the loss cannot be recovered. **Q: Is a zero-excess policy worth it?** Rarely on expected value. You pay a certain, repeated premium loading to avoid an uncertain, affordable cost - and because small claims damage your no-claims discount, you would often not claim for small amounts anyway. It can make sense if your claim frequency is genuinely high or an unexpected excess would force you to borrow. **Q: Should my home insurance excess follow the same rule?** Yes - the same break-even division applies, just with home claim frequencies, which for most households are lower than motor. Watch the per-peril compulsory excesses (escape of water is commonly higher) since your voluntary excess stacks on top of those. ## SourcesThe expected value of any insurance-like product is simple: what it pays out, times the probability it pays out, minus what it costs. For an extended warranty:
EV = (probability of a covered failure x average benefit) - premium
Everything the salesperson says at the till is an attempt to make you feel the first term and forget the second. So put numbers on it. Which? found UK extended warranties cost an average of £84.80 a year, and a big-brand washing machine plan from AO starts at £6.99 a month - £419 if held for the full five years, close to the price of the machine it covers. Currys' multi-year Care & Repair plans run from £90 to £250 depending on the product.
Against that premium, the benefit is a repair (typically £120-£220 for a domestic appliance call-out plus parts) or, if it can't be fixed, a replacement voucher. The failure probability is the number nobody quotes at the till - and it is doing all the work.
## How do the numbers actually fall out?Take a £450 washing machine with the usual one-year manufacturer warranty, and a five-year plan for a one-off £190 (mid-range for the plans above). The plan's incremental cover is years 2-5. Suppose the machine has some probability p of a breakdown in that window, an average repair costs £170, and total write-offs get a £450 replacement.
The breakeven failure probability if claims are repairs: £190 / £170 = 112%. Literally impossible - even a guaranteed single repair would not return the premium. If every claim were a full £450 replacement: £190 / £450 = 42%. The plan only breaks even if two in five machines die completely within five years.
Now plug in a generous real-world failure belief - say p = 20% in years 2-5, with claims split between repair and replacement (average benefit £250):
EV = 0.20 x £250 - £190 = -£140
You pay £190 for something worth £50 in expectation. Even doubling the failure rate to 40% leaves the cover £90 underwater. Retailers are not confused about this: warranty margins are among the best in electrical retail precisely because the premium is set several multiples above expected claims - a textbook negative expected value product from the buyer's side.
**The one-line test** Before saying yes at the till, estimate: what are the odds this breaks after the free warranty but within the plan, and what would the repair cost? Multiply them. If the premium is bigger than that number (it almost always is, usually by 3-5x), decline. If it is smaller, you have found the rare exception - buy it. ## When is an extended warranty actually worth buying?The multiplication can flip positive when three things stack: a genuinely high failure rate, expensive repairs, and a cheap, long plan. A £35 five-year plan on a £1,200 appliance with a 25% failure probability and £400 average repairs has EV = 0.25 x £400 - £35 = +£65. Deals like that exist - usually manufacturer promotions (free or token-price extended cover for registering the product) rather than till-point retail plans.
Two other cases deserve a mention. First, genuinely fragile, heavily-used, costly-to-fix items (a laptop that commutes daily) can carry failure-x-cost numbers high enough to justify some price of cover - though accidental-damage insurance and the till-point warranty are different products, and the till price is rarely the cheap way to buy it. Second, if a surprise £400 bill would be genuinely unaffordable rather than merely annoying, insurance can be rational even at negative EV - you are buying protection from ruin, not maximising expectation. A washing machine rarely qualifies.
## Why do we keep buying -EV cover? Loss aversion makes a possible £450 loss feel roughly twice as painful as a £450 gain feels good, so we overpay to delete small, survivable risks. Warranty pricing is that bias with a barcode. The tell: nobody insures their £40 kettle through a broker, but plenty of people say yes to £3.50 a month for it at the checkout, where the pain of imagining the loss is freshest. The rational frame is self-insurance. Put the £7 a month you would have spent on cover into a repairs pot instead. Across all your appliances and gadgets, some will break and most will not - and because you are diversified across many items, your pot almost certainly ends ahead of the premiums. You keep the float; the retailer was keeping it before. Insurance belongs on risks that could ruin you (house, liability, income), not on anything you could rebuy with a grimace. ## What free protection do you already have?UK law quietly covers a large slice of what extended warranties sell back to you:
An extended warranty's incremental value is only what it pays beyond all of the above - mostly wear-and-tear failures that are not inherent faults, plus convenience (no arguing, one phone line). Convenience is worth something. It is rarely worth 3-5x the expected claims.
## Frequently asked questions **Q: Are extended warranties ever worth it on cheap items?** Practically never. On a £60 toaster, any plan priced above a few pounds exceeds the item's expected repair value - and you would just replace it anyway. The cheaper and more replaceable the item, the more certainly the cover is negative EV. Self-insure everything you could rebuy without wincing. **Q: Is the Consumer Rights Act really as good as a warranty?** It covers inherent faults (manufacturing defects that were present at purchase) for up to 6 years in England and Wales, with the burden of proof shifting to you after 6 months. It does not cover wear and tear or accidental damage. So it overlaps heavily with what warranties actually pay out on, but it is not identical - the gap is the honest part of what a warranty sells. **Q: What failure rate should I assume for appliances?** Treat it as your own estimate, not gospel: reliability surveys consistently put major-appliance breakdown rates for years 2-5 well below the 40%+ a typical plan needs to break even. Our model shows the conclusion barely moves between 10% and 30%, so you don't need a precise figure - the premium-to-benefit ratio does the deciding. **Q: Does paying monthly change the maths?** It usually makes it worse: £6.99 a month feels smaller than £419, but that is what it totals over five years - and monthly plans quietly continue past the point where the appliance's value justifies any cover at all. If you do buy cover, a fixed-term plan with a known total beats an open-ended direct debit. **Q: What about accidental damage and gadget insurance?** Different product, same test. Phones and laptops have higher incident rates and repair costs, so the failure-x-cost term is genuinely bigger - but premiums are bigger too, and excesses cut the payout. Run the same multiplication with the excess subtracted from the benefit before deciding, and check whether your home contents policy already covers it.Worked figures are our own calculations from the premiums and assumptions stated in the text; they illustrate the arithmetic and are not financial advice.
--- # What Investment Platform Fees Really Cost Over 20 Years Source: https://expectedvalue.co.uk/blog/investment-platform-fees-long-term-cost/ # What Investment Platform Fees Really Cost Over 20 Years A 0.35% platform fee sounds tiny. Compounded over 25 years it can cost £30,000+. The maths of fee drag, and where flat fees beat percentages. **Answer** Percentage platform fees are a guaranteed drag on every year's return. Investing £500/mo from a £20,000 start at 5% real, a 0.35% platform fee costs roughly £32,000 of terminal wealth over 25 years, versus about £19,000-21,000 on flat-fee or capped platforms. Flat fees overtake percentage fees once your pot passes roughly £20,000-£48,000. ## Why does a 0.35% fee cost so much more than it looks?A platform fee quoted as 0.35% a year sounds like a rounding error. The trap is in what it is 0.35% of. It is charged on your whole pot, every year, whether markets rise or fall - and your pot is (hopefully) growing. A fee that costs you £70 on a £20,000 pot costs £875 a year by the time you reach £250,000. Percentage fees are the only bill in your life that automatically inflates with your success.
The second effect is worse. Every pound taken in fees this year is a pound that can never compound for you again. In expected value terms, a percentage fee is a guaranteed negative edge applied to every single year's return: returns are uncertain, but the fee is certain. Reduce a 5% expected real return by 0.57% of all-in charges and you have not lost 0.57% of your final wealth - you have lost 0.57% of the growth rate, compounded for decades. That is a much bigger number.
## What does 25 years of fee drag actually look like?Take a UK investor with a £20,000 starting pot, adding £500 a month, earning 5% a year in real (after-inflation) terms for 25 years, and holding a typical global tracker fund with a 0.22% ongoing charge (OCF). We modelled the same journey on three real fee structures, using the rates in force as of July 2026:
With no fees at all, this saver finishes with £361,161. Here is what each route actually delivers:
| Item | Terminal pot after 25 years | Lifetime fees paid | Total drag vs a free ride | Drag as % of final wealth | | --- | --- | --- | --- | --- | | Hargreaves Lansdown (0.35% tiered) | £329,180 | £20,878 | £31,981 | 8.9% | | Vanguard Investor (0.15%, capped) | £340,182 | £13,655 | £20,979 | 5.8% | | interactive investor (flat fee) | £342,510 | £11,841 | £18,652 | 5.2% |Every figure includes the 0.22% fund OCF, which on its own accounts for £12,817 of drag - more on that below. The gap between the dearest and cheapest platform here is £13,330 of terminal wealth, for holding identical investments with identical returns. The only difference is the fee structure.
Note that the drag is not the same as the fees paid. HL collects £20,878 in fees, but the saver ends £31,981 behind the frictionless benchmark: the missing £11,103 is growth those fee payments would have earned had they stayed invested. That is the compounding penalty working against you.
**The 1% rule of thumb** Run the same model with a 1% all-in annual charge (a typical actively-managed fund on a percentage platform) over 30 years and the saver loses 17.7% of their terminal wealth to costs - £51,105 gone from a £288,306 pot. A single percentage point, applied for long enough, quietly consumes a sixth of the outcome. ## When do flat fees beat percentage fees? The crossover logic is simple: a flat fee is a fixed number, a percentage fee is a line that rises with your pot. Divide the annual flat cost by the percentage rate and you get the pot size where the two lines cross. Above it, flat wins - and keeps winning by more every year your pot grows. Here is the annual platform cost (excluding fund OCF) at different pot sizes: | Item | £10,000 | £32,000 | £50,000 | £100,000 | £250,000 | £500,000 | | --- | --- | --- | --- | --- | --- | --- | | HL (0.35% tiered) | £35 | £112 | £175 | £350 | £875 | £1,500 | | Vanguard (0.15% capped) | £48* | £48 | £75 | £150 | £375 | £375 | | ii (flat Core/Plus) | £72 | £72 | £72 | £72 | £180 | £180 |*Vanguard's £4/month minimum applies below £32,000.
The crossovers fall out directly:
Notice the honest wrinkle at the small end: on a £10,000 pot, HL's percentage fee (£35) is actually the cheapest of the three. Percentage fees are genuinely fine for small pots. The problem is that they do not stay fine, and most people never revisit the decision as their pot grows.
## Which fee structure wins at your pot size?As of July 2026, for a fund-based portfolio on the three structures modelled here:
The general principle: re-run the crossover check every time your pot roughly doubles. A platform choice that was right at £15,000 is often costing you three figures a year by £80,000 - a classic case of a decision that was positive expected value when made, silently turning negative because the inputs changed.
## Does the fund fee matter more than the platform fee? In our 25-year model the 0.22% fund OCF alone cost £12,817 - comparable to the entire platform bill on the cheaper routes. Unlike platform fees, the OCF has no cap, no flat-fee alternative and no crossover point: it is a pure percentage on the whole pot, forever. That has two practical consequences. First, the gap between a 0.12% developed-world tracker and a 0.50% (or 1%+) active fund dwarfs most platform-fee differences - if you are optimising costs, the fund line usually deserves attention before the platform line. Second, quoted 'all-in' comparisons that bundle the two can hide which lever you actually control. Keep them separate: platform structure is a crossover decision, fund OCF is a minimisation decision. ## What about dealing charges and zero-fee platforms? Flat platform fees often come with per-trade dealing charges, and percentage platforms are adding them too: from March 2026 HL charges £1.95 per fund deal unless you invest via a regular monthly plan. A £500/month saver dealing ad-hoc would add roughly £23 a year; set up the regular-investing option and it is £0. ii's plans include monthly trades on the Core plan and more on Plus, so a monthly saver typically pays nothing extra either. There is also a newer generation of platforms with no platform fee at all - Trading 212 and InvestEngine both charge £0 for a DIY stocks and shares ISA. The catch is range and packaging rather than headline cost. We compare them directly in Vanguard vs InvestEngine and Trading 212 vs eToro if a zero-fee route suits your portfolio. ## Why are fees the one number worth obsessing over?You cannot choose next year's market return. You can choose your cost structure. Returns arrive as a distribution - some sequences kind, some brutal, and as we cover in ergodicity, you only get to live one path through them. Fees are different: they are deterministic. Cutting 0.3% of annual drag raises your growth rate on every possible path, good or bad, with zero risk attached.
That is what makes fee optimisation the rare free lunch in investing: a guaranteed improvement in expected value with no offsetting downside. Almost everything else you can do to a portfolio trades expected return against risk. This does not.
**The one-line summary** Percentage fees rent your pot; flat fees rent the service. While your pot is small, renting the pot is cheaper. From roughly £20,000-£48,000 upwards the arithmetic flips - and the longer you leave the switch, the more terminal wealth the difference quietly consumes. ## Frequently asked questions **Q: Is a 0.35% platform fee expensive?** Not at small pot sizes - on £10,000 it is £35 a year, cheaper than most flat-fee plans. It becomes expensive as the pot grows: £350 a year at £100,000 and £875 at £250,000, versus £72-£180 on a flat structure. The fee rate never changed; the base it applies to did. **Q: Are flat-fee platforms always cheaper for large portfolios?** For the structures compared here, yes above roughly £120,000 - even Vanguard's capped £375 stays above ii's £179.88. But always include dealing charges: a frequent trader on a flat plan with per-trade fees can pay more than the headline suggests. For a monthly regular investor, the flat fee usually dominates. **Q: Do fees really matter if my returns are strong?** More than ever, because percentage fees scale with the pot your strong returns create. The drag compounds exactly like returns do - our model shows a 1% all-in charge consuming 17.7% of terminal wealth over 30 years regardless of whether the underlying return was strong or weak. **Q: Should I switch platforms purely to save on fees?** Check three things first: exit or transfer-out charges on your current platform (many are now £0, but not all), whether your funds transfer in-specie or must be sold and rebought, and any cashback offers that offset switching friction. If the annual saving exceeds any one-off cost within a year or two, the switch is positive expected value. **Q: What assumptions does the model use?** £20,000 starting pot, £500/month contributions, 5% annual real return, 25 years, 0.22% fund OCF, and platform fee schedules in force as of July 2026 (HL 0.35% tiered, Vanguard 0.15% capped at £375 with the £4/month small-account minimum, ii Core £5.99/month to £100,000 then Plus £14.99/month). Fees are charged annually on the average pot in the model; real platforms charge monthly, which changes the totals by a rounding error.All worked examples and terminal-wealth figures are our own calculations from the fee schedules above; they are illustrations of the arithmetic, not financial advice or a personal recommendation.
--- # Premium Bonds: What's the Real Expected Value? Source: https://expectedvalue.co.uk/blog/premium-bonds-expected-value/ # Premium Bonds: What's the Real Expected Value? Premium Bonds advertise 3.8%, but that average hides a skewed reality. Here's the real expected (and typical) return for £1k, £10k and £50k holdings. **Answer** Premium Bonds pay a headline 3.8% prize fund rate (from the July 2026 draw), but that is an average dragged up by rare big prizes. The typical saver wins less, and with £1,000 invested you most often win nothing. Only holdings near the £50,000 maximum get close to the headline rate.The real expected value of Premium Bonds is the gap between a tempting headline and what most savers actually receive. NS&I (National Savings and Investments, the government-backed savings provider) advertises a prize fund rate of 3.8% from the July 2026 draw, with odds of 22,000 to 1 on each £1 Bond. That 3.8% is a genuine average, but it is one of the most misleading averages in personal finance, because the prize distribution behind it is wildly lopsided.
Two savers can hold identical bonds for a year and have completely different experiences: one wins nothing, the other lands a £100 prize, and a vanishingly rare third scoops £1 million. Understanding why the typical outcome sits so far below the average is the whole game, and it is a textbook lesson in thinking about distributions rather than headlines.
## How do Premium Bonds actually work?You buy bonds in £1 units, from a minimum of £25 up to a maximum holding of £50,000. Every £1 bond is entered into a monthly prize draw with odds of 22,000 to 1. Instead of paying interest, NS&I pools what would have been the interest and pays it out as tax-free prizes ranging from £25 to £1 million, with two £1 million jackpots each month and well over six million prizes in total.
The prize fund rate, currently 3.8%, describes the size of that pool relative to the total money invested. If every pound earned its mathematically fair share, a holding would return 3.8% a year. The catch is that prizes are lumpy and indivisible: you cannot win 3.8% of a pound, you either win a £25 prize or you win nothing. That lumpiness is where the average and the typical experience part company. Bonds are also fully backed by HM Treasury, so your capital is completely safe.
## What is the real expected value of Premium Bonds?The expected value of a Premium Bonds holding is, by construction, the prize fund rate: 3.8% a year. But it helps to see where that comes from. With odds of 22,000 to 1 per bond per month, each £1 has roughly a 1-in-1,833 chance of winning something in a year. The average prize works out at about £70 (the prize pool divided by the number of prizes).
Multiply expected prizes by the average prize and every holding lands back at 3.8%. That is the expected value. The problem is that expected value describes the long-run average across millions of bonds, not what any single saver is likely to see in a given year.
## Why does the typical saver win less than the headline rate?Because the prize distribution is extremely right-skewed. The overwhelming majority of prizes are the smallest, £25, while two enormous £1 million prizes and a scattering of mid-sized ones pull the average upwards. A handful of huge winners collect a large slice of the pool, leaving everyone else below the mean.
This is the difference between the mean and the median. The odds tell you a win is rare for a small holding; the skew tells you that even when you do win, it is almost always the £25 minimum. So the 50th-percentile saver - the one with average luck - earns noticeably less than 3.8%, because the average is being held aloft by jackpot winners they will almost certainly never join. Quoting the mean as if it were the result any single saver should expect is the single most common way Premium Bonds are misunderstood.
## What can you realistically expect from £1,000, £10,000 or £50,000?With average (median) luck, the picture looks very different from the headline:
The pattern is the whole lesson: the more bonds you hold, the closer your actual return creeps to the expected value. Small holders are playing a near-lottery; only large holders experience anything like the advertised rate.
## Do the tax-free prizes change the maths?Sometimes, and this is where Premium Bonds earn their keep. All prizes are tax-free, which matters depending on your Personal Savings Allowance. Basic-rate taxpayers can already earn £1,000 of ordinary savings interest tax-free each year, higher-rate taxpayers £500, and additional-rate taxpayers nothing.
For a basic-rate taxpayer with modest savings, the tax-free wrapper adds nothing they did not already have, and a normal easy-access account paying around 4.5% beats Premium Bonds on both expected return and certainty. But for a higher-rate taxpayer who has used their £500 allowance, a tax-free 3.8% is equivalent to roughly 6.3% of taxable interest; for an additional-rate taxpayer with no allowance left, it is closer to 6.9%. That gross-equivalent return is genuinely competitive - but only for someone holding near the £50,000 maximum, where their actual return approaches the mean rather than languishing below it.
## So are Premium Bonds worth it?For a basic-rate taxpayer with a small holding, Premium Bonds are a weak expected-value choice: a lower average return than a standard savings account, and a median return that is frequently zero. The appeal there is purely the fun of the draw and the safety, not the maths.
For a higher or additional-rate taxpayer who has exhausted their Personal Savings Allowance and can hold close to £50,000, the case is genuinely reasonable. The tax-free gross-equivalent return is competitive, the law of large numbers makes the outcome reliable, the money is fully government-backed with no upper protection limit, and the tiny chance of a life-changing prize is a free lottery ticket on top. The honest verdict is that Premium Bonds are not a scam and not a slam-dunk: they are a tax-and-scale play dressed up as a game of chance.
**The one-line takeaway** A headline average is only your expected experience when you make the bet many times over. With £1,000 in Premium Bonds you make it once a month and the typical result is nothing; with £50,000 you make it fifty times over and the average finally shows up. ## Frequently asked questions **Q: What is the expected return on Premium Bonds?** The expected (average) return equals the prize fund rate, which is 3.8% a year from the July 2026 draw. But because the prize distribution is heavily skewed by rare large prizes, the typical saver wins less than this, and small holdings often win nothing in a given year. **Q: How much would you realistically win with £1,000 in Premium Bonds?** With £1,000 you can expect roughly one prize every two years, almost always the £25 minimum. In a typical year you win either nothing or £25, giving a realistic return of about 0% to 2.5%, well below the 3.8% headline rate. **Q: Are Premium Bonds better than a savings account?** For a basic-rate taxpayer, usually not: a top easy-access account paying around 4.5% offers a higher and guaranteed return. Premium Bonds mainly suit higher and additional-rate taxpayers who have used their Personal Savings Allowance and hold close to the £50,000 maximum. **Q: Are Premium Bond prizes tax-free?** Yes. All prizes are free of UK income tax and capital gains tax. This is most valuable to higher and additional-rate taxpayers who have already used their Personal Savings Allowance on other interest. **Q: Why do small Premium Bond holdings keep winning nothing?** With odds of 22,000 to 1 per bond and most prizes being the £25 minimum, a small holding simply does not buy enough draws to win often. Only as your holding approaches £50,000 does the law of large numbers bring your results close to the average rate. ## Related guides --- # UK FIRE Movement 2026: The Complete Roadmap Source: https://expectedvalue.co.uk/blog/uk-fire-movement-2026/ # UK FIRE Movement 2026: The Complete Roadmap A UK FIRE roadmap for 2026: the 4% rule, your real number, the tax-wrapper order, and the ISA bridge to pension access age. Honest maths, no hype. **Answer** FIRE (Financial Independence, Retire Early) means saving and investing enough that your portfolio can cover your living costs indefinitely. The rough target is 25 times your annual spending. In the UK the key twist is timing: you bridge the years before pension access age (57 from 2028) with ISAs.The FIRE movement (Financial Independence, Retire Early - a strategy of saving aggressively to escape paid work decades early) has a simple core and a UK-specific catch. The core is a single number: save roughly 25 times your annual spending, invest it sensibly, and you can in principle live off the portfolio forever. The catch is that the UK tax system rewards a particular order of saving, and locks much of your money away until your late fifties.
This guide lays out the 2026 numbers, the account sequence that makes UK FIRE work, and a concrete roadmap from a standing start. It is written for someone who wants the maths to be honest rather than motivational: financial independence is mostly arithmetic, and the arithmetic is more demanding, and more achievable, than the headlines suggest.
## What is the FIRE movement?FIRE is the idea that financial independence - having enough invested that work becomes optional - can be reached far earlier than the traditional retirement age, often in your forties or even thirties. The lever is your savings rate: the percentage of your take-home pay you keep rather than spend. A high savings rate does double work, because every pound saved both grows your pot and lowers the lifestyle you need to fund.
The movement has grown several flavours worth knowing:
All of them rest on the same engine: spend less than you earn, invest the gap, and let time and compounding do the heavy lifting.
## How much money do you need to retire early?The standard answer is the 4% rule: you can withdraw 4% of your portfolio's starting value in year one, increase that amount with inflation each year, and have a high chance of the money lasting 30 years. Flip it around and the rule gives you a target: your FIRE number is your annual spending multiplied by 25. Spend £25,000 a year and you need roughly £625,000; spend £40,000 and you need £1,000,000.
The 4% rule comes from US market history and a 30-year retirement. Early retirees face a longer horizon - potentially 40 to 50 years - so many in the UK community adopt a more cautious 3.5% withdrawal rate, which lifts the multiple to roughly 28 to 29 times spending. The difference matters: at £25,000 a year, moving from 4% to 3.5% raises the target from £625,000 to about £715,000. Neither figure is a guarantee. The rule is a probability statement about an uncertain future, not a promise, which is why flexibility about how much you withdraw in a bad year is worth more than any single percentage.
## Why does UK FIRE need a different playbook?Most FIRE writing is American, and three big differences make UK FIRE its own game. First, healthcare: the NHS removes the enormous medical-insurance cost that dominates US early-retirement plans. Second, the state pension: the full new State Pension is £241.30 a week in 2026/27, around £12,548 a year, according to gov.uk, and it arrives from State Pension age (currently 66, rising to 67 and then 68). That guaranteed income late in life genuinely shrinks the private pot you need.
Third, and most importantly, access age. Money inside a pension is locked until the normal minimum pension age, which rises from 55 to 57 on 6 April 2028. ISAs, by contrast, can be accessed at any age. So a UK FIRE plan is really two plans stitched together: a pension pot you cannot touch until your late fifties, and a separate, accessible pot that has to carry you from the day you stop working until that pension unlocks.
## Which accounts should you fill, and in what order?Tax wrappers are not interchangeable, and the order you fill them changes how much you keep. A sensible default priority for most UK FIRE savers looks like this:
The exact split depends on your retirement age. The earlier you want to stop, the more weight shifts towards the ISA, because you need more money you can actually reach before 57.
## How do you start in your first year?The path from interested to independent begins with a few unglamorous moves that matter more than any fund choice. In rough order:
None of this requires a high income to begin. The savings rate, not the salary, is what starts the clock - and the clock rewards starting badly today over starting perfectly next year.
## What is the ISA bridge, and why is it the heart of UK FIRE?Imagine you reach financial independence at 45 but cannot touch your pension until 57. That is a twelve-year gap you have to fund entirely from accessible savings. This is the ISA bridge: a pot, mostly held in a stocks and shares ISA, sized to cover your spending across the years between early retirement and pension access age.
The bridge reframes the whole plan. It is not enough to have a big total net worth; you need the right money in the right place at the right time. A saver with £700,000 locked in a pension and nothing accessible cannot retire at 45, however large the number looks. Practically, that means an early retiree front-loads the ISA in their thirties and early forties, then leans more on pension contributions once the bridge is funded. Get the sequencing wrong and you can be a paper millionaire who still cannot afford to stop working.
## A concrete roadmap: from £0 to financial independencePut numbers to it. Suppose you want £25,000 a year in retirement, giving a FIRE number of roughly £625,000 at a 4% withdrawal rate. The dominant variable is not your investment returns; it is your savings rate. On common assumptions of around 5% real annual returns, the years-to-independence look broadly like this:
The reason the savings rate dominates is that it compounds from both ends at once: a higher rate builds the pot faster and shrinks the pot you need. A worked path might look like this. You start at 30 earning a middle income and commit to a 50% savings rate. Around age 47 you hit your number. But because pension access is 57, you spend your thirties and forties deliberately overweighting the ISA, building roughly twelve years of spending - about £300,000 - in accessible accounts, while still capturing your employer pension match. At 47 you stop full-time work, draw on the ISA bridge until 57, then let the pension and, later, the State Pension take over.
## Where do UK FIRE plans most often go wrong?The failures tend to rhyme. The most common is ignoring the bridge: pouring everything into a pension for the tax relief, then realising the money is locked until 57 and there is nothing to live on in the meantime. The second is lifestyle inflation - letting spending creep up with every pay rise, which quietly raises your FIRE number and pushes the finish line further away even as your income grows.
A third mistake is holding too much in cash out of fear. Over a multi-decade horizon, a large cash pile is almost certain to lose purchasing power to inflation, and after April 2027 idle cash inside a stocks and shares ISA will also face a 22% charge on its interest. A fourth is forgetting the State Pension entirely and over-saving, working years longer than the maths required. The last, and most human, is treating a spreadsheet forecast as destiny: building a plan around one assumed rate of return and then panicking when reality refuses to follow the line. A good plan survives a range of outcomes and expects to be revised.
## What could derail your plan?Two risks deserve respect. The first is sequence-of-returns risk: a market crash in the first few years of retirement is far more damaging than the same crash later, because you are selling assets to live while prices are low. The 4% rule already accounts for average bad luck, but the defence that works best in practice is flexibility - trimming withdrawals in down years rather than mechanically taking the inflation-adjusted figure.
The second is over-precise planning. A roadmap built on a single assumed return is a forecast dressed as a certainty. Treat your number as a moving target, revisit it as your spending and the markets change, and favour decisions that stay good across a range of outcomes rather than ones that only work if everything goes to plan. Financial independence is reached on the strength of your savings rate and your behaviour in bad years, far more than on picking the perfect withdrawal percentage.
**Your savings rate is the lever that matters** Chasing an extra 1% of investment return is tempting, but it moves your timeline far less than raising your savings rate. Cutting spending does double duty: it frees up money to invest and lowers the total you ever need. Focus your energy there before optimising fund choices to the third decimal place. ## Frequently asked questions **Q: How much do I need to retire early in the UK?** As a rule of thumb, multiply your target annual spending by 25 for a 4% withdrawal rate, or by about 28 to 29 for a more cautious 3.5% rate. Spending of £25,000 a year implies a FIRE number of roughly £625,000 to £715,000, before allowing for the State Pension later in life. **Q: At what age can I access my pension in the UK?** The normal minimum pension age is 55, rising to 57 from 6 April 2028. ISAs can be accessed at any age, which is why early retirees rely on them to bridge the years before their pension unlocks. **Q: Should I prioritise a pension or an ISA for FIRE?** Capture any employer pension match first, since it is free money. After that, early retirees usually prioritise a stocks and shares ISA because it is accessible before 57, then return to pension contributions for the tax relief once the bridge is funded. **Q: Is the 4% rule safe for a 40-year retirement?** The 4% rule was tested on 30-year retirements. For a longer early-retirement horizon, many UK savers use 3.5% to add a margin of safety. Neither is a guarantee; staying flexible about withdrawals in poor years matters more than the exact percentage. **Q: Does the State Pension count towards FIRE?** Yes. The full new State Pension is about £12,548 a year in 2026/27 and reduces the private income your portfolio must generate, though it only starts from State Pension age (66, rising to 67 and 68). Many plans treat it as a top-up that arrives in later life rather than relying on it early. ## Related guides --- # Stocks and Shares ISA Limit 2026: Rules and 2027 Changes Source: https://expectedvalue.co.uk/blog/stocks-and-shares-isa-limit-2026/ # Stocks and Shares ISA Limit 2026: Rules and 2027 Changes The 2026/27 stocks and shares ISA limit is £20,000. Here's how the allowance works, what the April 2027 cash ISA cut means, and how to use it. **Answer** For the 2026/27 tax year you can put up to £20,000 into a stocks and shares ISA, the same as the total ISA allowance. The stocks and shares limit stays at £20,000 from April 2027, but the cash ISA limit falls to £12,000 for under-65s.The stocks and shares ISA limit for the 2026/27 tax year is £20,000. An ISA (Individual Savings Account, a wrapper that shelters your money from UK tax) shares one annual allowance across all its forms, and a stocks and shares ISA can take the whole £20,000 if you want it to. That headline figure has not moved, but the rules around it are about to change more than they have in years.
At the Autumn Budget on 26 November 2025, the Chancellor confirmed the biggest reshaping of the ISA system since 2017. Almost every part of it points in the same direction: it pushes savers out of cash and towards investing. If you hold a stocks and shares ISA, or you are deciding where this year's allowance should go, the changes coming in April 2027 matter to you now, because the rational move is to plan around them early.
## What is the stocks and shares ISA limit for 2026/27?For the tax year running from 6 April 2026 to 5 April 2027, you can pay up to £20,000 into ISAs. There is no separate, lower cap on the stocks and shares portion, so you can direct the full £20,000 into shares, funds, investment trusts and bonds held inside a stocks and shares ISA.
This is one combined allowance, not a separate £20,000 for each ISA type. If you put £8,000 into a cash ISA, you have £12,000 left for a stocks and shares ISA. Any growth, dividends or interest earned inside the wrapper is free of UK income tax and capital gains tax, and you never declare ISA holdings on a tax return. That tax shelter is the entire point, and it is why using the allowance each year is one of the few genuinely free wins in personal finance.
## How does the £20,000 allowance work across ISA types?The £20,000 is split however you like across four adult ISA types: cash ISAs, stocks and shares ISAs, innovative finance ISAs (which hold peer-to-peer loans) and the Lifetime ISA. Two of those carry their own internal caps you need to know about.
So a saver could, in a single year, pay £4,000 into a Lifetime ISA, £16,000 into a stocks and shares ISA, and £9,000 into a child's Junior ISA, and stay fully within the rules.
## What's changing for ISAs from April 2027?The 2026/27 year runs under today's rules. The reforms announced at the Autumn Budget take effect from 6 April 2027, and they target cash, not investments. According to the government's ISA reform factsheet, the key changes are:
The message from the Treasury is blunt: cash ISAs are for shorter horizons and modest balances, and the system is being re-pointed towards investing for the long term.
**Watch your idle cash after 2027** From April 2027, parking large amounts of uninvested cash inside a stocks and shares ISA stops being a free hiding place: the interest it earns faces a 22% charge. If you hold cash for an upcoming purchase, a cash ISA or an instant-access account may serve you better than leaving it idle in an investment wrapper. ## Why the 2027 reforms reward investingLook at the reforms through an expected-value lens and they stop looking like a tax grab and start looking like a nudge towards the higher-expected-return choice. Over long horizons, a globally diversified equity fund has historically delivered a meaningfully higher average annual return than cash, even after inflation. Cash protects the number on your statement but quietly loses purchasing power; equities are volatile year to year but carry a positive long-run expected return.
Holding a large cash balance for decades is, in expected-value terms, choosing the option with the lower expected outcome to avoid short-term discomfort. That is a textbook case of loss aversion overriding the maths. None of this means cash is wrong; an emergency fund and money you need within a few years belong in cash. But for money you will not touch for ten years or more, the reforms simply make the tax system agree with what the long-run numbers already said. Investing is one of the clearest places ordinary savers can find positive expected value.
## How should you split this year's £20,000 allowance?There is no single right answer, but a horizon-based rule of thumb works for most people. Money you might need within three years, such as an emergency fund, a house deposit you are about to use, or a wedding next summer, belongs in cash where its value is certain. A cash ISA or a high-interest easy-access account is the sensible home for it.
Money you will not touch for ten years or more is where a stocks and shares ISA earns its keep. Over that kind of horizon the higher long-run expected return of a diversified equity fund usually outweighs the short-term wobbles, and the wrapper means none of the growth is ever taxed. If you are saving for a first home and you are under 40, a Lifetime ISA bolted on top adds a 25% government bonus to the first £4,000 each year, a guaranteed uplift no cash account can match.
A worked example shows why the wrapper matters. Suppose £20,000 grows at an illustrative 6% a year. Inside an ISA, all of that growth compounds untouched. In a taxable account, the same returns could be chipped away each year by dividend tax above the £500 allowance and capital gains tax above the £3,000 exemption. Over a couple of decades, sheltering the money rather than leaving it exposed can be worth several thousand pounds: the kind of quiet, compounding edge that defines a positive-expected-value decision.
## When is the ISA deadline, and does unused allowance roll over?The ISA year ends at midnight on 5 April, and the new £20,000 allowance starts the next day, on 6 April. Crucially, the allowance does not roll over. If you only use £5,000 of this year's £20,000, the unused £15,000 is gone for good on 6 April; you do not start next year with £35,000 to play with.
That use-it-or-lose-it design is why many investors drip-feed money in across the year and then top up before the April deadline if they have spare cash. Leaving everything to the final week works, but it risks missing the cut-off because of a bank transfer delay or a provider's processing time. Setting up a monthly standing order removes that risk entirely.
## Can I pay into more than one stocks and shares ISA?Yes. Since 6 April 2024 you can open and pay into more than one ISA of the same type in the same tax year. That means you can hold and contribute to two different stocks and shares ISAs at once, perhaps to try a new platform, as long as your total contributions across all ISAs stay within the £20,000 limit.
The Lifetime ISA is the exception: you can still only pay into one Lifetime ISA per year. You can also now make partial transfers of money you have paid in during the current year, rather than being forced to move the whole lot. For most people, fewer, lower-cost accounts are still easier to manage, but the flexibility is there if you need it.
## What is 'bed and ISA' and is it worth it?"Bed and ISA" is the process of selling investments you hold outside an ISA and immediately rebuying them inside one, using that year's allowance. It moves existing holdings into the tax shelter in a single, broker-assisted transaction.
It has become more valuable as the tax-free allowances outside ISAs have shrunk. The capital gains tax annual exempt amount is now just £3,000, and the dividend allowance is £500, both a fraction of what they were a few years ago. Holdings left in a taxable account can therefore generate a tax bill that an ISA would have wiped out. The one cost to weigh is that selling can crystallise a capital gain in the year you do it, so it is worth spreading larger moves across tax years to stay within that £3,000 exemption.
**Start early, not at the deadline** Investing your allowance in monthly chunks from the start of the tax year gives your money more time in the market and smooths out the price you pay through pound-cost averaging. It also means you are not scrambling in early April, when missing the deadline costs you a full year's allowance you can never get back. ## Frequently asked questions **Q: What is the stocks and shares ISA limit for 2026?** For the 2026/27 tax year (6 April 2026 to 5 April 2027) you can pay up to £20,000 into a stocks and shares ISA. This is the same as the total ISA allowance, which is shared across all ISA types you hold. **Q: Is the stocks and shares ISA allowance changing in 2027?** No. From 6 April 2027 the stocks and shares ISA limit stays at £20,000. It is the cash ISA allowance that falls, to £12,000 for under-65s, while those aged 65 and over keep the £20,000 cash limit. **Q: Will I be taxed on cash held in my stocks and shares ISA?** From April 2027, interest on uninvested cash held inside a stocks and shares ISA or innovative finance ISA faces a 22% charge. Your shares, funds and dividends are unaffected; the charge applies only to idle cash and to all account holders regardless of age or income. **Q: Does unused ISA allowance carry over to next year?** No. The ISA allowance resets every 6 April and any unused portion is lost. You cannot combine this year's unused allowance with next year's. **Q: Can I have a cash ISA and a stocks and shares ISA in the same year?** Yes. You can split your £20,000 allowance across both. For 2026/27 there is no minimum split; from April 2027, an under-65 who wants to use the full £20,000 will need to put at least £8,000 of it into investments rather than cash. ## Related guides --- # Where Can You Actually Find Positive Expected Value? Source: https://expectedvalue.co.uk/blog/where-to-find-positive-expected-value/ # Where Can You Actually Find Positive Expected Value? Almost every bet is negative-EV by design. Here are the few places an ordinary person can genuinely find positive expected value. **Answer** Most everyday bets are negative expected value by design - the operator builds in a margin. Genuine positive EV is rarer but real: bookmaker promotions (matched betting), long-run index investing, skill games where you have a true edge, and insurance you actually need. The hard part is telling a real edge from a good story.If you understand expected value, the first thing it teaches you is humbling: almost everything marketed as a chance to win money is built to lose you money slowly. So the obvious question is the interesting one. If the lottery, the casino and the bookmaker are all negative-EV, where can an ordinary person actually find positive expected value? There are a handful of honest answers, and a lot of stories dressed up to look like them.
## Why almost everything is a negative-EV betCommercial gambling works because the operator never offers fair odds. A roulette wheel pays 35-to-1 on a number that has a 1-in-37 chance, so every spin quietly skims the difference. A bookmaker prices a two-way market so the implied probabilities add up to more than 100% - that surplus is the over-round, and it is the reason the average punter loses over time. A lottery returns less than half of ticket sales as prizes, which is why a ticket is the textbook example of negative expected value.
None of this is bad luck. It is arithmetic, and the law of large numbers guarantees it shows up the more you play. A positive-EV opportunity, then, is one where that margin has been removed, handed to you, or never existed in the first place.
## So where does positive expected value actually hide?Four places, broadly. Each works for a different reason, and each comes with a catch that the marketing tends to skip.
## 1. Bookmaker promotions, and the maths of matched bettingTo win new customers, bookmakers give away free bets and bonuses. A free bet is stake you did not pay for, which means its expected value to you is positive before you even place it - the bookmaker has handed you the margin instead of charging it. The technique people use to capture that value is matched betting: you back an outcome with the bookmaker using the free bet, and simultaneously lay the same outcome on a betting exchange, so that whichever way the event goes, the result largely cancels out and you keep most of the free-bet value as cash.
It is a hedging method, not a system that beats the odds, and it is important to be honest about the catches. It is not a guaranteed income and it is not literally risk-free: exchange commission eats a slice, a mistimed or mismatched bet can cost you real money, the offers eventually run out, and bookmakers routinely restrict or close the accounts of customers who only ever bet on promotions (the slang is being "gubbed"). Done carefully it is a low-variance way to extract a one-off edge from each offer; treated as a money machine it disappoints.
If you want to see the mechanics worked through properly, extraplace.co.uk is a UK resource that walks through the back-and-lay steps step by step.
## 2. Owning the market: the long-run equity premiumThe least glamorous positive-EV opportunity is also the most reliable. Over long horizons, a broad basket of shares has historically returned more than cash or government bonds - the gap is called the equity risk premium, and it is compensation for tolerating volatility along the way. You do not need an edge over other investors to collect it; you only need to own the whole market cheaply and wait, which is exactly what a low-cost index fund does.
The catch is the volatility itself. Positive expected value over twenty years says nothing comforting about any single year, and selling in a panic converts a positive-EV position into a realised loss. This is where expected value and lived experience pull apart, which is the subject of expected value versus expected utility.
## 3. Skill games where the edge is yours, not the house’sIn poker, you are not playing against the house - you are playing against other people, and the house simply takes a cut (the rake). That changes everything: if you are genuinely better than the opposition by more than the rake costs you, your expected value is positive. The same logic applies to sports modelling or any market where a real informational edge exists. The two conditions are strict, though: the edge has to be real and measurable, and it has to be larger than the margin you pay to play.
Even a real edge is not enough on its own. Stake too much of your bankroll on each bet and a normal losing streak can wipe you out before the edge pays off - which is what the Kelly criterion and its safer cousin, fractional Kelly, exist to prevent.
## 4. Insurance: negative EV, positive expected utilityInsurance is deliberately negative expected value - the premium has to cover the insurer’s costs and margin, so on average you pay in more than you get back. Buying it can still be the right decision, because money is not linear. Losing your house is not a thousand times as painful as losing a thousandth of its value; it is catastrophic. Paying a small, certain, negative-EV premium to remove a small chance of a ruinous loss is a textbook case of maximising expected utility rather than expected value. The trick is buying it only where the downside is genuinely unaffordable, and self-insuring the small stuff.
**The catch nobody mentions: variance** A positive edge is necessary but not sufficient. Bet too big relative to your bankroll and variance can ruin you before the edge arrives - try the numbers in the risk-of-ruin calculator. Size matters as much as edge. ## How to tell a real edge from a good storyMost "systems" fail one of four tests. Run any opportunity through them before you risk money:
Availability cascades drive bubbles + panics in markets. Recognising them is one of the most useful investment skills.
## What is an availability cascade?Defined by Kuran + Sunstein (1999):
In markets:
Stage 1 - Initial salient event:
Stage 2 - Media amplification:
Stage 3 - Social transmission:
Stage 4 - Cognitive availability ≠ statistical certainty:
Eventually:
Tech bubble (1995-2000):
2008 housing bubble:
2020-2021 COVID-era retail investing:
2023-2024 AI narrative:
Buying into cascades:
Identifying cascade phases:
Buying after cascades break:
Avoiding mistakes:
Cascade:
Genuine paradigm shift:
How to distinguish (real-time difficult):
Mental accounting helps us budget but also creates systematic financial errors. This guide covers the bias + how to mitigate it.
## What mental accounting isDefinition:
Identified by:
The contradiction with rational economics:
The pattern:
Examples:
Why this happens:
Cost:
The classic puzzle:
Why people do this:
Cost:
Quick test for mental accounting: are you saving money at a lower rate than you're paying on debt? If yes, you're being mentally-accounted-out of wealth. The fungibility test: 'Would the money I have in savings, transferred to pay down debt, save me more than the savings interest forgone?' If yes (almost always), pay the debt. The mental account 'holiday fund' is a psychological construct; the money is fungible + has higher economic value reducing debt.
## Mechanism 3 - categorising by purchase contextThe pattern:
Why this matters:
Application:
Position-level vs portfolio-level tracking:
'Risk capital' vs 'safe capital':
'House money' effect:
Mental accounting isn't entirely irrational - it can serve useful purposes:
The skill:
The conjunction fallacy is the error of estimating that 'A AND B' is more likely than 'A alone' - which is mathematically impossible (P(A AND B) ≤ P(A)). Classic illustration: Kahneman + Tversky's 'Linda problem' showed that people rate 'Linda is a bank teller AND active in feminism' as more probable than 'Linda is a bank teller'. The conjunction is necessarily LESS probable but feels more specific = more believable. Why it matters in business: compound business plans get over-believed because they sound concrete. 'We'll grow at 20% AND raise GBP 5M AND launch in 12 months AND hire 30 people' feels achievable when each individual step is. But the conjunction of all four is far less probable than each. Real-world cases: startup roadmaps that compound 5+ uncertain steps; investment theses that require multiple 'ifs' to hold; project plans that compound technical, hiring, market timing assumptions. Detection signal: when a business plan reads like 'we'll do X AND Y AND Z AND succeed', estimate each conjunct's probability + multiply. The product is usually much lower than gut feel. Mitigation: identify the conjuncts; estimate each independently; multiply (or use Bayesian model); compare to gut feel. Often reveals optimism bias of 2-10x.
The conjunction fallacy distorts business planning by making complex plans feel more probable than their parts. This guide covers detection + mitigation.
## What is the conjunction fallacy?The classical Linda problem (Kahneman + Tversky, 1983):
'Linda is 31 years old, single, outspoken, very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination + social justice + participated in anti-nuclear demonstrations.'
Which is more probable?
~85% of people choose B, the classic result from Tversky and Kahneman's Linda problem.
The math:
Why we get it wrong:
1. Startup roadmaps:
2. Investment theses:
3. Project plans:
4. M&A acquisition plans:
The fallacy is most pronounced when:
Why narratives amplify it:
Why group settings amplify it:
1. Decompose + multiply:
2. Identify weakest link:
3. Premortem analysis:
4. Reference class forecasting:
5. Decision journal:
Important caveat: not all multi-step plans are conjunction-fallacy-prone.
Necessary sequential dependencies:
Strategic optionality:
Mixed dependencies:
Different decisions call for different frameworks. This guide compares the four classical approaches + shows when each is most useful.
## 1. Expected Utility - the workhorseHow it works:
When it dominates:
When it fails:
Practical use:
How it works:
Why it matters:
When it dominates:
When it fails:
How it works:
Why it matters:
When it dominates:
When it fails:
Real-world example:
How it works:
Why it matters:
When it dominates:
When it fails:
Real-world examples:
Use EXPECTED UTILITY when:
Use MINIMAX REGRET when:
Use SATISFICING when:
Use MAXIMAX when:
Real-world decisions often benefit from hybrid framework use:
Why hybrids work:
In poker tournament play, expected value (EV) calculation differs fundamentally from cash games because of the Independent Chip Model (ICM). Cash game EV: every chip is worth the same dollar value; calling a +EV bet (positive chip EV) is +EV in cash. Tournament EV: chip value is NON-linear - each additional chip is worth less than the last. A 10x chip stack isn't worth 10x the prize pool. Why this matters: ICM analysis shows that calls that are +EV in chips can be -EV in actual prize pool money, especially near the money bubble + final-table jumps. The 4 phases of tournament EV thinking: (1) Early stages - chip EV = tournament EV (no ICM impact yet). (2) Money bubble - ICM most aggressive; tighten ranges substantially. (3) Pay jumps + final table - ICM still strong but less than bubble. (4) Heads-up - returns toward chip-EV (winner takes most). Practical applications: push/fold ranges tighten near bubble; calls require more equity vs short stacks (they're risking less); aggression toward medium stacks (incentive to avoid bust). Tools: ICMIZER, ICM Trainer Pro, HoldemResources for offline analysis; in-game ICM intuition develops through practice. Caveat: ICM analysis assumes 'no skill edge' going forward; aggressive players have higher cEV than ICM suggests.
Tournament poker EV is more complex than cash games because chips don't equal money. This guide covers the Independent Chip Model + practical tournament EV decisions.
## The cash vs tournament EV distinctionCash games:
Tournament games:
Implication:
What ICM does:
Worked example - bubble scenario:
Doubling up vs busting at the bubble:
Phase 1 - Early stages (well above money):
Phase 2 - Money bubble + early money:
Phase 3 - Pay jumps + final table:
Phase 4 - Heads-up:
Small stack at the bubble:
Medium stack at the bubble:
Big stack at the bubble:
Practical bubble pattern:
Cash game (no ICM):
Bubble situation (max ICM):
Pay-jump situation (moderate ICM):
Approximation rules:
ICMIZER:
HoldemResources Calculator:
Equilab:
In-game intuition:
Probability-weighted utility is the framework for combining probabilities with subjective value (utility) to compare different options under uncertainty. Why it matters more than 'expected value': expected value treats all pounds (or units) as equal. Probability-weighted utility recognises that the value of GBP 50,000 to your annual income matters far more than the value of an extra GBP 50,000 once you have GBP 5M. The structure: Expected Utility = Σ (probability of outcome × utility of outcome). Choose the option with highest expected utility. Where this beats simple EV thinking: high-stakes single decisions (career, business venture, marriage); decisions involving non-linear value (insurance, downside protection, retirement portfolio); situations where average outcome differs from typical outcome. Common applications: career pivot (probability of failure × bad outcome vs probability of success × good outcome, weighed against current path); business venture decisions; portfolio risk tolerance; major life decisions. Caveat: assessing your own utility function requires honesty + introspection. Most people overestimate near-term utility + underestimate long-term utility. For routine decisions: expected value alone is often sufficient. Reserve probability-weighted utility for major decisions where the stakes justify the analysis.
For high-stakes decisions, raw expected value isn't enough - you need to weight outcomes by how much they actually matter to you. This is probability-weighted utility theory.
## Why expected value sometimes failsClassic example:
Most people prefer Option A (the certain gain).
Why:
This is rational, not biased - your utility function genuinely treats certainty differently. Probability-weighted utility captures this.
## The probability-weighted utility formulaMathematical statement:
The challenge:
Simplified procedure:
Decision: Take a startup job (Option A) or stay at current corporate role (Option B)?
Option A - Startup job:
Option B - Stay at corporate:
Decision:
Single decisions with non-linear value:
Repeated decisions with consistent stakes:
Practice probability-weighted utility on lower-stakes choices to build intuition:
Why repeated low-stakes practice matters:
Excel / Google Sheets:
Decision-analysis software:
Decision journals (manual):
Risk aversion and loss aversion are two distinct behavioural concepts often conflated. Risk aversion is the preference for certain outcomes over uncertain ones with the same expected value - e.g. preferring a guaranteed GBP 50 to a 50/50 chance of GBP 0 or GBP 100. It's rational + symmetric (applies equally to gains and losses). Loss aversion is the tendency to weight losses roughly twice as heavily as equivalent gains - e.g. the pain of losing GBP 100 feels equivalent to the pleasure of gaining GBP 200. It's emotional + asymmetric (applies more to losses). Why the distinction matters: they predict different behaviour. Risk-averse + non-loss-averse: avoids volatility but treats gains and losses symmetrically. Loss-averse: takes MORE risk to avoid recognising a loss (e.g. holding a losing stock too long), takes LESS risk after a gain (e.g. selling winners too early). Investment implications: loss aversion explains the disposition effect (selling winners, holding losers); endowment effect (overvaluing what you own); reluctance to rebalance portfolios. Risk aversion is rational; loss aversion is a cognitive bias. Behavioural finance treats them as complementary explanations of observed behaviour, not interchangeable terms.
Risk aversion + loss aversion get confused often - even by experienced investors. This guide clarifies the distinction + shows how each affects real-world decisions.
## Risk aversion - the rational preferenceDefinition:
It's rational:
It's symmetric:
Measuring risk aversion:
Definition:
It's asymmetric:
It's emotional / behavioural, not rational:
Measuring loss aversion:
Pure risk aversion (no loss aversion):
Pure loss aversion (no excess risk aversion):
Combined (most people):
1. Disposition effect:
2. Endowment effect:
3. Reluctance to rebalance:
4. Over-insurance:
1. Reframe losses as 'opportunity costs':
2. Pre-commit to rules:
3. Use decision frameworks:
4. Diversify your reference points:
5. Practice on small-scale decisions:
Risk tolerance:
Risk aversion:
Loss aversion:
How they interact:
Bayesian updating is the formal process of revising beliefs in response to new evidence - moving away from 'this is true / this is false' thinking toward 'this is N% likely, and I'll update as evidence arrives'. The basic structure: (1) Prior probability - your starting belief BEFORE seeing the new evidence (based on history, base rates). (2) Likelihood - how probable the new evidence is, assuming your belief is true. (3) Posterior probability - your updated belief after combining prior + likelihood via Bayes' theorem. Why this matters in daily life: most decisions involve incomplete information; sticking with first impressions when contradicting evidence appears is irrational. Bayesian updating gives you a framework for proportional belief change. Three real-world UK examples: (a) Investment - new earnings report changes your view of a stock's prospects. (b) Health - a symptom changes your view of which condition you might have. (c) Career - rejection from one role tells you something (but not everything) about whether you'll get the next. Common mistakes: updating too much on weak evidence (anecdotes); updating too little on strong evidence (confirmation bias); ignoring base rates entirely. Mental shortcut: 'How strong is this new evidence relative to what I already believed?' If much stronger, update significantly; if equally strong, modest shift; if weaker, minor adjustment.
Bayesian updating is one of the most useful mental skills you can build - it works across investment, health, relationships, career, and parenting. This guide covers how to apply it in everyday UK life.
## The basic Bayes' theorem structureThree components:
Plain-English formula:
Why it's intuitive once you see it:
Scenario: You believe a particular UK FTSE 100 stock has 40% probability of beating market over next year (your prior).
New evidence: Earnings report is significantly above analyst expectations.
Bayesian update:
What this tells you:
Where humans go wrong:
Scenario: You have a chronic headache that started yesterday. You believe (prior) there's 95% probability it's a tension headache, 5% probability it's a more serious condition.
New evidence: A specific symptom appears (e.g. blurred vision).
Bayesian update:
What this tells you:
Scenario: You believe (prior) you have 30% probability of getting a particular type of role (e.g. senior product manager at a UK tech startup).
New evidence: Rejected after final-round interview for one specific role.
Bayesian update:
What this tells you:
Rule of thumb #1 - Strength of evidence vs belief:
Rule of thumb #2 - 3x test:
Rule of thumb #3 - Likelihood ratio approximation:
Decision journals:
Calibration training:
Bayesian software / spreadsheets:
Community + practice:
High-stakes single decisions:
Repeated decisions:
Less useful for:
The compounding benefit:
The Monty Hall problem is the most famous probability puzzle of the modern era. It looks trivial - three doors, one prize, one decision - but the right answer is so counter-intuitive that thousands of people, including PhD mathematicians, refused to believe it when it first hit the mainstream press in 1990. Switching doors wins the car two-thirds of the time. Sticking wins one-third. The 50/50 instinct is wrong, and understanding why teaches you more about how probability really works than any textbook chapter.
## What is the Monty Hall problem?You're a contestant on a game show. The host, Monty Hall, shows you three closed doors. Behind one is a car. Behind the other two are goats. You pick a door - say, door 1. You don't open it yet.
Monty, who knows where the car is, then opens one of the two doors you didn't pick. He always opens a door with a goat. Suppose he opens door 3 and reveals a goat. He now offers you a choice: stick with door 1, or switch to door 2.
Should you switch? Or does it not matter?
**The answer** Switch. Switching wins the car with probability 2/3. Sticking wins with probability 1/3. The two remaining doors are not equally likely, even though only two doors are left. ## Why almost everyone gets this wrongThe intuitive argument runs like this: there are two doors left, one has a car, one has a goat, so the chance of the car being behind either door is 1/2. Switching can't matter. This argument is wrong, but it's wrong for a subtle reason that almost no one spots on first hearing.
The two remaining doors are not symmetric. One of them - the door you originally picked - was chosen before Monty did anything. The other - the door Monty left closed - survived a deliberate filtering process. Monty's choice carried information. Treating the two doors as interchangeable throws that information away.
The clearest way to see this is to enlarge the problem. Imagine 100 doors instead of three. You pick door 1. Monty then opens 98 doors, every one of them revealing a goat, leaving only door 1 (your original pick) and one other door - say, door 47 - closed. Would you stick with door 1, or switch to door 47?
Door 47 is now obviously the better bet. Your original pick had a 1/100 chance of being right. The other 99 doors collectively had a 99/100 chance. Monty, by opening 98 of those 99 doors and avoiding the car, has effectively concentrated that 99/100 probability onto the single door he chose to leave closed. Switching wins 99 times out of 100.
The three-door version is the same trick at smaller scale. Your original pick has a 1/3 chance of being right. The other two doors collectively have a 2/3 chance. Monty opens one of them - guaranteed to be a goat - and concentrates that 2/3 probability onto the remaining unopened door.
## Proof 1: enumerationThe most bullet-proof proof is to list every possible game and count outcomes. Suppose, without loss of generality, you always pick door 1. There are three equally likely worlds, depending on where the car actually is.
World A: Car is behind door 1 (probability 1/3). You picked correctly. Monty opens either door 2 or door 3 (he can pick either - both have goats). If you stick, you win. If you switch, you lose.
World B: Car is behind door 2 (probability 1/3). You picked the wrong door. Monty must open door 3 - door 2 has the car, so he can't open it. If you stick, you lose. If you switch to door 2, you win.
World C: Car is behind door 3 (probability 1/3). You picked the wrong door. Monty must open door 2. If you stick, you lose. If you switch to door 3, you win.
Sticking wins in 1 of 3 worlds. Switching wins in 2 of 3 worlds. The strategy is settled.
**Why the host's constraint matters** In two of the three worlds (B and C), Monty has no choice about which door to open - he is forced. That forced behaviour is precisely what carries information about the car's location, and it's what the 50/50 intuition misses. ## Proof 2: Bayes' theoremIf enumeration feels like a trick of small numbers, Bayes' theorem gives you the same answer mechanically. We want the probability that the car is behind door 2, given that you picked door 1 and Monty opened door 3.
Let:
C₂ = car is behind door 2C₁ = car is behind door 1C₃ = car is behind door 3M₃ = Monty opens door 3The prior probability of each Cᵢ is 1/3. We need the likelihoods P(M₃|Cᵢ):
P(M₃|C₁) = 1/2 - the car is behind your door, so Monty can open either of the other two with equal probability.P(M₃|C₂) = 1 - the car is behind door 2, you picked door 1, so Monty must open door 3.P(M₃|C₃) = 0 - Monty never opens the door with the car.By Bayes' theorem:
P(C₂|M₃) = P(M₃|C₂) · P(C₂) / P(M₃)
= (1 · 1/3) / (1/2 · 1/3 + 1 · 1/3 + 0 · 1/3)
= (1/3) / (1/2)
= 2/3And similarly P(C₁|M₃) = (1/2 · 1/3) / (1/2) = 1/3. Switching gives you 2/3. Sticking gives you 1/3. Same answer.
If the Bayesian machinery feels rusty, the conditional probability deep-dive walks through the formula step by step, including a full tree-diagram derivation.
## Proof 3: simulationIf you still don't believe it, simulate it. Here's the logic in pseudocode:
wins_stick = 0
wins_switch = 0
for trial in range(10_000):
car = random.choice([1, 2, 3])
pick = random.choice([1, 2, 3])
# Monty opens a door that is neither the pick nor the car.
available = [d for d in [1, 2, 3] if d != pick and d != car]
monty_opens = random.choice(available)
# The switch target is the remaining door.
switch_to = [d for d in [1, 2, 3] if d != pick and d != monty_opens][0]
if pick == car:
wins_stick += 1
if switch_to == car:
wins_switch += 1
print(wins_stick / 10_000) # ≈ 0.333
print(wins_switch / 10_000) # ≈ 0.667This is the gold-standard sanity check for any disputed probability claim. Simulation collapses the dispute: the empirical frequencies converge on 1/3 and 2/3 as you run more trials. There is no analytical room left for doubt. Building this same kind of intuition through repetition is exactly what probability calibration training is built around.
## Proof 4: combine the unpicked doorsHere's the proof that finally clicks for most people. Forget the door Monty opened. Think of your decision as a choice between two groups of doors:
The car is in Group A with probability 1/3, and in Group B with probability 2/3. Monty isn't changing where the car is - he's just telling you, for free, which of the two doors in Group B definitely doesn't have it.
When you switch, you're effectively saying: "I want whatever is in Group B. Monty, please show me which door in Group B to open." The 2/3 probability that the car is somewhere in Group B all gets concentrated on the single remaining unopened door in Group B.
## Proof 5: generalise to N doorsThe general rule for N doors, where Monty opens N−2 goat doors after your initial pick, is:
1/N(N−1)/NWith 3 doors the gap is 1/3 vs 2/3. With 10 doors it's 1/10 vs 9/10. With 100 doors it's 1/100 vs 99/100. At any door-count the switch probability is always (N−1)/N - overwhelming for large door-counts. The reason the three-door case feels close to 50/50 is that 1/3 and 2/3 happen to look symmetric, even though they aren't. The 100-door version is the same problem, but the intuition has nowhere to hide.
This is also a useful technique to add to your toolkit: when a probability puzzle confuses you, scale it up. Many cognitive traps that whisper "50/50" stop whispering when you change three to a hundred. The same trick exposes the false positive paradox in medical testing: a 99% accurate test for a 1-in-1000 disease produces mostly false alarms, and that becomes obvious only when you think in terms of ten thousand patients rather than one.
## The vos Savant controversyThe Monty Hall problem entered popular culture through Marilyn vos Savant's Ask Marilyn column in Parade magazine in September 1990. A reader posed the question. Vos Savant - at the time listed in the Guinness Book of Records for the highest recorded IQ - gave the correct answer: switch.
The reaction was extraordinary. Parade received an estimated 10,000 letters from readers insisting she was wrong. Roughly 1,000 of them were from people with PhDs, including academic mathematicians and statisticians. Many were openly hostile. One, from a maths PhD at the University of Florida, read in part: "You blew it, and you blew it big! Since you seem to have difficulty grasping the basic principle at work here, I'll explain. ... There is enough mathematical illiteracy in this country, and we don't need the world's highest IQ propagating more."
She was right. They were wrong. Vos Savant published a follow-up column with a clearer enumeration, and asked schools to run classroom simulations. The empirical results - students switching and winning roughly two-thirds of the time - eventually settled it. Many of the academics later wrote in to apologise; some never did. Paul Erdős, one of the most prolific mathematicians of the 20th century, reportedly remained unconvinced until shown a computer simulation.
The episode is now a textbook example in books on the psychology of reasoning. It shows that mathematical training does not by itself protect you against motivated bad intuitions - and that even experts can be confidently wrong when a problem's structure clashes with how their brain wants to slice it. The same vulnerability shows up in the Dunning-Kruger effect literature: confidence and accuracy decouple far more easily than people expect.
## What changes if Monty's behaviour changesThe most important thing to notice about the Monty Hall problem is that the 2/3 answer depends on Monty's specific protocol. Change the rules and the answer changes too. Three variants worth knowing:
1. Monty Fall ("Ignorant Monty"). Monty doesn't know where the car is. He picks a random door from the two you didn't pick and opens it. By luck, it has a goat. Now the probability really is 1/2 for each remaining door. The reason: in this version, the cases where Monty would have accidentally revealed the car are excluded after the fact, and that conditioning erases the asymmetry. Monty's knowledge is what made the original puzzle work; remove it and the information vanishes.
2. Monty Hell ("Malicious Monty"). Monty only opens a door and offers a switch when your original pick was correct (he's trying to talk you out of the car). If he offers a switch, you should stick. The opening-and-offering itself is now a signal that you picked right.
3. Monty doesn't always offer. If Monty is allowed to choose whether to offer the switch - and might, for instance, offer it more often when you're wrong - then your switching odds depend on his unknown policy. Without knowing the policy, the problem is no longer well-defined.
The general lesson: a probability puzzle is only as well-defined as the protocol. "Three doors, one car, one revealed" isn't enough information. You need to know what rule generated the reveal. This is the same trap behind base rate neglect and behind many real-world probability mistakes - people focus on the numbers and forget to ask what process produced them.
## Why the puzzle matters beyond game showsThe Monty Hall problem is famous because it's a clean demonstration of three ideas that show up everywhere:
Conditional probabilities are not symmetric. Knowing one thing changes the probability of another in ways that depend on how you came to know it. The same observation, generated by a different process, can carry very different information. "Monty opened door 3" means one thing if Monty knows where the car is, and something else if he doesn't.
The structure of evidence beats the quantity of evidence. One door opened by an informed host is more useful than ninety-nine doors opened at random. Information value depends on how the information was generated, not just how much there is. This same principle is why probability and odds give different intuitions for the same underlying numbers - the framing changes what you notice.
Frequencies are clearer than fractions. Most people who get the Monty Hall problem wrong on first encounter get it right when shown a 100-door version, or when walked through a frequency table of "out of 300 games, 200 are won by switching." Switching from probabilities to frequencies is one of the most reliable de-biasing techniques in the cognitive psychology toolkit.
If you internalise just one of these - the asymmetry of conditional probability - the Monty Hall puzzle becomes obvious. And once it's obvious, a surprising number of other puzzles get easier too: the boy-or-girl paradox, the two-envelope problem, the prosecutor's fallacy, even the way Bayesian medical testing actually works in practice.
## Frequently asked questions **Q: Does the Monty Hall answer depend on which door I originally pick?** No. By symmetry, the analysis is identical for any starting door. The 2/3 switching advantage holds whichever of the three doors you choose first. **Q: What if Monty doesn't know where the car is?** Then it's a different problem (sometimes called "Monty Fall"). If Monty opens a random door and it happens to be a goat, the remaining two doors are genuinely 50/50. Monty's knowledge is what creates the asymmetry - without it, the information content of the reveal collapses. **Q: Why does my intuition keep saying 50/50?** Because it ignores the conditional structure. Your brain treats "two doors, one car" as symmetric, but the two doors got there by different processes - one was your free pick, the other survived Monty's deliberate filter. Same outcome, different conditional history, different probabilities. **Q: Is the Monty Hall problem useful outside game shows?** Yes. The same conditional-probability mistake shows up in courtroom statistics (the prosecutor's fallacy), medical diagnosis (false positive paradox), and Bayesian reasoning generally. Anywhere new information arrives via a non-random process, the Monty Hall asymmetry is present in some form. **Q: How can I convince a sceptic?** Don't argue - simulate. Run a Python or spreadsheet simulation of 10,000 games and show the empirical frequencies converging on 1/3 and 2/3. The 100-door variant of the puzzle is the second-best argument: scaling the problem up to 100 doors makes the switching advantage so obvious that the 50/50 intuition breaks on its own. **Q: Did the real Monty Hall on 'Let's Make a Deal' actually run this protocol?** No, not consistently. The historical Monty Hall used a more flexible game design and could choose whether to offer a switch. The puzzle named after him is a stylised version with a fixed rule, designed to make the conditional-probability point cleanly. The real game show was strategically richer, and partly because of that, the puzzle's answer wouldn't have applied directly. ## Bottom lineSwitch. Always switch. Switching wins 2/3 of the time when Monty knows where the car is and follows the standard protocol - and that ratio gets stronger, not weaker, as you scale to more doors.
But the puzzle's lasting value isn't the answer. It's the lesson: probability is a property of how you came to know things, not just of the things themselves. Two situations that look identical can have different probabilities if they were generated by different processes. Once that idea sinks in, a lot of confusing probability puzzles stop being confusing - and a lot of everyday reasoning starts being more careful.
--- # Survivorship Bias: The Hidden Data That Changes Everything Source: https://expectedvalue.co.uk/blog/survivorship-bias/ # Survivorship Bias: The Hidden Data That Changes Everything Survivorship bias hides the failures behind every success story - from WWII bombers to mutual funds. How to spot the missing data and decide better. **Answer** Survivorship bias is the analytical error of drawing conclusions from a sample that has been filtered by an outcome - typically by counting only the survivors. The classic example: WW2 statistician Abraham Wald examined returning bombers and observed where bullet holes clustered. The intuition was to armour those spots; Wald's insight was the opposite - the planes that took hits there returned, so armour the spots with NO bullet holes (because those planes didn't make it back). Practical implication: studying successful startups, traders, or athletes only tells you what survivors did, not what worked.In 1943, the US Air Force had a problem. Bombers were being lost over Europe at an alarming rate, and the engineers wanted to know where to add more armour. They studied the planes that came home and mapped every bullet hole. The damage clustered around the wings, the tail, the fuselage. The obvious answer was to reinforce those areas. A statistician named Abraham Wald told them they had it exactly backwards.
Wald's insight is the cleanest illustration of survivorship bias: the systematic error you make when you study the things that survived and ignore the things that didn't. It is one of the most powerful cognitive biases in decision-making, and once you start looking for it, you see it everywhere - investing, business, self-help, history.
## What is survivorship bias?Survivorship bias is the logical error of focusing on the people, things, or data points that made it past some selection process while overlooking those that did not. The selection process is usually invisible, which is what makes the bias so persistent. You see the winners; you do not see the losers, because they have left the dataset.
The result is a distorted picture of reality. Properties that looked like the cause of success may turn out to be irrelevant - or worse, properties that actively predict failure can look like recipes for success when you only sample the survivors.
It belongs to a wider family of base-rate errors: the failure to ask, "out of everyone who tried, what fraction succeeded?" When you ignore the denominator, the numerator can tell any story you want.
## What is the Abraham Wald bullet-hole story?Wald was working at the Statistical Research Group at Columbia, a wartime team of mathematicians advising the US military. The Air Force gave him a dataset of bullet damage on returning bombers and asked: where should we add armour to bring more planes home?
The intuitive answer was to armour the parts of the plane with the most bullet holes. Wald's answer was the opposite. The damage map showed only the planes that had returned. Bullet holes in the wings, tail and fuselage were survivable - those planes flew home with damage in those areas. The places with fewer bullet holes were the places where a hit was fatal: the engines and the cockpit. Planes hit there did not return, so they were missing from the dataset.
Wald recommended armouring the engines. He was right, and the Air Force adopted his advice. Decades later, his memo is still used to teach analysts that the most important data point is often the one that is missing.
**The general form** Whenever you study the survivors of a process, ask: what would the failures have looked like, and how would they change my conclusion if I included them? If you cannot answer, your conclusion is built on the wrong dataset. ## How does survivorship bias warp fund performance numbers?Mutual fund performance tables are a survivorship-bias factory. The standard published list of "top funds over the last 10 years" almost always excludes funds that closed, merged or were quietly buried during that period. Funds underperform, lose investors, get rolled into a sister fund and disappear from the records. Only the survivors stay in the published table.
Studies of US equity mutual funds have estimated that this drop-out rate is around 3-5% per year. Over a decade, that means roughly a third of the funds that existed at the start are gone - and they are gone disproportionately from the bottom of the performance distribution. The reported "average fund return" is therefore systematically too high, by an estimated 1-2 percentage points per year compared with the true return on a starting cohort.
This matters for two reasons. First, it makes active management look better than it is when compared with index funds. Second, it makes any individual fund's track record harder to interpret: the 10-year-old fund you are looking at survived a 10-year filter that 30% of its peers failed.
The corrective is to look for survivorship-bias-free data. The CRSP Survivor-Bias-Free US Mutual Fund Database is the standard academic source; it includes the corpses. Most retail-facing performance tables do not.
This connects to regression to the mean - the funds at the top of last decade's table are unusually likely to revert to average over the next, even before survivorship bias is considered.
## Why is the startup-founder template a survivorship-bias trap?Every business book and YouTube channel about successful founders has the same structural problem. They study Bezos, Musk, Zuckerberg and a hundred others, look for shared traits, and conclude that those traits caused the success. Common findings: drop out of college, work 80-hour weeks, ignore conventional wisdom, take huge risks early.
The problem is that the dataset of founders who did all those things and failed - which is much, much larger - never gets sampled. Most college dropouts who started a business are not running a trillion-dollar company. Most people who took a huge risk early lost. Most contrarians turned out to be wrong. We do not interview them because they are not famous.
If your selection process is "made it onto the cover of Forbes", every trait of the survivors looks like a success factor. The rigorous version of the question - "of all the founders who exhibited trait X, what fraction succeeded?" - almost never gets asked. When researchers do ask it, the picture is usually that the famous traits have weak or no predictive power, while boring ones (industry experience, well-capitalised launch, picking a growing market) have strong ones.
**The advice problem** If you take advice only from people who succeeded at what you are about to do, you are sampling exclusively from the survivors. Their advice will be a mix of things that helped them, things that did not matter, and things that almost killed them but did not. You cannot tell which is which without the failures. ## How does survivorship bias pollute self-help advice?The same machine runs self-help. Memoirs and biographies sample exclusively the people whose unconventional choices ended well. "I quit my job and followed my passion" is a story you only hear from the people for whom it worked. The much larger group who did the same thing and ended up broke do not write books - they go back to their old industry quietly and try not to talk about it.
The fix is not to ignore success stories. They contain useful information. The fix is to actively seek out the failures: the people who tried the same thing and did not succeed, the businesses that copied the strategy and went under, the famous "contrarian" calls that turned out to be wrong. Those data points exist; they just do not have publishers chasing them.
This is why keeping a decision journal matters. Your own decision history is a survivor-bias-free dataset of one - every decision is in there, the ones that worked and the ones that did not, with the reasoning you used at the time.
## How do you spot survivorship bias in your own thinking? - **Where is this dataset coming from?** - Before you draw a conclusion from any list, ask how things got onto the list. "Top performers of the last decade" usually means "top performers of the last decade who are still around to be measured". The selection criterion is the bias. - **What would the failures have looked like?** - Imagine the missing data points. If they would have changed your conclusion, the conclusion is unstable. If they would have reinforced it, you can be more confident - but you still need to find the failures to be sure. - **Why am I hearing this story?** - If a story about a successful approach reaches you, it is because the approach succeeded. The same approach failed many times in stories that never reached you. The signal-to-noise ratio of "famous success" is not the same as the signal-to-noise ratio of "the underlying strategy". ## How do you actually correct for survivorship bias?Knowing the bias exists is not enough - you have to actively counteract it. Three practical methods:
Find the obituaries. Whatever class of thing you are studying - funds, startups, careers, products - look for explicit lists of the failures. They are harder to find but they exist: business obituaries, fund-closure records, post-mortems. The CFA's mutual fund attribution literature is one example; CB Insights' startup post-mortems are another. The dataset of failures is rarely as well-curated as the dataset of successes, but it is rarely zero either.
Estimate the base rate. Even rough numbers help. If a strategy is reported to "work" in case studies, ask: out of how many attempts? If the answer is "out of every 100 founders who tried this, 3 succeeded", the case studies suddenly look very different than if the answer is 50 in 100. This is the same instinct that drives Bayesian thinking: the prior matters as much as the headline.
Pre-mortem your own decisions. Before committing, run the pre-mortem: assume the decision has failed, and write the story of why. This forces you to populate your own dataset of failure modes, the ones that would not appear in any biography of a survivor. It is the closest thing to making your own survivorship-bias-free corpus, in real time, before it can hurt you.
## Frequently asked questions **Q: Is survivorship bias the same as confirmation bias?** No. Confirmation bias is selectively interpreting evidence to support a belief you already hold. Survivorship bias is a structural problem with the dataset itself: the evidence you have access to is not a random sample of reality. They often co-occur - confirmation bias makes you accept the survivor data uncritically - but they are different errors with different fixes. **Q: Did Abraham Wald actually exist, or is the bomber story a myth?** Wald was real (1902–1950), worked at the Statistical Research Group at Columbia, and authored a 1943 memo titled "A Method of Estimating Plane Vulnerability Based on Damage of Survivors". The popular telling has been simplified, but the core insight - that the missing damage on returning planes pointed to where the fatal hits occurred - is genuinely Wald's. **Q: How big is the survivorship-bias effect on mutual fund returns?** Academic estimates put it in the range of 1–2 percentage points per year for US equity mutual funds, depending on the time period studied. Over a 30-year compounding window that is the difference between roughly doubling your money and roughly tripling it - a much larger gap than most investors appreciate. **Q: Can survivorship bias work in my favour?** Occasionally. If you are the one who survived (a long-running business, a long-tenured employee, a successful trader), the people you compare yourself with are also survivors, so the bias roughly cancels. But the comparison set you draw advice from - books, podcasts, biographies - almost never cancels, because it is curated for you to read it. **Q: What is the easiest fix in everyday decisions?** Ask one question before drawing any conclusion: "how were the cases I'm looking at selected?" If the answer involves a filter you cannot fully describe ("things that happened to be famous enough for me to hear about"), treat any pattern you see as a hypothesis to test, not a conclusion to act on. --- # Overconfidence Bias: Why Active Traders Underperform Source: https://expectedvalue.co.uk/blog/overconfidence-bias/ # Overconfidence Bias: Why Active Traders Underperform Overconfidence bias is investing's most expensive cognitive error. How overestimation, the planning fallacy and overprecision hide it, and what fixes it. **Answer** Overconfidence bias is the systematic gap between how confident we feel and how often we are actually right. In investing, the consequence is hyperactive trading: studies of retail brokerage accounts consistently show that the most active traders underperform passive index investors by 2-5% per year after costs. Confidence is not the same thing as accuracy.Overconfidence bias is the systematic tendency to overestimate our own skill, knowledge, and the accuracy of our forecasts. It is, by a wide margin, the most expensive cognitive bias for investors. The Barber and Odean research from the late 1990s found that the most active retail traders earned roughly 6.5 percentage points less per year than the market - not because trading is unprofitable in principle, but because the people most willing to trade frequently were the most certain they had an edge they did not actually have.
## What is overconfidence bias, in plain English?Overconfidence bias is not one thing but three closely linked errors. Psychologists distinguish between overestimation (thinking you're better than you are at an absolute level), overplacement (thinking you're better than other people), and overprecision (placing too much faith in the precision of your own forecasts). All three matter for investors, but overprecision is the most dangerous because it makes the other two harder to detect.
The textbook example is asking experienced drivers to rank themselves on driving skill: a famous Swedish study found that 88% of US drivers and 77% of Swedish drivers rated themselves above the median - a statistical impossibility. The same pattern shows up in surgeons predicting their operative success rates, in fund managers predicting their alpha, and in every retail trader who has ever told themselves "this stock can't go any lower."
## Why does the planning fallacy keep catching everyone?The planning fallacy is overconfidence applied to schedules. People consistently underestimate how long their own projects will take, even when they have explicit data on how long similar projects took in the past. Daniel Kahneman tells the story of an Israeli textbook committee where, asked to estimate completion time, members guessed 18-30 months. The expert in the room had also reviewed comparable committees and quietly estimated 7-10 years; about 40% of those committees never finished at all. The textbook took eight years.
Investing has its own version. The new investor who plans to "give it a year and see how it goes" almost always underestimates how much time they will actually spend, how often they will be tempted to deviate from the plan, and how much they will trade in moments of stress. The remedy is to use reference-class forecasting: ignore your inside view of why your project is special and look at what actually happened to comparable people doing comparable things.
## What is overprecision and why is it dangerous?Overprecision is the part of overconfidence that does the most damage and is the easiest to test. Ask someone to give a 90% confidence interval - a range they're 90% sure contains the true answer - for, say, the length of the Nile in kilometres or the year Wolfgang Mozart was born. Then check the answers. The result, replicated across thousands of subjects, is that fewer than 50% of "90% intervals" contain the right answer. People give themselves a tight range to look smart, and they are wrong far more often than they realise.
For investors, overprecision is what makes a forecast like "the S&P will return 8.5% next year" feel like analysis when it should feel like a guess. The honest version of the same statement is something closer to "between -22% and +28% with 80% confidence," which is almost useless as a basis for action. That uselessness is the point: tight forecasts are not analysis, they are the appearance of analysis.
**Quick test** Pick five questions where you don't know the answer (length of the Amazon, year Pluto was discovered, population of New Zealand). For each, give a range you're 90% sure contains the truth - wide enough that you'd be willing to bet £20. Then look up the answers. Most people get 1-2 out of 5 in their range, not the expected 4-5. That gap is overprecision. Repeating this exercise daily for two weeks is, for most people, the fastest way to feel overconfidence in their own thinking. ## Why do active traders underperform?The Barber and Odean studies of 78,000 retail trading accounts at a discount brokerage in the 1990s remain the cleanest empirical demonstration of how expensive overconfidence is in markets. Their headline finding: the 20% of accounts that traded most actively earned 6.5 percentage points less per year than the buy-and-hold market return. The mechanism wasn't that active trading is intrinsically broken - it was that costs (commissions, taxes, bid-ask spreads) compounded against an edge that, for most traders, didn't exist.
A separate paper by the same authors found a stark gender difference: men traded 45% more than women, and men's returns suffered correspondingly more. The interpretation is straightforward and well-documented in psychology: men, on average, exhibit higher overconfidence in domains they perceive as masculine, and trading was (and largely still is) one of those domains. The lesson is not about gender - it's that the people most certain they have an edge are the people most likely to trade themselves into the ground.
For most investors most of the time, the conclusion is to calibrate your confidence and then trade less than your calibrated confidence suggests.
## How does Dunning-Kruger connect to overconfidence?Overconfidence and the Dunning-Kruger effect are closely related but not identical. Dunning-Kruger says specifically that people with low skill in a domain lack the meta-cognitive ability to recognise that they are unskilled - incompetence and the inability to recognise incompetence are the same skill. Overconfidence is a broader pattern that affects experts too; in fact, some research suggests that experts are more overconfident than novices in their narrow specialism, because they overestimate the transferability of their expertise.
The practical merge of the two ideas: novices systematically don't know what they don't know, and experts systematically overestimate how much their expertise generalises. Investors fall into the second trap regularly - a successful career in software engineering or surgery doesn't carry over to picking stocks, but the confidence absolutely does.
## Is calibration the only real antidote to overconfidence?The only well-evidenced remedy for overconfidence is calibration training: the deliberate practice of making probability estimates and comparing them to outcomes. The intelligence-community work of Philip Tetlock and the Good Judgment Project showed that calibration is a learnable skill - the best forecasters in his studies got measurably better with practice and feedback, and the gap between them and ordinary forecasters was driven mostly by calibration, not by access to better information.
The concrete techniques are mundane and effective:
- Keep a decision journal - write down what you predict, when, and how confident you are. Revisit predictions when the outcome resolves, not when you remember caring about them. We cover the format in detail in our piece on decision journals. - Pre-register confidence intervals - when forecasting a number, write the range first and the point estimate second. Force yourself to live with the range your real uncertainty implies. - Run pre-mortems - imagine the project has already failed and write the post-mortem. Our pre-mortem framework walks through the technique. - Use base rates - when forecasting an outcome, start with the base rate of similar outcomes happening to similar people, then adjust based on the specifics of your situation. Most overconfident forecasts skip the first step entirely. - Track your hit rate - for predictions made at 80% confidence, do roughly 80% of them come true? If not, your scale is miscalibrated. Score yourself, not your reasoning. ## How do you spot overconfidence in your own forecasts?Three sentence-level signals that you've drifted into overconfidence territory:
Overconfidence sits at the centre of a cluster of related cognitive errors we cover in detail. Start with probability calibration training for the practical antidote, then read the Dunning-Kruger effect for the related novice-vs-expert dynamic. Pair both with decision journals as the tracking system that makes calibration measurable, and hindsight bias as the failure mode that quietly rewrites the journal in your favour. Thinking in Bets by Annie Duke is the single best book-length treatment.
--- # 12 Cognitive Biases That Wreck Probability Estimates Source: https://expectedvalue.co.uk/blog/cognitive-biases-probability/ # 12 Cognitive Biases That Wreck Probability Estimates The 12 cognitive biases that most consistently distort probability judgements - what each one does and how to defend against it. **Answer** Twelve cognitive biases reliably distort probability estimates: base-rate neglect, the availability heuristic, anchoring, confirmation bias, the gambler's fallacy, hindsight bias, overconfidence, the conjunction fallacy, survivorship bias, the prosecutor's fallacy, the framing effect, and recency bias. Each one nudges the brain toward a different kind of error - some inflate small probabilities (availability, recency), some deflate large ones (base-rate neglect, overconfidence in alternatives), some confuse conditional probabilities (the prosecutor's fallacy), and some make probabilities look more or less attractive depending on how the same number is presented (framing). The single biggest defence is explicit reasoning: write the probability down before you read further evidence, then update from it rather than from your gut.Probability isn't intuitive. Kahneman and Tversky spent decades demonstrating that even statisticians get it wrong under time pressure, because the brain runs probability judgements through pattern-matching shortcuts (System 1) rather than explicit Bayesian arithmetic (System 2). The biases below are not character flaws - they're the predictable outputs of a fast, energy-cheap inference engine. Knowing the failure modes is the first step in catching them in your own thinking.
## 1. Base-rate neglect: ignoring the priorBase-rate neglect is the habit of weighting specific evidence about a case more than the underlying prevalence of the outcome. Told that someone is shy, bookish and detail-oriented, most people guess they are more likely to be a librarian than a farmer - even when farmers outnumber librarians by 20 to 1. The personality description is doing all the work; the base rate is doing none.
The probability error this produces is consistent: rare conditions get overestimated whenever a description sounds like it 'fits' them, and common conditions get underestimated whenever the description doesn't pattern-match cleanly. In medical screening, base-rate neglect is the engine of the false-positive paradox - a 99%-accurate test for a disease that affects 1 in 10,000 people produces vastly more false alarms than true detections, because the prior matters more than the test's specificity.
Our deep-dive: Base Rate Neglect: Why Your Intuitions Are Wrong.
## 2. The availability heuristic: vivid examples misleadThe availability heuristic substitutes 'how easily can I think of an example?' for 'how common is this?'. The two questions feel similar but answer different things. Plane crashes are easy to recall because they are reported intensively; car crashes are forgotten because they happen constantly. Asked which is more dangerous per mile travelled, most people get the answer wrong - in the direction the heuristic predicts.
For probability estimates, this matters whenever a rare event is also memorable. Shark attacks, terrorist incidents, lottery wins, founder unicorns, viral product launches - all of them are over-weighted in mental probability tables because the examples come to mind faster than their actual base rates support. The fix is to pause and ask explicitly: how many trials produced these vivid cases?
Our deep-dive: The Availability Heuristic: Why Vivid Examples Mislead.
## 3. Anchoring: the first number hijacks the restAnchoring is the tendency to start from whatever number was put in front of you and adjust insufficiently. The classic experiments spin a wheel labelled 0 to 100, then ask participants what percentage of African countries are in the UN. People who saw 65 give higher answers than people who saw 10 - even though the wheel was visibly random.
The probability cost is straightforward: any time you encounter a probability estimate, that number becomes the gravitational centre of your own estimate, even when you 'know' it shouldn't. Negotiation, forecasting, planning poker, project deadlines, market valuations - anywhere the first plausible number gets spoken aloud, the rest of the conversation orbits it. Writing your own estimate down before you hear anyone else's is the most reliable defence.
Our deep-dive: Anchoring Bias: How First Numbers Hijack Judgement.
## 4. Confirmation bias: selective evidenceConfirmation bias is the tendency to seek, weight and remember evidence that supports a hypothesis you already hold, while dismissing or forgetting evidence against it. For probability judgements, this produces estimates that drift systematically toward whichever direction you started leaning - because each new piece of incoming evidence is filtered before it can update your prior.
The most insidious form is asymmetric scepticism: confirming evidence is accepted at face value, disconfirming evidence is interrogated until it can be dismissed. In investment decisions, this is how a thesis you wrote in 2022 still feels correct in 2026 despite the underlying facts having shifted. Pre-mortems and explicit decision journals are the standard counter-measures.
Our deep-dive: Confirmation Bias in Investing: Fight Your Own Brain.
## 5. The gambler's fallacy: false patterns in randomnessThe gambler's fallacy is believing that a sequence of independent random outcomes 'owes' the opposite outcome to balance out. Eight reds in a row at the roulette table feel like black is 'due'. Five girls in a row feel like the next baby is more likely to be a boy. Both intuitions are wrong: independent events do not have memories.
The probability error is structural - it confuses 'long-run frequencies tend toward the underlying probability' (true) with 'individual outcomes adjust to enforce balance' (false). The long run is achieved by averaging over an enormous number of trials, not by reverse-engineering individual outcomes. Recognising independence is the only fix; once you can name the trials as independent, the fallacy dissolves.
Our deep-dive: The Gambler's Fallacy: Why You're Wrong About 'Due' Outcomes.
## 6. Hindsight bias: 'I knew it all along'Hindsight bias is the systematic distortion of past probability estimates after the outcome is known. Asked before the 2016 US election how likely a Trump win was, most pundits gave probabilities below 20%; asked the same question after the election, the same pundits recalled having given probabilities of 40% or higher. The brain quietly rewrites the prior to match the posterior.
For probability estimation, hindsight bias does two damaging things. First, it makes us feel better calibrated than we actually are, because we remember being closer to right than we were. Second, it teaches us false lessons: we update our forecasting habits as though the signal in retrospect was obvious in foresight, which it usually wasn't. Recorded forecasts - dated, with confidence intervals, kept in a journal - are the only durable defence.
Our deep-dive: Hindsight Bias: Why Everything Looks Obvious After the Fact.
## 7. Overconfidence: tight intervals on uncertain thingsOverconfidence in probability estimation usually shows up as confidence intervals that are too narrow. Asked for a 90% confidence range on a numeric question - the length of the Nile, say, or next quarter's revenue - most people give a range that turns out to contain the true answer about 40% of the time, not 90%. The intervals look reasonable from the inside; from the outside, they are wildly under-calibrated.
This bias compounds across probabilistic decisions because each individual overconfident estimate looks defensible on its own. It's only when you record a track record across many decisions - calibration training, essentially - that the pattern becomes visible and fixable. Forecasters who do this regularly become measurably better; everyone else stays roughly as overconfident as they started.
Our deep-dive: Overconfidence Bias: Why Active Traders Underperform.
## 8. The conjunction fallacy: 'A and B' feels more likely than 'A'The conjunction fallacy is judging the probability of two events occurring together as higher than the probability of one of them alone. The classic example: Linda is bright, single, outspoken, and concerned with social justice. Which is more probable - that Linda is a bank teller, or that Linda is a bank teller AND active in the feminist movement? Most people pick the conjunction, even though logically it must be smaller (or equal).
For probability estimation, the conjunction fallacy is what makes detailed forecasts feel more credible than vague ones. A startup pitch that includes specific stages and milestones feels more likely to succeed than the same pitch without them - because each added detail makes the story more vivid, even though each added detail strictly lowers the joint probability. Narrative fluency is not evidence.
There's no dedicated post on the conjunction fallacy yet; the closest treatment is in our guide to probabilistic thinking in everyday life, which covers how to spot it in product roadmaps and life plans.
## 9. Survivorship bias: only the winners reply to the surveySurvivorship bias distorts probability estimates by sampling only from cases that made it to the present. The famous WWII example: Abraham Wald looked at bullet patterns on bombers returning from missions and recommended armouring the parts that didn't have bullet holes - because the planes hit in those areas had crashed and never came back to be measured.
For probability work, this bias is everywhere. Mutual fund track records exclude the funds that closed. Startup founder advice excludes the founders who went bust. Diet success stories exclude the people the diet didn't work for. The cumulative effect is to overestimate success rates - sometimes by an order of magnitude - because the failures aren't visible in the data set. The fix is to ask explicitly: what does the denominator actually contain?
Our deep-dive: Survivorship Bias: The Hidden Data That Changes Everything.
## 10. The prosecutor's fallacy: swapping P(A|B) for P(B|A)The prosecutor's fallacy is the conditional-probability error that has wrongfully imprisoned people. A DNA match has a 1-in-a-million false-positive rate, the prosecution argues - so the probability the defendant is innocent is 1 in a million. That's a sleight of hand: 1-in-a-million is P(match given innocent), not P(innocent given match). Reversing the conditioning requires Bayes' theorem and the prior probability of guilt before the test, which is rarely close to 50%.
For everyday probability estimation, the prosecutor's fallacy shows up whenever someone quotes a test's accuracy rate as though it answered the question 'how likely am I to have this condition given a positive result?'. The answer requires the base rate of the condition in your population - and once you include it, the posterior probability is often dramatically lower than the test's headline accuracy suggests.
Our deep-dive: The Prosecutor's Fallacy: How Courts Get Statistics Wrong.
## 11. The framing effect: identical numbers, different decisionsThe framing effect distorts the perceived attractiveness of a probability without changing the probability itself. A surgery described as having a 90% survival rate is preferred to the same surgery described as having a 10% mortality rate. A discount described as 'save 20%' is preferred to the same discount described as 'avoid a 20% surcharge'. The numbers are equivalent; the preferences aren't.
For probabilistic decision-making, framing means the same expected-value calculation can produce different choices depending on how the inputs are described. The most robust defence is to translate every framing into both the gain and the loss formulation and check whether your preference flips - if it does, at least one of the two framings is anchoring you onto an irrelevant feature of the description.
Our deep-dive: Framing Effect: Why How You Hear It Changes Your Choice.
## 12. Recency bias: the last few examples dominateRecency bias is the over-weighting of recent observations relative to the longer baseline they sit in. After three months of falling markets, market participants on average estimate the next month's odds of decline higher than the long-run base rate justifies. After a sports team wins a few in a row, fans estimate the probability of the next win above what their season-long record supports. Recency dominates because it's the easiest data to retrieve.
For probability estimation, recency bias produces the same structural error as the availability heuristic but for a specific reason - recent events are available because they're recent, not because they're representative. The defence is to formally specify the window your estimate should use (e.g. 'season-to-date win rate', 'rolling 12-month inflation') before looking at the recent data, so the data feeds a defined estimator rather than your gut.
Our deep-dive: Recency Bias in Investing and Sports Predictions.
## Defending against all twelve at onceAcross all twelve biases, three defences keep showing up in the research on improved calibration:
The deeper move is to treat probability estimation as a separate cognitive task from intuitive judgement - one that benefits from explicit tools (Bayes' theorem, calibration scoring, expected-value formulas) rather than reflex. Our decision-journal guide covers the practical version of this for everyday decisions, and the calibration training guide covers the structured-practice version for people who want to get measurably better at the underlying skill.
Further reading: Kahneman's Thinking, Fast and Slow is the canonical synthesis of the underlying research. Annie Duke's Thinking in Bets applies the same ideas to probabilistic decision-making under uncertainty. Both are worth reading in order if you want to deepen this beyond the twelve-bias surface.
## Sources and primary referencesThe underlying research for the twelve biases above:
Buy a stock and something strange happens to your reading habits. Bullish articles suddenly seem more compelling. Bearish ones look poorly researched. Quarterly results are interpreted through a filter that makes the good news loud and the bad news quiet. The market disagrees with you for a while, and you decide the market is wrong.
This is confirmation bias. Raymond Nickerson, in his 1998 review in Review of General Psychology, called it “perhaps the best known and most widely accepted notion of inferential error” - and noted it operates beneath conscious awareness in almost every domain humans make decisions in. Nowhere is that more financially expensive than investing, where every position is a live bet on a thesis you can keep defending forever if you are not careful.
## What is confirmation bias, in plain English?Confirmation bias is the systematic tendency to seek, interpret, and remember information in ways that confirm what you already believe. It is not one mechanism but a cluster of three - and the worst investing mistakes typically involve all three operating at once.
Daniel Kahneman, in Thinking, Fast and Slow, frames it as a default behaviour of System 1: the fast, associative part of the mind that builds a coherent story from whatever evidence is in front of it. Coherence feels like truth. The mind actively recruits supporting evidence and filters out the rest.
**The one-line definition** Confirmation bias is the unconscious tendency to treat evidence that supports your existing view as more credible, more relevant, and more memorable than evidence against it. ## How does confirmation bias show up in different ways?Nickerson distinguishes three distinct ways confirmation bias operates. Each is independently dangerous; together they form a feedback loop that is almost impossible to break from the inside.
You search asymmetrically. After buying a stock, you Google the ticker followed by “buy” rather than “sell.” You read the bulls on Twitter and mute the bears. The set of evidence you gather is structurally biased before you start interpreting it.
Peter Wason's 1960 rule-discovery task showed this experimentally: when participants tried to identify a hidden numerical rule, almost all tested only sequences they expected to confirm it, and almost none tested sequences that might have falsified it. Most got the rule wrong.
The same data point looks different depending on what you already believe. A 5% revenue miss is “a temporary headwind” if you own the stock and “a structural deceleration” if you don't. Lord, Ross, and Lepper's 1979 Stanford study on capital punishment is the classic demonstration: showing identical mixed evidence to people on both sides of an issue made each side more confident in their original view. Same data, opposite interpretations.
You remember the bits that fit your thesis and forget the bits that don't. The CEO's quote on margin expansion is vivid; the throwaway line about competitive pressure has gone by Friday. When you review the position six months later, you reconstruct the case for owning it from a sample of memories that has already been filtered.
## Why is investing the worst domain for confirmation bias?Most domains where confirmation bias operates have some natural correction mechanism. A doctor who misdiagnoses gets sued. A scientist whose theory contradicts experiments has the experiments to answer to.
Markets are different. The feedback is noisy, delayed, and ambiguous. A stock can fall 30% because your thesis was wrong, because the market is irrational, or because of an unrelated macro shock. By the time the truth is obvious, you've had months to construct a story that protects the original decision. Worse, you're sometimes rewarded for being wrong - a stock held for the wrong reasons can still go up. Confirmation bias is hardest to fix when the feedback loop occasionally pays you for not fixing it.
The bias also compounds with two cousins. Loss aversion makes us reluctant to admit a losing position is broken because closing it makes the loss real. Sunk cost thinking says we've already done the research, so we should stick with it. Confirmation bias provides the intellectual cover that lets the other two operate undisturbed.
## What are some real-world market examples?The thesis that won't die. An investor buys an EV battery start-up at IPO on a TAM-expansion story. Eighteen months later, three production guidance misses and a shrinking cash runway. The investor reads every analyst note that uses “long-term” and dismisses every short report as a hit piece. The thesis was never updated, only re-defended.
The forecaster who's always right in retrospect. Newsletter writers and finance Twitter accounts routinely interpret whatever happened as having confirmed their prior call. If the market rallies after they were bullish, they were right. If it rallies after they were bearish, it was a bear-market bounce that proves their broader thesis. There is no possible evidence that would falsify the worldview, which is the structural definition of pseudo-science as Karl Popper described it.
The “quality compounder” label. Once a stock has been classified as quality, subsequent earnings are filtered through that label. Margins compressing for two years are explained as “investing for growth.” A genuinely quality business and one that has stopped being one look identical from the outside if your priors will not update.
## How does a pre-mortem fight confirmation bias?The pre-mortem is the most powerful debiasing tool in the investing toolkit, and the easiest to actually use. Developed by psychologist Gary Klein, the technique is straightforward: before you act, imagine you are at your desk a year from now and the trade has been a disaster. Write down - in detail - the most plausible reasons it failed.
The mechanism is subtle. Forecasting why a position might fail invites optimistic dismissals (“sure, in theory, but…”). Imagining a failure that has already occurred bypasses the optimism filter. Klein's research found that prospective hindsight - assuming the bad outcome is real and explaining it - increased the ability to identify reasons for future outcomes by roughly 30%.
A five-minute pre-mortem before pressing buy might surface: “The new product cycle slipped two quarters,” “Insider selling accelerated and we ignored it,” “A regulatory change we knew was possible actually happened.” Anything you cannot comfortably dismiss belongs in the thesis as a tracked risk, with a pre-committed action if it materialises. Deeper walk-through: pre-mortem decision making.
**Pre-mortem prompt** It is one year from today. This investment has lost 40% of its value and you are explaining to yourself why. Write the three most likely reasons in past tense. Anything you can write convincingly belongs in your risk list before you buy. ## How do you steel-man the opposite thesis?Steel-manning is the opposite of straw-manning. Instead of caricaturing the opposing view so it is easy to dismiss, you construct the most intelligent, evidence-grounded version you can - and then ask whether your own view still survives.
For investing, this means writing the short report on every long position you own. Not a vague “risks include macro headwinds” paragraph: a real document, with the strongest specific argument an intelligent bear could make, backed by data you have checked. If you cannot articulate why a thoughtful analyst might short the stock, you do not understand the position well enough to own it.
Charlie Munger's rule is sharper: “I never allow myself to have an opinion on anything that I don't know the other side's argument better than they do.” Clearing that bar for even one or two positions reduces conviction-by-default - the silent assumption that because you bought the stock, you must already have thought through the bear case.
## How does a decision journal cut confirmation bias?The decision journal is the antidote to memory bias and to the related hindsight bias - the tendency to remember your past predictions as more accurate than they were. Without a written record, you reconstruct the reasons you bought a stock from a position of knowing how it turned out. With a journal, you can't.
The minimum useful entry has five fields:
Review entries quarterly and ask one question: given what I know now, would I make the same decision? Not whether it worked out - whether the reasoning was sound. Detailed guide: decision journals.
## Which smaller habits compound against confirmation bias?Read your bears first. Before reading the bull case for any position you are considering, deliberately read the strongest bear case. Search for the ticker plus “short thesis” before you search for “buy.” Order matters: the case you read first becomes the anchor against which the second is judged - one of the few ways to use anchoring in your own favour.
Update your prior properly. When new evidence arrives, the question is not “does this still fit my thesis?” but “how would my probability of being right change if I were updating from scratch?” This is the heart of Bayesian thinking. A thesis that survives every possible piece of evidence is not a thesis; it is an emotional commitment dressed as analysis.
## Which debiasing techniques don't actually work?“Just be aware of the bias.” Awareness alone does almost nothing. Nickerson notes the bias persists even in subjects explicitly warned about it. It operates pre-consciously; you cannot disable it by intending to.
Reading more. Consuming more content from sources that already align with your view deepens the bias by widening the asymmetry between confirming and disconfirming evidence.
Talking it through with someone who agrees with you. Group-polarisation research shows that discussing a view with people who share it makes the view more extreme, not better calibrated. The useful conversations are with people who disagree - ideally people whose intelligence you respect enough that you cannot easily dismiss them.
## Frequently Asked Questions **Q: Is confirmation bias the same as motivated reasoning?** They overlap heavily but are not identical. Confirmation bias is a general cognitive tendency to favour confirming evidence regardless of motive. Motivated reasoning is a stronger version where the bias is driven by an emotional or financial stake in the conclusion. Most investing decisions involve both, which is why they are particularly resistant to correction. **Q: Does it affect professional investors as much as retail?** Yes, and often more. Behavioural finance research - including work by James Montier - finds professionals are vulnerable at similar or higher rates because reputation, career, and bonus are tied to a published thesis. Specialisation reduces some biases but tends to amplify confirmation bias. **Q: How long should I hold a thesis before reviewing it?** Time-based reviews matter less than evidence-based reviews. Set the falsifiers in advance - specific quarterly numbers, customer events, regulatory changes - and review when those trigger, not on a calendar. A quarterly journal review catches things the falsifiers miss. **Q: Will index investing protect me from this bias?** Largely, yes - for individual stocks. A diversified index strategy removes stock-specific confirmation bias because there's no single thesis to defend. But the asset-allocation decision and the meta-decision of whether to stick with indexing during a long underperformance stretch are both still vulnerable. **Q: Is there any advantage to confirmation bias?** In some domains it has evolutionary value: efficient cognition requires shortcuts, and once a working mental model is in place, defending it is faster than rebuilding it from scratch every time. The cost-benefit only flips badly negative in domains with delayed feedback and high stakes. Markets are exactly that domain. ## The Honest Bottom LineYou will not eliminate confirmation bias. Nickerson, Kahneman, and almost every researcher who has worked on it agree it is a feature of how the mind handles evidence, not a defect that can be patched out. The realistic goal is to reduce the cost: catch the biggest mistakes before they compound, hold positions with calibrated rather than emotional conviction, and exit when the falsifiers fire instead of when the pain becomes unbearable.
The three techniques here - the pre-mortem before you buy, the steel-manned bear case for everything you hold, and the decision journal that makes your reasoning impossible to revise - are the highest-leverage defences. None takes more than twenty minutes per position. Over a decade, that is the difference between perpetually defending past decisions and visibly improving your process. Your brain is not on your side here, but a few cheap habits, applied consistently, can do most of the work for it.
--- # Regression to the Mean: Why Extremes Don't Last Source: https://expectedvalue.co.uk/blog/regression-to-the-mean/ # Regression to the Mean: Why Extremes Don't Last Why extreme performance - sporting peaks, market gains, viral hits - almost always reverts to average. Galton's discovery, examples, and how to spot it. **Answer** Regression to the mean is the statistical phenomenon where extreme outcomes are followed by less extreme outcomes closer to the long-run average. A football team that wins its first 5 matches by huge margins is likely to revert to a less extreme record across the next 20. The mechanism is partly skill, partly luck: extreme observations bundle both, and the luck component doesn't persist. Practical implication: ignore short-track-record extremes; the manager who 'beat the market 3 years running' has a much higher chance of underperforming the next year than the data alone suggests.Regression to the mean is the statistical phenomenon where unusually extreme measurements - outstanding performance, terrible results, surprise winners - tend to be followed by more average ones. It is one of the most underappreciated ideas in statistics, partly because it is invisible until you understand it, and partly because once you do, you cannot stop seeing it everywhere. Misreading regression to the mean is behind a huge fraction of bad business decisions, false faith in unproven medical treatments, and confused commentary about sports, markets, and education.
This guide covers what the effect actually is, where Francis Galton discovered it in 1886, four common real-world examples (Sports Illustrated cover jinx, fund manager performance, the placebo trap, regression to mediocrity in school cohorts), and a checklist for spotting it in your own thinking.
## What regression to the mean actually meansMost measurable outcomes - sales figures, exam scores, sporting performances, blood pressure readings - combine a stable underlying signal with random noise. Skill, conditions, sample peculiarities, and luck all bake into the number you observe. When the number you measure is unusually high, it is high partly because the underlying signal is good, and partly because the noise was favourable. Repeat the measurement, and the noise re-rolls. The next observation tends to be closer to the underlying signal than the extreme reading was.
The same logic runs the other way. If a fund manager has a terrible year, some of that is genuine underperformance and some is bad luck. The next year, the bad luck does not necessarily continue, so the manager looks better. Crucially, this is true even when nothing about the manager has changed.
Two technical points are worth holding on to. First, regression to the mean only happens when there is some genuine noise - if outcomes were entirely deterministic, extreme readings would just keep happening. Second, it works most strongly when the correlation between two measurements is weakest. The less the underlying signal explains, the more the second measurement looks like the population average rather than a near-copy of the first.
## Where the idea came from: Galton's sweet peasSir Francis Galton - Charles Darwin's half-cousin and the founder of much of modern statistics - discovered the effect in the 1880s while studying the seeds of sweet peas. He sorted parent peas by size and tracked the offspring. To his surprise, the children of giant peas were on average smaller than their parents, and the children of tiny peas were on average bigger. Both groups had drifted toward the population average.
He found the same thing in human heights. Tall fathers had sons who were tall but not as tall on average. Short fathers had sons who were short but not as short. He called this 'regression toward mediocrity' - a slightly miserable label that has stuck in the maths world ever since, although modern statisticians prefer the neutral 'regression toward the mean'.
Galton initially mistook this for an inheritance phenomenon, as if biology were pulling everyone toward an average size. The mathematician Karl Pearson later showed it was simply a property of imperfectly correlated measurements: any time two measurements share less than perfect correlation, the extreme values on one will drift toward average on the other. No biological mechanism required.
**The simple test for whether regression applies** Ask: are these measurements perfectly correlated? If yes (e.g. measuring the same height twice with a perfect ruler), no regression. If no (any noise, any change in conditions, any partial overlap of skill versus luck), regression applies. The weaker the correlation, the stronger the regression. ## Example 1: the Sports Illustrated cover jinxFor decades, fans have joked about the Sports Illustrated cover jinx - the apparent curse where athletes who appear on the magazine's cover go on to underperform in the following season. There are dozens of high-profile cases, from injured stars to hitters who lose their batting eye, that fuel the legend.
The jinx is real in the sense that the average post-cover season really does tend to be worse. But the cause is not magic. Athletes appear on the cover precisely because they have just had an extraordinary run - a hot streak, a record-breaking month, a championship win. That extreme performance combines genuine skill with favourable random variation. The next season, the variation re-rolls. Their underlying ability has not changed, but the universe is no longer cooperating to the same degree, and they look worse.
The same effect explains why rookies-of-the-year often have weaker sophomore seasons, why the Madden NFL cover has its own apparent curse, and why a basketball player who scores 50 points one night usually scores fewer the next. Each of these involves selecting on extreme observations and then watching the noise drift back to baseline.
## Example 2: fund manager performance and the star fund effectEvery year, a few mutual funds dramatically beat the index. Investors flock to them, magazines run features, the manager gets called a genius. The following year, those same funds usually drift back toward the average - not because the manager has changed, but because some of last year's outperformance was luck, and luck does not persist.
This pattern has been documented exhaustively. The S&P SPIVA scorecards consistently show that of the top 25% of US equity funds in any given year, only a small fraction stay in the top 25% the next year - barely better than chance. Pick the very top fund, and it is roughly as likely to land in the bottom half next year as the top half. The same regression-to-mean logic explains why active funds collectively cannot beat the index after fees: the year-on-year persistence in 'skill' is too weak to overcome the cost drag.
If you want a deeper dive into how to think about this kind of probabilistic edge in investing, we cover related ideas in our guide to risk vs uncertainty and the Kelly criterion for bet sizing.
The fund-manager case is also a textbook example of survivorship bias - the published "top funds" tables systematically exclude the funds that closed or were merged, inflating reported returns by an estimated 1–2 percentage points per year.
## Example 3: the placebo trap in medicineImagine you have a chronic back problem. Most weeks the pain hovers around a 4 out of 10. Occasionally it spikes to 8 - and on one of those bad days, you finally try a new supplement, an alternative therapy, or a fad treatment. A few days later, the pain is back to 4. You credit the treatment.
The trap is that you would be just as likely to feel better even if you had taken nothing. People typically seek treatment when symptoms are at their worst, and symptoms generally regress to the personal average over time, regardless of intervention. Because the treatment was applied at the peak, the natural drift back to baseline gets falsely attributed to the treatment.
This is one of the central reasons that medicine relies on randomised controlled trials with placebo arms. Without a control group experiencing the same regression to the mean without the active treatment, the apparent effect of any intervention will be inflated. Bayesian thinking is the main antidote - see our guide on Bayesian thinking for the underlying framework.
**Selection on extremes is the danger sign** If a study, anecdote, or business case selects subjects because they were already at an extreme (worst-performing teams, sickest patients, hottest stocks), regression to the mean will bias the apparent outcome of any intervention. Whenever you read 'we picked the worst 10% and applied X, and 80% improved', ask: would they have improved anyway? ## Example 4: the schools that 'turn around' versus the schools that 'collapse'Every year, a handful of schools post dramatically improved results, sometimes after a new headteacher has been appointed. Other schools see a sharp drop, often after a perceived management failure. The press attributes this to leadership, the new curriculum, or the latest reform. In reality, much of it is regression to the mean.
Consider a school that finished bottom of the league last year. That ranking reflects a mix of genuine challenges (deprivation, intake, staffing) and luck (a difficult cohort, an unusually low-achieving exam group). This year, the unlucky cohort has moved on; the new cohort is closer to the school's underlying baseline. The school's results bounce upward, and whoever is in charge takes the credit.
The reverse happens at top schools. A school riding a remarkable cohort to the top of the table will, all else equal, drop the following year as the next cohort regresses toward the school's actual baseline. This is one of the reasons that educational researchers warn against drawing causal conclusions from year-on-year league-table movements without a careful comparison group.
## How to spot regression in your own thinkingIf you are about to make a decision based on a recent extreme - a personnel change, an investment decision, a treatment switch, a strategy pivot - run the observation through this checklist before committing.
Three real-world patterns repay the effort of internalising the idea.
Hiring and promotion decisions. Companies often promote people on the back of one outstanding quarter or year. If that performance was partly luck, the promoted person regresses to a more average level - and the company concludes that promotion ruined them. The Peter Principle (people rising to their level of incompetence) overlaps heavily with regression to the mean. Better promotion decisions average performance over multiple cycles before drawing conclusions.
Investment chasing. Investors who switch out of a fund after a bad year and into one after a great year are buying high and selling low - exactly the wrong move when both will tend to regress. The discipline of dollar-cost averaging into broad indexes sidesteps this trap entirely. Combine it with a probabilistic framework for evaluating choices and most of the noise becomes manageable.
Performance feedback in sport, education, and business. Praising someone after an exceptional performance and then watching them regress can feel like the praise hurt them. Criticising after a terrible performance and watching them improve can feel like the criticism worked. Both are mostly regression. The classic study by Daniel Kahneman of Israeli flight instructors found exactly this effect, and the false lesson the instructors drew was that punishment works better than praise - when in fact neither response was driving the change.
## Frequently asked questions **Q: Is regression to the mean the same as the gambler's fallacy?** No - they are nearly opposite errors. The gambler's fallacy assumes random outcomes (like coin flips) become 'due' to balance after a streak, which is wrong because each flip is independent. Regression to the mean applies when measurements include both signal and noise, and the next noisy measurement tends back toward the underlying average. We have a separate guide on the gambler's fallacy if you want the contrast. **Q: Does regression mean my exceptional performance was luck?** Not necessarily - it means some of it was. If you score 90% on a hard exam, your underlying ability is probably above average and the noise was probably also favourable. The next score will tend to be lower, but still well above the population mean. Skill anchors the regression at your personal baseline, not the population baseline. **Q: How do I decide whether an effect is regression or a real change?** Three steps. First, check whether the case was selected on the original extreme - if yes, regression is the default explanation until proven otherwise. Second, look for a comparable control group that did not receive the intervention; their drift tells you what regression alone would predict. Third, look at the size of the change - a regression-only effect should pull the case roughly halfway back to the group average, no further. Larger effects suggest something genuine has happened. **Q: Can regression to the mean be used to make money?** Sometimes, in mean-reverting markets - pairs trading, statistical arbitrage, and certain commodities strategies all exploit regression. But the cost of misidentifying a structural change as 'just noise' is high, and the trade-offs are far more subtle than they look. Most retail attempts to time mean reversion in equities lose money to fees, slippage, and timing errors. **Q: Does regression apply to my own self-assessment over time?** Yes, and it is one of the reasons people overweight recent good or bad days. A great day at work is partly luck. A terrible day is partly luck. Tracking weekly or monthly averages catches the underlying signal far better than any single day, which is one reason habit-tracking apps and decision journals tend to beat one-off self-evaluation.Regression to the mean is one of those ideas that, once internalised, quietly improves the quality of every decision involving uncertainty. It is the default explanation for the cover jinx, the star fund, the alternative therapy that worked, the school that bounced back. None of those stories require a hidden mechanism - just an awareness that extreme observations are partly luck, and that luck does not persist.
For a wider toolkit on probabilistic reasoning, see our guides to correlation vs causation and thinking in probabilities.
--- # Status Quo Bias: Why We Stick with Defaults Source: https://expectedvalue.co.uk/blog/status-quo-bias/ # Status Quo Bias: Why We Stick with Defaults Status quo bias explained: the cognitive tendency to stick with defaults. UK financial impact: pension funds, insurance renewals, current accounts. **Answer** Status quo bias is the cognitive tendency to prefer current state over alternatives, even when alternatives would be better. In UK financial life, it's responsible for billions of pounds of suboptimal outcomes: pension auto-enrolment defaults that under-allocate to growth assets, insurance auto-renewals that increase year-on-year while customers pay 30-50% more than new-customer rates, current accounts that pay 0% interest while savings products elsewhere pay 4-5%. The underlying mechanism is loss aversion + decision fatigue + endowment effect. Practical fix: schedule annual reviews of pension allocation, insurance, and banking - one Sunday afternoon a year typically saves £500-£2,000 in UK households.Status quo bias is the most consequential cognitive bias in UK household finance, and the one most people underestimate. Insurers, banks, and pension providers all design products to exploit it - knowing that most customers will accept the renewal, the default, the existing arrangement, rather than do the work of comparison. This guide covers what status quo bias actually is, the specific UK financial consequences, and the practical schedule that prevents it from costing you thousands per year.
## What is status quo bias, in plain English?Status quo bias is the documented tendency for people to prefer their current state over alternatives, even when objective analysis suggests the alternative is better. It was named in a 1988 paper by William Samuelson and Richard Zeckhauser at Harvard, and has been replicated across decades of behavioural-economics research.
The underlying mechanism has three contributing components:
Loss aversion. Per Kahneman and Tversky's prospect theory, humans weight losses ~2-2.5x more heavily than equivalent gains. Switching from current state involves giving up something certain (the current arrangement, which you understand) for something uncertain (the alternative, which might be better but might be worse). The loss-aversion asymmetry tilts the decision toward staying.
Decision fatigue. Making decisions consumes cognitive resources. After a day of work decisions, family decisions, and minor consumer choices, the marginal decision to review your pension allocation is the one that gets deferred. Defaults benefit from being the path-of-least-resistance for an exhausted decision-maker.
Endowment effect. Humans value things they already have more than equivalent things they don't have. Your current pension fund, current insurance policy, current current-account are 'yours' in a psychological sense that newly-considered alternatives aren't. The bias against losing something you have reinforces the bias against changing.
These three mechanisms compound. The result is that 70-80% of UK consumers in standard financial categories never review their initial choice - they took whatever the default was at signup and have stuck with it for years, sometimes decades.
## What does status quo bias cost UK consumers?Pension auto-enrolment defaults. The UK's auto-enrolment system places new workers into a workplace pension's 'default fund' - typically a balanced lifestyle fund with conservative growth-asset weighting. Default funds optimise for the modal employee (someone who'll spend 40+ years in the scheme without active management) which means they're often more conservative than what a 25-year-old with 40 years of compounding ahead actually needs. The cost of staying in the default vs an age-appropriate growth-tilted fund: typically £100,000-£300,000 in eventual retirement assets, per various pension-industry analyses.
Insurance auto-renewals. Home, car, and pet insurance renewals price-walk by design: the renewal premium increases ~5-15% per year while the same insurer offers new customers a meaningfully lower rate. The FCA banned 'price walking' in 2022 - new customers and renewing customers must be quoted the same price - but the bias still operates because customers who never check assume the renewal IS the market rate. Annual UK household spend on insurance is around £1,200-£2,000; comparison-site switching typically saves 15-30%, i.e. £200-£500/year.
Current account interest. The default UK current account pays 0% interest. Even modest reform - moving £5,000-£10,000 of savings buffer to a savings account paying 4-5% - delivers £200-£500/year in tax-free interest (within the personal savings allowance). Most UK adults don't make this move, because the current account is the status quo.
Energy supplier. Pre-2022 energy crisis, switching energy supplier was the canonical UK status-quo-bias example - average savings of £200-£400/year for the cost of 15 minutes on a comparison site. Post-2022 the price-cap reality changed the maths but the underlying principle (defaults are designed to be more expensive than active alternatives) remains true for fixed-tariff comparisons in calmer market periods.
Investment platform fees. UK investment platforms charge 0.15-0.45% annually on AUM. The difference between Hargreaves Lansdown (~0.45%) and Vanguard UK (~0.15%) on a £100,000 portfolio is £300/year, or £15,000+ over 30 years of compounded fees. Most UK investors stay with the platform they signed up to first.
## Why does an 'opt-out' policy beat 'opt-in'?Status quo bias has a positive policy application: governments and employers can use defaults to nudge people toward better outcomes by making the desired action the default rather than the active choice.
The classic example is the UK pensions auto-enrolment introduced from 2012 onwards: workplace pensions changed from 'opt-in' to 'opt-out' for eligible workers, and pension participation rates rose from ~50% to ~88% within five years. The default flipped from 'no pension' to 'enrolled in pension', and most workers stuck with the new default just as they had stuck with the old one.
Organ donor registration uses the same mechanism in countries with opt-out (presumed-consent) frameworks - donation rates are dramatically higher than in opt-in (express-consent) systems, not because people in opt-out countries are more altruistic but because the default differs.
The practical implication for individual decision-making: whenever you're operating within a system that has defaults, ask whether the default is optimised for YOUR situation or for the system's modal user. If you're not the modal user (younger pension contributor, atypical insurance risk, savings-heavy current-account holder), the default is likely suboptimal and worth reviewing.
## Why does an annual-review schedule fight status quo bias? 1. **Schedule one Sunday afternoon per year** - Pick the same date each year (your birthday, end of tax year, or any memorable date). Block 3-4 hours. This is the only commitment that makes the schedule sustainable - making it ad-hoc never works because the next ad-hoc moment never comes. 2. **Review pension allocation** - Log into your workplace pension provider. Check your current allocation vs your age + retirement-date appropriate target (most UK pension providers have lifestyling/target-date funds that auto-adjust). If you're 25-40 and in a 'default' fund with 40%+ in bonds/cash, you're probably under-allocated to growth assets. Reallocating takes 15 minutes and may add £100k+ to eventual retirement savings. 3. **Quote-check home + car + pet insurance** - Use MoneySavingExpert, Compare the Market, or Confused.com to get new-customer quotes 3-4 weeks before each policy renewal. Even if you stick with your current insurer, telling them you have a cheaper quote often triggers a price-match. Annual saving: £200-£500 typical. 4. **Check current account + emergency-fund interest** - Calculate your average current-account balance (rough approximation - £3,000-£8,000 is typical UK). If you're not earning at least 4% on that, move the buffer to an easy-access savings account or money market fund. The Marcus by Goldman Sachs / Chase / Santander Edge are popular UK options paying 4-5% as of 2026. 5. **Review investment platform total fees** - Calculate your annual fees: platform fee + fund OCFs + any trading costs. If you're paying more than 0.4% all-in for passive ETF holdings, you can probably reduce this by switching platform (Vanguard UK, InvestEngine, Trading 212 ISA are competitive). Switching takes 2-4 weeks but is a one-time cost for years of compounded saving. 6. **Document the review** - Keep a one-page note of what you reviewed, what you changed, and what you decided not to change (with reasons). Next year's review starts from this note. After 3-4 years you'll have a pattern of decisions that lets you identify which categories actually need annual review vs which can move to every-2-years cadence. ## When is the status quo actually the right answer?This guide has been status-quo-bias-skeptical throughout, but the framing isn't 'always change everything'. Sometimes the status quo is genuinely the right choice and changing introduces real costs.
Long-term relationships (with people, with companies, with brands) often deliver compounding benefits that pure optimisation misses. Your bank knowing you for 20 years means easier mortgage applications. Your insurer knowing your claims history means fairer pricing in some categories. Your pension provider knowing your full contribution history matters for some withdrawal optimisations.
Decision costs are real. Changing your current account costs 8-15 hours of admin (updating direct debits, employer payroll, online accounts that have your card number). Changing your investment platform costs 2-4 weeks of out-of-market time during the transfer. These are real costs that must be netted against the gains.
Optimal portfolios are NOT continuously optimal. The fund that performed best last year is rarely the best this year. Constantly chasing the top-performer (the opposite failure mode of status quo bias) is well-documented as worse than just staying in a low-cost broad index.
The practical filter: status quo bias should be diagnosed by checking whether your current arrangement is good FOR YOU. If you've actively reviewed and decided the current arrangement is best, that's not status quo bias - that's an informed decision. The bias is the un-reviewed acceptance of arrangements that may or may not be appropriate.
## Frequently asked questions **Q: What's the difference between status quo bias and risk aversion?** Status quo bias is the preference for current state regardless of risk profile. Risk aversion is a specific preference for known outcomes over uncertain ones at any baseline state. They often co-occur (loss aversion is a contributing mechanism to both) but they're distinct. A risk-averse person who actively chose their current arrangement isn't displaying status quo bias - they're displaying considered risk aversion. The bias is the unconsidered default acceptance. **Q: How much does status quo bias actually cost UK households?** Based on published industry analyses: pension defaults (avg £100k-£300k in eventual retirement assets lost vs optimal allocation), insurance renewals (£200-£500/year per household), current-account interest forgone (£200-£500/year), platform fees (£100-£500/year on typical UK ISA portfolios). For a representative UK household this is £500-£2,000/year in annual cash flow + significantly larger in compounded retirement wealth. The annual-review schedule above recovers most of this. **Q: Does the FCA's 2022 ban on insurance price-walking mean status quo bias no longer affects insurance?** It reduces but doesn't eliminate the cost. The FCA ruled that insurers can't charge renewing customers more than new customers for equivalent cover. But (a) different insurers still vary on price, so comparison still pays; (b) the rule applies to motor and home insurance, not pet or other categories; (c) insurers compete on switching incentives outside the price-walking ban. Annual quote-checking still typically saves 15-30% in UK retail insurance. **Q: Is auto-enrolment pension default 'good' or 'bad' default?** Both. Auto-enrolment ITSELF (default = enrolled) is a clearly good use of status quo bias - retirement saving participation jumped from ~50% to ~88% which is a major welfare improvement. But the default FUND within auto-enrolment is typically too conservative for younger savers. The right move is: stay enrolled (don't opt out), but actively review and possibly change your investment selection within the scheme. The pension provider's lifestyling/target-date fund options usually exist and are the appropriate alternative. **Q: What's a good companion concept to read about next?**Decision frames - the broader cognitive-bias category that includes status quo bias, anchoring, and framing effects. Daniel Kahneman's Thinking, Fast and Slow covers the underlying psychology comprehensively. For specifically UK financial-decision applications, the Behavioural Insights Team's published work (the 'nudge unit' descended from Thaler and Sunstein's framework) is the most relevant practical resource. See our companion loss aversion guide for the related contributing mechanism.
## Related guides --- # Black Swan Events: How to Prepare for the Unpredictable Source: https://expectedvalue.co.uk/blog/black-swan-events/ # Black Swan Events: How to Prepare for the Unpredictable Black swan events explained: Taleb's definition, examples (2008, COVID, AI surge), and the barbell strategy + antifragility approach. **Answer**A black swan event is a rare, high-impact occurrence that appears unpredictable beforehand but is rationalised afterward as if it should have been obvious. Nassim Taleb's definition has three components: (1) outlier - outside the realm of regular expectations; (2) extreme impact; (3) retrospective predictability illusion. Recent examples: 9/11, the 2008 financial crisis, the COVID-19 pandemic, the 2024 NVIDIA AI valuation surge. You cannot predict specific black swans, but you can prepare for the category via Taleb's barbell strategy (extreme safety + extreme exposure, avoiding the middle) and by structuring decisions for antifragility (gains-from-disorder).
Nassim Taleb's 'black swan' is the most-borrowed and most-misused concept in modern financial writing. The point isn't to predict the next black swan - by Taleb's own definition that's impossible. The point is to structure your decisions so that unpredictable extreme events benefit you (or at minimum don't ruin you). This post covers the technical definition, what black swans actually look like in recent history, and the concrete tools (barbell strategy, antifragility, convexity) you can apply to decisions that have unknown unknowns.
## How does Nassim Taleb define a Black Swan event?Nassim Nicholas Taleb defined the black swan in his 2007 book of the same name. The phrase comes from the historic European belief that all swans were white - black swans were, by definition, impossible - until Dutch explorers found them in Australia in the 17th century. A theory that 'all swans are white' was destroyed by a single counterexample, however many white-swan observations had preceded it.
Taleb's three components:
1. Outlier. The event is outside the realm of regular expectations because nothing in the past suggests its possibility. The 2008 financial crisis wasn't an 'outlier' in the colloquial sense - many people were warning about housing market problems - but it was a true black swan in the structural sense: the magnitude of the global collateral chain collapse was outside the modelled risk envelope of every major bank.
2. Extreme impact. The event has consequences that dwarf normal-distribution outcomes. The 2008 crisis didn't just cause a recession - it destroyed several major banks, restructured global financial regulation, triggered the Eurozone debt crisis, and shifted global political dynamics.
3. Retrospective predictability illusion. After the event, narrative explanations make it seem predictable in hindsight. 'The housing bubble was obvious' becomes received wisdom even though it wasn't acted on at scale before the crisis. This component is the most important practical insight: humans construct post-hoc explanations that don't actually translate to predicting the next black swan, because the next one will be different.
All three components are required for Taleb's definition. A rare-but-predictable event (a 100-year flood in a flood-plain) is not a black swan - the impact may be extreme but predictability is normal. A common event with no consequence is also not a black swan. The combination of outlier + extreme + post-hoc-rationalisation is what makes the category structurally hazardous.
## What are some real-world Black Swan examples?The 2008 financial crisis. The collapse triggered by US subprime mortgages cascading through structured credit products. Most major banks' risk models priced housing-market correlation at near-zero across regions; the correlation went to ~1 during the crisis and the entire collateral chain failed simultaneously. This was the canonical black swan in Taleb's writing - the structure of the modelled-risk world made the unmodelled event invisible until it arrived.
The COVID-19 pandemic. A novel coronavirus pandemic was discussed in epidemiology academic literature for decades before 2020, but the global economic and social response - simultaneous lockdowns, supply-chain disruption, fiscal stimulus on a scale unseen since WWII - was outside the modelled scenario range of every major institution. Like 2008, the post-hoc rationalisation makes it seem inevitable; but the specific shape of the response was genuinely unmodelled in early 2020.
The 2024-2025 AI valuation surge. NVIDIA's market cap quadrupled in 18 months as large language model adoption accelerated. Many analysts had AI as a long-term thesis but the specific velocity of the 2024-2025 valuation expansion was outside published price targets. Positive black swans are less-discussed than negative ones but the structural properties (outlier, extreme impact, retrospective rationalisation) are identical.
9/11. The geopolitical and aviation-security restructuring that followed was outside the modelled scenario range. Subsequent rationalisation made it seem predictable - the underlying terrorist network had been tracked - but the specific event was not.
What these examples share: in retrospect, narrative explanations make each event seem predictable. The hard test is whether anyone with skin in the game was structured to BENEFIT from the event before it happened, in proportion to how much would have been lost if it didn't happen. Most weren't - which is the practical signal that the event was a real black swan.
## How do you prepare for Black Swans with the barbell strategy?Taleb's primary preparation tool is the barbell strategy: hold extreme safety on one end and extreme risk on the other, avoiding the medium-risk middle that intuition suggests is sensible.
Applied to investing: 80-90% of your portfolio in the safest available assets (UK Treasury gilts, FSCS-protected cash, short-term government bonds), 10-20% in extremely high-upside positions (early-stage equity, options, venture exposure). Avoid the medium-risk middle - balanced funds, dividend stocks, investment-grade corporate credit - because they're vulnerable to black swan downside without enough black swan upside to compensate.
The barbell logic: when a negative black swan hits, the 80-90% safety bucket limits your downside to ~10-15% of portfolio. When a positive black swan hits (some early-stage position 20-50x's), the 10-20% upside bucket can deliver meaningful aggregate returns. The medium-risk middle does neither - it loses 50-60% in negative black swans without delivering compensating upside.
For non-investing decisions, the same logic applies. Career: invest most of your time in safe income-generating work, allocate small bets to high-upside speculative work (side projects, startup equity, niche skill development). Relationships, health, learning - same structure. The principle is that nonlinearity in outcomes means you want maximum exposure to outsized positive outcomes while limiting maximum exposure to outsized negative ones.
## What is antifragility, the principle underneath?Antifragility is Taleb's broader concept (his 2012 book of the same name): the property of systems or strategies that benefit from disorder rather than being damaged by it. The relationship to black swans: black swans are the proximate cause of system failures; antifragility is the property that lets you not just survive them but gain from them.
Examples of antifragile structures:
The opposite is fragility - structures that fail catastrophically under stress. A heavily-leveraged business has fragility. A career dependent on a single skill in a declining market has fragility. A relationship without ongoing communication has fragility.
The practical implication: when you can't predict the next black swan, you can ask whether the structures you're operating within are fragile, robust (neither benefit nor harm from disorder), or antifragile. Move toward antifragile structures over time.
## What is convexity, the mathematical handle?Convexity is the mathematical property that makes the barbell strategy and antifragility work. A convex payoff function has limited downside and unlimited upside (or vice versa - concavity is the opposite). Options contracts are the canonical convex instrument: maximum loss is the premium paid, maximum gain is unbounded.
The black-swan implication: convex payoff structures benefit when extreme events occur, because extreme events are precisely where the convex curve diverges from the concave alternative. Linear payoff structures (most income, most balanced investment portfolios) get neither - they neither benefit nor catastrophically lose from black swans, but they also miss the asymmetric upside that convex structures capture.
For most non-financial decisions, you can ask the convexity question directly: 'If this works out 10x better than expected, do I capture that upside? If it goes 10x worse, am I bankrupted by it?' A career bet, a relationship investment, a learning investment - the convexity test applies.
## What is the hardest part of acting on Black Swan risk?The hardest practical part of preparing for black swans isn't the strategy choice - it's accepting that your current model is going to be wrong in ways you can't predict, and structuring decisions around that humility.
This runs counter to almost every consumer-facing financial product. Financial advisors recommend balanced 60/40 portfolios precisely because the medium-risk middle SOUNDS sensible. Career advisors recommend skill specialisation because 'becoming an expert' is the deliverable language they can sell. Health advisors recommend specific diets because 'the right diet' is the simplifying narrative.
The black-swan-aware framing pushes against all of these. Not because they're wrong on average - they're optimised for the modal outcome - but because the modal outcome isn't the dangerous one. The dangerous outcomes are the tail events. The medium-risk middle that's optimal at the mean is suboptimal at the tails.
This is a real ongoing cost of black-swan thinking. You will look slightly silly during normal-state periods (your portfolio underperforms the balanced fund; your career path looks meandering). The compensation is structural protection plus structural exposure to the events that actually move the long-run distribution.
## Frequently asked questions **Q: What's the difference between a black swan and a tail risk?** Related but distinct. Tail risk is a statistical concept - low-probability outcomes in the tails of a probability distribution. Tail risks are known unknowns (you know they exist; you can't precisely time them). Black swans are unknown unknowns by Taleb's definition - events outside your model entirely. The 100-year flood in a flood plain is tail risk, not a black swan; the simultaneous correlation breakdown in 2008's CDO market was a black swan because it was outside the modelled correlation envelope. **Q: Can you actually predict black swans?** By Taleb's own definition, no - if you could predict them they wouldn't be black swans. What you can do is observe categories: financial systems with extreme leverage, biological systems with single-point-of-failure structures, geopolitical systems with concentrated power. These categories are MORE LIKELY to produce black swans than diversified, distributed, lower-leverage systems. The practical implication: move away from black-swan-prone categories in your own exposures, even if you can't predict the specific event. **Q: Is the barbell strategy backed by published research?** The barbell strategy itself is Taleb's prescriptive framework rather than an empirically-tested investment thesis. The underlying mathematical principle (convex payoff structures outperform concave ones at the tails) is uncontroversial - it's basic option-theory mathematics. The applied investment recommendation (80-90% safe + 10-20% high-upside) is a heuristic implementation of that principle. Comparable to Kelly criterion sizing, it's mathematically sound but the specific percentages aren't single-study-tested. **Q: How does this apply to my career?** Black swans hit careers similarly to portfolios. AI automation, industry disruption, company-specific failures all create black-swan-shaped career risks. Practical applications: avoid skill concentration in a single firm or technology; build optionality via transferable skills and side projects; structure income so a single-source failure doesn't catastrophically affect you. The barbell career structure: a stable income job + speculative side projects with high upside. Same principle as the investing barbell. **Q: What should I read after this?** Taleb's full book series - 'Fooled by Randomness' (2001), 'The Black Swan' (2007), 'Antifragile' (2012), 'Skin in the Game' (2018). All readable, all build on each other. For the financial-engineering side, his 'Dynamic Hedging' (1997) is the technical foundation. For shorter reading, our companion guides on Taleb's ideas explained and risk vs uncertainty cover the related concepts in summary form. ## Related guides --- # Probabilistic Thinking in Daily Life: 7 Practical Uses Source: https://expectedvalue.co.uk/blog/probabilistic-thinking-everyday-life/ # Probabilistic Thinking in Daily Life: 7 Practical Uses How to apply probabilistic thinking to medical decisions, career bets, insurance, investing, dating, sports betting and travel. 7 worked examples. **Answer** Probabilistic thinking means treating uncertain outcomes as distributions rather than single predictions, and choosing actions by their expected value rather than their most-likely outcome. Across seven everyday domains - medical decisions, career bets, insurance, investing, dating, sports betting, and travel - the same three habits make the difference: estimate base rates first, separate the decision from the outcome, and update beliefs with new evidence rather than anchoring on prior intuitions.Most everyday decisions are gambles. We make them sequentially, often quickly, and we judge them mostly by how they turned out. Probabilistic thinking is the discipline of judging them instead by the quality of the decision itself - the available information, the estimated probabilities, the expected value across all outcomes - rather than by the single outcome that happened to land. This guide covers seven concrete domains where the discipline pays back the time it takes to apply it.
**The three habits underneath all seven** Before the worked examples: every probabilistic-thinking application in this article uses the same three underlying habits. (1) Estimate the base rate before anchoring on the specific case. (2) Separate the decision from the outcome - good decisions can produce bad outcomes and vice versa. (3) Update beliefs with new evidence in proportion to its strength (Bayes' rule, informally applied). Everything else is variation on these. ## How does probabilistic thinking change medical decisions?The most consequential probabilistic-thinking application most people make is around medical screening. The classic example: a screening test for a rare disease (1-in-1,000 base rate) has 99% sensitivity and 99% specificity. A positive result is wrong, intuitively, ~1% of the time. The actual posterior probability of having the disease given a positive result is much lower than that intuition suggests.
Worked through formally: out of 100,000 people, 100 have the disease (base rate 1/1000). The test catches 99 of them (99% sensitivity). Of the 99,900 who don't have the disease, the test wrongly says positive for ~999 (99% specificity = 1% false-positive rate). So 1,098 people test positive but only 99 actually have the disease - ~9% true-positive rate among positives. The intuitive 99% is off by an order of magnitude because the base rate dominates.
This pattern - the prosecutor's fallacy, or base-rate neglect - applies wherever a test, signal, or red flag is rarer than the false-positive rate of the detector. See our base-rate neglect explainer for the wider framing. The practical takeaway: when a positive medical test arrives, ask 'what was the base rate for this condition in someone like me?' before assuming the test result is the answer.
## 2. Career bets: switching jobs and the option value of waitingCareer-switch decisions tend to be evaluated as binary choices ('take the new role' vs 'stay'), but they're better modelled as bets across a distribution of outcomes for each path. The new role might land at salary X with promotion timeline Y; the current role compounds to known-X with known-Y. The probabilistic version surfaces three questions the binary frame hides.
First, what's the variance? A new role with a higher expected salary but wider variance (e.g. an early-stage startup) is a different proposition from one with a higher expected salary and lower variance (a senior IC role at an established firm), even at identical means. Most career-switch advice ignores the variance.
Second, what's the option value of waiting? If the new role's offer is unlikely to be the last similar offer you'll receive in the next 12 months, declining preserves the option to take a better one. Decision-theory people frame this as the optimal stopping problem - the same maths as the secretary problem, in a different guise.
Third, what's the recovery cost if the bet fails? Switching to a role that you exit within 12 months has a real cost - lost compensation, CV damage, return-cost to a previous employer. Higher recovery cost narrows the range of bets worth taking.
## 3. Insurance: when the expected-value frame breaks downInsurance is one of the cases where naive expected-value calculations give the wrong answer. The expected value of an insurance contract is always slightly negative to the buyer (the insurer needs to cover claims plus its costs plus a margin), so by pure EV insurance is a money-losing bet.
The probabilistic-thinking move is to recognise that EV is the wrong yardstick when outcomes have different utility curves across the loss distribution. A £5,000 unexpected expense is meaningfully more painful than a £500 unexpected expense times ten, because the marginal utility of money declines as you have less of it. The relevant question isn't 'what's my EV?' but 'what's the cost of the unhedged worst-case outcome, and can I absorb it?'
Practical application: insure against catastrophic risks (home, life, liability) that would meaningfully disrupt the rest of your finances. Don't insure against minor risks (extended warranties on consumer electronics, mobile phone accidental damage) where the worst case is annoying but absorbable - the negative EV isn't compensated by utility-curve protection. See our insurance explainer for the actuarial side.
## 4. Investing: the role of diversificationSingle-stock vs index investing is the canonical probabilistic-thinking application in retail finance. A single stock has high variance around its expected return; a broad index has the same expected return at much lower variance because the unsystematic risk diversifies away.
The mathematical point is that variance scales sub-linearly with the number of independent positions, so for the same expected return you get less variance with more positions. The practical point: most retail investors hugely under-diversify, partly because of recency bias (chasing recent winners) and partly because the upside-skew of individual stocks (lottery effect) is more salient than the variance.
Probabilistic thinking suggests two adjustments: (1) hold a global market-cap-weighted index rather than picking sectors or single stocks for the bulk of long-term wealth, and (2) use Monte Carlo simulation rather than single-path projections when planning retirement, because the variance around the median-case 30-year path is meaningful enough to change the right savings rate.
## 5. Dating: thinking in distributions, not single matchesModern dating apps create the same single-path-vs-distribution problem as retail stock-picking. Each match presents as a binary 'this person or not'; the probabilistic frame treats matches as draws from a wider distribution, and decisions as how long to sample before committing.
The optimal-stopping literature has a clean answer for sequential search with no recall: sample roughly 37% of the expected pool, then commit to the next candidate who exceeds the best seen in the sample. That's the secretary problem's analytical solution. Real dating breaks the no-recall assumption - candidates can sometimes be re-approached - but the deeper point is that the optimal strategy involves an explicit sampling phase before a commit phase, not an attempt to score each candidate against a fixed template.
The probabilistic-thinking move here isn't to apply the maths literally - it's to recognise that early-pool decisions have lower commit value than later-pool decisions, and to act accordingly. The intuition that 'this match feels right' is usually low-information early in the sampling phase.
## 6. Sports betting: where the edge actually comes fromMost retail sports bettors think they're profitable when they're not, because they remember wins more strongly than losses (recency bias again) and because the variance of small-stake betting masks negative-EV results for hundreds of bets before the true mean asserts itself.
The probabilistic-thinking move is to evaluate each bet by the gap between your estimated probability and the implied probability of the bookmaker's odds, not by whether the bet won or lost. A bet at 3.0 odds (implied 33.3%) where you estimated 40% probability has positive EV regardless of whether that specific bet wins. After 1,000 such bets, the bettor who consistently identifies 6-point edges will be ahead; the bettor who picks 'good bets' on intuition won't.
The other element is bankroll management. Even known-positive-EV bets have variance; the Kelly criterion formalises the optimal-fraction-of-bankroll bet size as a function of edge and odds. Practical retail bettors typically use fractional Kelly (half or quarter) to reduce variance at a modest cost to long-run growth.
## 7. Travel: should you book the connecting flight?The smallest of the seven domains but illustrative because the maths is easy. A connecting flight saves £200 vs the direct option but has a 15% chance of misconnecting (delay, cancellation, baggage). A misconnect costs roughly £150 in rebooking + £200 in lost time (one wasted day). EV of the connecting option:
EV(connect) = 0.85 × £200_saved + 0.15 × (-£350_cost) = +£170 - £52.50 = +£117.50 vs direct
The connecting flight is positive-EV by about £117. But the variance is wide: 85% of the time you save £200, 15% of the time you lose £350. If a misconnect would derail a tightly-scheduled business trip or an irreplaceable family event, the variance matters more than the EV. The probabilistic move is to take the connecting flight when variance is tolerable (a casual holiday) and pay the £200 premium when it isn't (a wedding day).
This is the same pattern as the insurance case (#3) - EV and utility curves diverge as the worst-case impact increases. The wider point: a single EV number isn't enough; the variance and the utility-impact of the tail matter too.
## How to actually use probabilistic thinking day to day 1. **Start with the base rate, not the specific case** - Whenever new information arrives, anchor on what's true for the broader reference class before considering what's specific to your situation. The medical-screening example is the canonical demonstration - base-rate neglect is the most expensive probabilistic-thinking error in everyday decisions. 2. **Separate the decision from the outcome** - A good decision can produce a bad outcome (15% chance of misconnect arrived); a bad decision can produce a good outcome (a recency-bias-driven stock pick happened to triple). Evaluate decisions by their inputs - estimated probabilities, expected value, variance, utility impact - not by the single outcome that landed. 3. **Run the rough EV calculation, then check the variance** - For any uncertain decision, do the basic expected-value arithmetic (probability × outcome, summed across cases). Then ask whether the variance is tolerable - or whether the worst-case tail is bad enough to override the EV. Both numbers matter. 4. **Update with new evidence, not all-at-once** - Bayes' rule says new evidence should shift beliefs in proportion to its strength, not flip them completely. A friend's anecdote about a stock isn't the same evidence weight as a peer-reviewed study. Calibrate the update size to the evidence quality. 5. **Calibrate your forecasts by checking them later** - Track explicit forecasts on a notebook or spreadsheet. After 50-100 forecasts, you'll see your systematic biases - typically over-confidence in your own predictions, and under-weighting of base rates. Calibration improves with feedback; intuition alone doesn't. ## Frequently asked questions **Q: What is probabilistic thinking?** Probabilistic thinking is the discipline of treating uncertain outcomes as distributions of possible results rather than single predictions, and choosing actions by their expected value (and variance) across that distribution rather than by their most-likely outcome. It contrasts with deterministic thinking, which collapses uncertainty into a single guess and judges decisions by whether that guess turned out right. **Q: Is probabilistic thinking the same as expected-value thinking?** Expected-value thinking is a key component but not the whole picture. EV uses the probability-weighted average of outcomes; full probabilistic thinking also considers variance (how wide the distribution is) and the utility curve (how outcomes feel at different points). Insurance is a clear case where EV alone gives the wrong answer because the utility curve matters more than the headline EV. **Q: How do I improve at probabilistic thinking?** Three steps: (1) make explicit forecasts and track them - calibration improves with feedback; (2) read the foundational behavioural-finance literature (Kahneman's Thinking, Fast and Slow, Taleb's Fooled by Randomness, Tetlock's Superforecasting); (3) practice the three habits (base rates first, decision-vs-outcome separation, Bayesian updating) deliberately for a few weeks until they become reflexive. Tetlock's research on superforecasters shows the skill is learnable and matters more than raw intelligence. **Q: Is probabilistic thinking useful in domains without numbers?** Yes. The framework doesn't require precise numbers - it requires explicit estimation. Even rough probability estimates ('I'd say 70% likely') are dramatically more useful than implicit unstated probabilities, because they're falsifiable and improvable. The dating-decisions example uses no real numbers but the optimal-stopping framework still gives better decisions than implicit pattern-matching. **Q: What's the biggest mistake people make with probabilistic thinking?** Base-rate neglect - ignoring the broader reference class and over-weighting the specific case. The medical-screening example shows this dramatically (a 99% positive feels like 99% probability of disease, when the actual probability is closer to 10% because the base rate is low). Recency bias is the next biggest - over-weighting recent observations relative to long-run base rates. ## Related guides --- # Recency Bias in Investing and Sports Predictions Source: https://expectedvalue.co.uk/blog/recency-bias/ # Recency Bias in Investing and Sports Predictions Recency bias: why investors chase last quarter's winners and bettors back the form team. The mechanics, the cost, and the four tools that defuse it. **Answer** Recency bias is the tendency to overweight recent observations when predicting future outcomes - last quarter's stock returns, the last five games of a football team, the most recent news cycle. It's costly in finance (momentum-chasing investors systematically underperform the index) and sports betting (the 'form team' is priced as if recent results predict future results more strongly than they do). The defence is process: long-window base rates, explicit time-decay weighting, and pre-committed rules that override the pattern your brain wants to see.Recency bias is one of the most expensive cognitive biases an individual decision-maker can have because it operates at scale in two domains where money flows: investment markets and sports betting. The fix isn't to think harder about each decision (that doesn't work for system-1 biases); it's to install structural rules that override the urge to extrapolate the most recent data.
## What is recency bias?Recency bias is a tendency to weight recently observed outcomes more heavily than older ones when forecasting what comes next. The bias was formally described in the behavioural-finance literature by Daniel Kahneman and Amos Tversky (whose Prospect Theory work won the 2002 Nobel Prize in Economics) as part of the broader research programme on heuristics and biases. The mechanism is straightforward: human memory has a recency advantage, recent events are easier to retrieve, and ease-of-retrieval gets misinterpreted as predictive weight.
The bias is distinct from gambler's fallacy (which goes in the opposite direction - assuming after a streak the opposite outcome is 'due') and from base-rate neglect (which is a broader failure to anchor on long-run frequencies). Recency bias and gambler's fallacy can coexist in the same person: the bias affects skill-based judgments (which teams are good now) while gambler's fallacy affects pure-chance judgments (which roulette number is overdue).
## How does recency bias hurt investors?In equity markets, recency bias drives two well-documented behavioural patterns.
Performance-chasing. A fund manager who outperformed last year sees inflows; a manager who underperformed sees outflows. The Morningstar 'Mind the Gap' studies have measured the investor return gap (the difference between fund returns and the returns actual investors earned in those funds) at roughly 1.5 to 2 percentage points annually across major fund categories. The gap is largely explained by recency-driven flows - money arrives after performance, leaves after underperformance, and structurally captures less of the long-run return than buy-and-hold.
Sector and theme rotation. The same effect operates at the sector level. Investors pile into whichever sector or theme led last quarter (commodities, AI, biotech, whatever) on the implicit assumption that the recent trend persists. Academic work on momentum factor behaviour establishes that momentum is real (Jegadeesh and Titman 1993, and the broader literature on factor investing) but the typical retail momentum-chase enters late, captures a smaller fraction of the move than systematic momentum strategies, and exits during the inevitable mean-reversion phase.
## How does recency bias warp sports prediction?Sports markets price recency bias into team odds with remarkable consistency. The 'form team' - a team on a winning streak - is priced shorter than its underlying skill would suggest, and the 'out-of-form' team is priced longer. This creates persistent value on the contrarian side when the streak doesn't reflect underlying ability changes.
The hot-hand fallacy is the related but distinct phenomenon at the player level: the belief that a basketball player who's hit several shots in a row is more likely to hit the next one. The original 1985 Gilovich, Vallone and Tversky study found no statistical evidence of hot-hand effects in NBA shooting; more recent work (Miller and Sanjurjo 2018) has identified a small selection bias in the original methodology that, when corrected, restores some real hot-hand signal - but the magnitude is tiny compared to the prediction markets' implicit weighting. Betting odds price the perception of hot-hand much more strongly than the underlying data supports.
The five-game form sample size is also too small to be statistically informative for most football leagues. Five matches is roughly 5% of a Premier League season, and the variance in five-match outcomes is large relative to the true team-quality signal. Recency bias treats this small sample as predictive; it isn't.
## Why is recency bias so hard to fix?Three factors make recency bias resistant to debiasing through awareness alone.
First, memory weighting is unconscious. The bias operates before the deliberative system gets involved. By the time you're consciously thinking about a forecast, the recency-weighted retrieval has already happened. 'Trying not to be biased' rarely works because the bias has already shaped the inputs.
Second, recency is often genuinely predictive on short horizons. Momentum is a real factor in equity markets. Form does carry some signal in sports. The bias isn't a complete miss - it's a systematic over-weighting. Mild recency bias would beat completely ignoring recent data; what makes recency bias costly is the magnitude of the over-weighting and the failure to fade as horizon lengthens.
Third, social reinforcement. Financial media, sports commentary, and casual conversation all weight recent events. If you're consuming the same recent-event-heavy information stream as everyone else, you're constantly being re-anchored on the latest data point. Debiasing requires intentional separation from the high-recency information stream.
## What are the four practical defences against recency bias? 1. **Anchor on long-window base rates first** - Before looking at recent performance, ask what the long-run base rate is. For equity sectors: the 10-year or 20-year return rather than the 1-year. For sports: full-season or multi-season win rates rather than the last 5 games. Long-window base rates aren't always more predictive than recent data, but they're a structural counterweight that makes the recency overweighting visible. 2. **Use explicit time-decay weighting** - If you genuinely think recent data carries more signal than older data, encode that with explicit weights (e.g. half-life of 6 months, weighted moving average over 24 months) rather than implicit recency bias. Explicit decay lets you adjust the weighting and audit your forecasts; implicit recency bias does whatever your brain happens to do that day. 3. **Pre-commit to rules that override the urge** - If you're an investor: pre-commit to fixed asset allocation rebalancing on a calendar schedule (quarterly or annually). The rule forces you to buy underperformers and sell outperformers - the structural opposite of recency-chasing. If you're a bettor: pre-commit to value criteria (price-vs-fair-odds gap) rather than form-driven team selection. Both approaches use rules to override the bias that conscious effort can't fix. 4. **Audit your wins and losses for recency-bias patterns** - After 50 to 100 decisions, look back at which ones were recency-driven (you held / bought / bet because of recent performance) versus base-rate-anchored (you held / bought / bet because of long-run characteristics). The recency-driven decisions are usually meaningfully worse on average. Seeing the pattern in your own historical data is the most effective form of evidence for behaviour change. ## Frequently asked questions **Q: What is recency bias?** Recency bias is a cognitive tendency to overweight recently observed outcomes when forecasting future outcomes. It causes investors to chase last quarter's winners (and exit losers near the bottom), and causes sports bettors to back the form team or fade the out-of-form team beyond what underlying skill suggests. It was formally described in the behavioural-finance literature by Daniel Kahneman and Amos Tversky. **Q: How is recency bias different from gambler's fallacy?** Opposite direction. Recency bias extrapolates a streak forward ("the team has won 5 in a row, they'll win the next one too"). Gambler's fallacy expects a streak to reverse ("the roulette wheel has hit red 5 times, black is due"). Recency bias is more dangerous in skill-based domains like investing and sports prediction; gambler's fallacy is more dangerous in pure-chance domains like casino games. See our gambler's fallacy guide for the deeper comparison. **Q: Does recency bias mean recent data is worthless?** No. Recent data does carry signal on short horizons - momentum is a real factor in equity markets, and form does mean something in sports. The problem is the magnitude of the implicit weighting. Recency bias overweights recent data, not 'weights recent data above zero'. The fix is explicit time-decay weighting, not ignoring recent data entirely. **Q: What's the financial cost of recency bias?** Roughly 1.5 to 2 percentage points annually for the average retail investor, per Morningstar's recurring 'Mind the Gap' studies measuring the investor return gap across major fund categories. Compounded over a 30-year investing horizon, that's a difference of 60-90% in terminal wealth - the difference between £200k and £350k from the same monthly contributions and the same fund choices, depending on the entry and exit timing pattern. **Q: How do I tell if my decision is recency-driven?** Two tests. First, would you have made the same decision 12 months ago with the same long-run information? If no, the change is likely recency-driven. Second, can you state the decision rule in advance? Pre-committed rules ('I rebalance to 60/40 every January') are structurally non-recency-biased; on-the-fly decisions almost always have a recency-weighting baked in. See Monte Carlo thinking for a related framework that defuses single-path recency anchoring. ## Related guides **Editorial note** This guide cites the behavioural-finance literature (Kahneman and Tversky's heuristics-and-biases programme; Jegadeesh and Titman's momentum work; Gilovich, Vallone, and Tversky's hot-hand study; Miller and Sanjurjo's 2018 correction) as primary sources. Morningstar's 'Mind the Gap' investor-return-gap data is updated annually. Refreshed periodically as new research lands. --- # Decision Journals: Track Your Thinking to Improve It Source: https://expectedvalue.co.uk/blog/decision-journals/ # Decision Journals: Track Your Thinking to Improve It A decision journal is the only honest record of how well you actually think. What to record, when to review, and the mistakes that ruin the practice.Most of us judge our decisions by how they turn out. The job offer we accepted that worked out was a 'good decision'; the investment that lost money was a 'bad decision'. This is the most common mistake in everyday thinking - it conflates decision quality with outcome quality, and the two are not the same thing. A decision journal - a written record of decisions and the reasoning behind them, captured before you know the outcome - is the single most effective tool for separating one from the other. By logging what you knew, what you assumed, and what you expected at the moment of choice, you build the only honest data you'll ever have about how well you actually think. This guide covers what to record, how to review entries, and why the people who try it rarely go back to deciding without one.
**TL;DR** Outcomes are noisy. Bad decisions get good results, and good decisions get bad ones. Without a written record taken before the outcome is known, hindsight bias rewrites your memory and you can't tell which is which. A decision journal - six fields, five minutes per entry, scheduled review - is the antidote. The most important field is your stated confidence number; everything else is supporting context. ## What is a decision journal?A decision journal is a written log of decisions you make, recorded at the moment of decision, capturing your reasoning, assumptions, expected outcomes, and confidence levels. The entries are short (five to ten minutes each) and structured. The goal isn't to write beautifully - it's to lock in your thinking before hindsight rewrites it.
The format you use matters less than the discipline of using one. A paper notebook, a spreadsheet, a Notion database, or a dedicated app all work. What matters is consistency: every meaningful decision (career moves, big purchases, investments, business choices, important relationships) gets an entry, and every entry is reviewed at a defined point in the future.
The technique is most associated with Annie Duke, whose book Thinking in Bets is the modern reference. The deeper roots run through the calibration-training research of Phil Tetlock and the Good Judgment Project, and through the financial-decision writing of Peter Bevelin and Charlie Munger.
## Why outcome-based learning is brokenImagine watching a poker player call a big bet on a marginal hand. They get lucky and win the pot. Did they make a good decision? You can't tell from the outcome of one hand. Across ten thousand hands, the win rate of a strategy is data. Across one hand, it's noise.
Most life decisions are more like one poker hand than ten thousand. You don't accept the same job ten thousand times to find out whether the move was sound. You accept it once, the world's randomness picks an outcome from the distribution of things that could have happened, and that's the only data point you ever get. Judging the decision by that single outcome is exactly as misleading as judging a poker strategy by one hand.
Three forces conspire to make this worse:
Without a record taken before the outcome was known, you have no way to separate the contribution of your reasoning from the contribution of the world's randomness. The journal is the record.
## What to record: the six fieldsEach entry has six fields. Most take a sentence; the whole entry should fit on one screen.
The confidence number is the most important field. Without it, every 'I sort of expected that' becomes 'I knew that all along'. With it, you can compare your stated confidence to base rates over time and see exactly where you're overconfident - the same calibration loop that superforecasters use to push their accuracy ahead of subject-matter experts.
## The review cadenceReviews are where the value lands. Without them, the journal is data collection without analysis.
At each review, ask three questions:
You're judging the decision, not the outcome. A decision that produced a bad outcome can still have been a good decision; the question is whether the reasoning was sound given what you knew at the time. Positive-expected-value choices still lose half the time at 50% confidence - that's the maths, not the failure.
## A worked exampleTake a concrete decision: should you accept a new role at £15k more salary, with a longer commute and less-interesting work?
Decision: Accept the new role. Alternatives considered: stay (counter-offer not requested), use the offer to renegotiate, decline both.
Reasoning: Net £15k pre-tax improves my savings rate from 18% to 26%. Less-interesting work is recoverable; the pay gap closes career-pivot regret faster than it accumulates. The new firm is in a more stable industry segment.
Assumptions: New role's stability matches old. Commute won't materially affect health or relationships. No recession in the next 12 months. Manager is a normal-distribution manager (didn't check references).
Expected outcome and confidence: 70% confident I'm still in the role at 18 months. 60% confident I'll consider the move worth it at 24 months.
What would change my mind: Counter-offer of £10k+ from old job. Specific bad signal on new manager from a trusted reference.
Review date: 18 months from start.
At the 18-month review you write the outcome alongside the original: still in the role, savings progress matched the prediction, but the commute was a bigger drag than predicted (only 50% retrospectively confident at the 18-month mark). Both predictions were broadly correct - calibration roughly right. With the same information today you'd make the same call, but you'd flag 'manager reputation' as a load-bearing assumption to actually verify next time. That last sentence - the one calibration entry - is the practice's whole compounding return.
## Common mistakesThe practice is simple to describe and easy to do badly. The patterns that ruin it:
Don't over-engineer the tool. Three options that all work:
The medium is irrelevant if the discipline isn't there. Pick the one you'll actually use this week and stop optimising.
## Where it fits with other thinking toolsThe journal is a calibration tool. It pairs naturally with:
The journal alone doesn't make you a better decision-maker overnight. What it does is build a feedback loop between your reasoning and your reality, the absence of which is the reason most people don't improve at decisions over a lifetime.
## Frequently asked questions **Q: How often should I write entries?** A weekly average is plenty for most people. The point is volume of calibration data, not exhaustive coverage. Better to log 50 small decisions a year than to demand of yourself that you log every choice - the over-strict version dies inside a fortnight. **Q: Will my memory of past decisions improve?** Calibration improves; raw memory does not. The journal is a tool against hindsight bias, not a tool against forgetting. The compounding return comes from the calibration loop, not from recall. **Q: How long does each entry take?** Five to ten minutes if you keep to the six-field format. Longer entries are usually a sign you're rationalising - strip back to: decision, reasoning, assumptions, expected outcome and confidence, what would change my mind, review date. **Q: Should I share my journal with anyone?** No. The value depends on candour, and candour requires that you're not performing for an audience. The only acceptable audience is your future self. **Q: Can teams use a decision journal?** Yes - group decisions are well-served by team pre-mortems and a shared decision log. The personal-calibration value is reduced (you can't tell whose reasoning was correct in a group), but the institutional record is even more valuable. Many investing firms maintain shared decision logs for exactly this reason. **Q: What if I don't trust my own confidence numbers?** That distrust is the calibration training kicking in. Write the number anyway. After a year of entries, you'll know whether your 70%-confidents come true 70% of the time, 50%, or 90%. Once you know your bias, you can correct for it - and that correction is impossible without the recorded numbers to compare against. **Bottom line** Six fields, five minutes per entry, scheduled review. Confidence numbers are mandatory. Don't edit past entries; never review early. The compounding return is calibration - knowing the gap between what you think you know and what you actually know - and over a lifetime there is no other tool that produces it. --- # Framing Effect: Why How You Hear It Changes Your Choice Source: https://expectedvalue.co.uk/blog/framing-effect-bias/ # Framing Effect: Why How You Hear It Changes Your Choice The framing effect: identical numbers presented differently produce opposite decisions. Asian Disease Problem, gain vs loss framing, defences. **Answer** The framing effect is a cognitive bias where identical information produces opposite decisions depending on how it is presented. Most famously: a treatment described as having a 90 percent survival rate is chosen more often than the same treatment described as having a 10 percent mortality rate. The effect is one of the most reliably replicated findings in behavioural economics. The defence is mechanical, not motivational - translate gain-framed and loss-framed versions of the same decision into base rates before choosing. ## What is the framing effect?The framing effect describes a systematic bias in human decision-making first formalised by Daniel Kahneman (a Princeton psychologist who won the 2002 Nobel Prize in Economic Sciences for his work on judgement under uncertainty) and Amos Tversky in a 1981 paper in Science. The finding: when the same underlying choice is described in different terms, people change their preferences in ways that violate basic rationality.
The mechanism sits inside prospect theory (a Kahneman-Tversky model of how people evaluate gains and losses asymmetrically). Loss-framed information triggers risk-seeking behaviour - people will gamble to avoid a guaranteed loss. Gain-framed information triggers risk-aversion - people accept a smaller certain gain over a larger uncertain one. The two responses are predictable, near-universal, and survive even when participants know the framing-effect literature.
## What is the Asian Disease Problem?The canonical demonstration is the Asian Disease Problem, run by Tversky and Kahneman in 1981. Imagine an unusual outbreak is expected to kill 600 people. Two response plans are proposed.
**Frame 1 - gains** Program A: 200 people will be saved.Programs A and C are identical (save 200 = let 400 die). Programs B and D are identical (1/3 chance of saving all = 1/3 chance nobody dies). The decision under the two framings reverses entirely - the gain-frame produces a strong preference for the certain option; the loss-frame produces a strong preference for the gamble. The shift was reliable across student samples and replicated in physician samples making medical-policy choices.
## Where does the framing effect show up in real decisions?Three domains where the effect is large enough to change observed behaviour.
- Medical decisions. Surgical consent rates change dramatically depending on whether the surgeon describes a procedure as having a 90 percent five-year survival rate or a 10 percent five-year mortality rate. McNeil et al. 1982 in the New England Journal of Medicine documented this in patients choosing between surgery and radiation for lung cancer. The clinical implication: how informed-consent forms phrase risk literally changes which treatment patients pick. - Investment and insurance. The same volatility number framed as 'historically returns 8 percent over 20 years' versus 'has lost 30 percent in 1 in 8 years' produces very different allocation choices. The 1 in 8 framing makes risk concrete in a way the 20-year geometric-mean framing hides. The framing effect is one reason why investors who agreed in writing to a 60/40 portfolio under one framing repeatedly capitulate at the bottom of drawdowns where the loss-frame becomes salient. Our loss aversion guide covers the related Kahneman-Tversky finding that losses feel roughly twice as bad as equivalent gains feel good. - Marketing and pricing. 'Save £20 by switching today' (gain frame) and 'You're overpaying £20 by staying' (loss frame) describe the same transaction; the loss frame consistently generates higher switch rates. Behavioural-economics-influenced UK regulators (FCA, Ofcom) now require certain consumer-finance disclosures to surface the loss frame explicitly, on the basis that the gain frame's lower urgency contributes to consumer inertia in markets with frequent better deals available. ## Why does the framing effect persist?Three reinforcing mechanisms keep the effect robust against introspection.
- The asymmetry of gains and losses. Prospect theory's central claim is that the psychological impact of a £100 loss is roughly twice the impact of a £100 gain. A frame that highlights losses therefore triggers a much stronger affective response than a numerically identical gain-frame. The brain is not solving for expected value; it is solving for a value function with a kink at the reference point. - The reference point is sticky. Once you anchor on 'I currently have X', losing to 'X minus Y' feels qualitatively different from gaining to 'X plus Y'. Reframing means moving the reference point, which is cognitively costly and rarely happens without prompting. The classic anchoring bias compounds the framing effect by making the reference point even harder to dislodge. - System 1 produces the answer before System 2 audits. Kahneman's two-system framing in Thinking, Fast and Slow applies here: the gut response to a loss-frame ('avoid the loss!') arrives before the slow, deliberative system has time to translate gain-frame and loss-frame into base rates. By the time System 2 catches up, the preference has formed. ## How do you defend against the framing effect?Awareness is not enough. Tversky and Kahneman tested whether telling participants about the bias before showing them framed problems reduced its size; the size dropped slightly but did not disappear. Practical defences are mechanical rather than motivational.
1. **Convert frames to base rates** - Whenever you see a percentage, translate the inverse before deciding. '90 percent survival' → 'so 10 percent mortality'. '70 percent satisfied' → 'so 30 percent dissatisfied'. Both numbers now coexist in working memory and the asymmetry between gain-frame and loss-frame collapses to a single base rate. 2. **Re-state the decision in both frames** - Before committing, force yourself to articulate the choice once in the language of what you gain and once in the language of what you lose. If the two articulations point at different choices, the framing effect is active and the choice is unstable - reach for the underlying expected value before deciding. 3. **Trust written rules under pressure** - Loss-framed information is most powerful when emotions are high (mid-drawdown, bad medical news, regulatory deadline). Pre-commit to a rule under calm conditions ('I rebalance my portfolio quarterly regardless of recent returns', 'I read both the 90 percent and 10 percent versions of any medical statistic before signing'), then defer to the rule when the loss frame surfaces. 4. **Watch for the asymmetric framing in pitches against you** - Sales, regulatory and political communications often deliberately exploit the framing effect. A 'limited time offer' is a loss frame engineered to push you toward a sub-optimal certain outcome; a 'pay 30p more per month' is a gain frame engineered to obscure the per-year cost. Reading both the gain-frame and loss-frame versions of any pitch you receive surfaces the asymmetry. 5. **Use a decision journal** - Writing down the choice, the framing you noticed, and the base-rate translation in our decision journal creates a paper trail that retrospectively shows whether your decisions skew toward the gain-frame or loss-frame answer over time. The pattern, once visible, is the most reliable nudge toward base-rate thinking. ## Frequently asked questions **Q: What is the framing effect in psychology?** The framing effect is a cognitive bias where the same information produces different decisions depending on how it is presented - typically as a gain or as a loss. Formalised by Tversky and Kahneman in 1981 as part of prospect theory. The most reliable demonstration is the Asian Disease Problem, where identical outcomes described in lives-saved versus lives-lost terms produce opposite preferences. **Q: Is the framing effect the same as anchoring?** No. Both are cognitive biases that distort decisions away from rational expected-value calculations, but they operate differently. Anchoring is the tendency to over-weight the first number you see; framing is the tendency to switch preferences based on whether information is presented as a gain or a loss. Our anchoring bias guide covers anchoring in detail; they often appear together but the defences are different. **Q: How is the framing effect related to loss aversion?** Loss aversion is the deeper finding - that losses feel approximately twice as painful as equivalent gains feel good. The framing effect is the practical consequence: because losses outweigh gains, a frame that highlights losses (a 10 percent mortality rate) triggers a stronger response than a numerically identical gain-frame (a 90 percent survival rate). See our loss aversion explainer for the underlying asymmetry. **Q: Can awareness of the framing effect eliminate it?** Largely no. Tversky and Kahneman tested this directly - explicitly warning participants about the framing effect before showing them framed problems reduced the size of the effect somewhat but did not eliminate it. The robust defences are mechanical (translating gain-frames and loss-frames into base rates) rather than motivational (trying to be more rational). **Q: Where is the framing effect deliberately used against me?** Sales copy frequently uses loss-frames to drive urgency ('limited time', 'don't miss out'). Marketing frequently uses gain-frames to obscure full costs ('only 30p more per month' instead of '£3.60 more per year'). Insurance pitches lean heavily on loss-frame language. Political campaigns alternate frames depending on the issue. Reading both the gain-frame and loss-frame of any pitch is the cheapest defensive habit you can develop. ## Related guides **Editorial note** This article is a research-led editorial explainer drawing on Tversky and Kahneman's original 1981 paper in Science, McNeil et al.'s 1982 study in the New England Journal of Medicine on medical-decision framing (PubMed 7078603), and Kahneman's Thinking, Fast and Slow. Examples are illustrative; we cite the original studies for the headline findings. --- # Endowment Effect Explained: Why You Overvalue What You Own Source: https://expectedvalue.co.uk/blog/endowment-effect/ # Endowment Effect Explained: Why You Overvalue What You Own The endowment effect makes you value the things you own more highly than identical things you do not. Here is why it happens, with worked examples. **Answer** The endowment effect is the tendency to value something you already own more highly than an identical item you do not own. Richard Thaler showed in 1990 that Cornell students given a coffee mug demanded roughly twice the price to sell it as students without one would pay to buy it.Hand someone a coffee mug, give them a few minutes to hold it, then ask what they would sell it for. They will ask for about double what a comparable buyer offers for the same mug seconds later. This is the endowment effect, and it shapes prices, portfolios, and break-ups in ways most people never notice.
## What is the endowment effect?The endowment effect is a cognitive bias first named by economist Richard Thaler (a University of Chicago economist who won the 2017 Nobel Memorial Prize for his work on behavioural economics). It describes a stubborn gap between the price an owner will accept to part with a possession and the price a non-owner will pay to acquire the same thing.
In the standard formulation, the minimum price an owner will accept (Willingness To Accept, or WTA) tends to run two to three times the maximum price a buyer will pay (Willingness To Pay, or WTP) for the same item. Classical economics says these two numbers should match - if a mug is worth £5 to you, it should not matter whether you currently own it or not. Endowment effect studies show that classical assumption breaks the moment ownership enters the picture.
The effect is robust across goods that are not investments, traded in laboratory and field settings, and visible from undergraduate experiments through to multi-million-pound housing markets. The full endowment effect entry on Wikipedia tracks the literature back to Thaler's 1980 paper and forward through more than four decades of replications.
## Where did the endowment effect come from?Thaler introduced the term in a 1980 paper titled Toward a positive theory of consumer choice. He was working through anomalies that the standard economic model could not explain: people refusing to sell tickets at face-value multiples, jurors awarding higher compensation for losses than gains, and a hundred small everyday observations that all pointed the same way.
The intellectual scaffolding came from prospect theory, developed by Daniel Kahneman and Amos Tversky (Israeli-American psychologists whose 1979 prospect theory paper underpins most of modern behavioural economics) in their landmark 1979 paper. Prospect theory says people evaluate outcomes relative to a reference point, and losses below that reference point hurt about twice as much as equivalent gains feel good - a phenomenon called loss aversion.
Ownership shifts the reference point. Once you own the mug, parting with it registers in your mind as a loss. Buying the same mug registers as a gain. Loss aversion then makes the loss feel bigger than the equivalent gain, so the price you require to give it up is higher than the price a stranger would pay to receive it. The endowment effect is essentially loss aversion measured at the moment ownership changes hands.
## What did Thaler's coffee mug experiments show?The most-cited demonstration of the endowment effect is the Cornell coffee mug study, run by Daniel Kahneman, Jack Knetsch, and Richard Thaler in 1990. It is worth walking through because it is so simple that it is hard to argue with.
Twenty-two students were given a Cornell University mug worth about $6 at the campus shop. Another twenty-two students got nothing. The owners were asked the lowest price they would accept to sell their mug. The non-owners were asked the highest price they would pay to buy one. If both groups valued the mug at its market price of $6, the trade prices should have clustered near that number.
They did not. The median selling price was $5.25. The median buying price was $2.75. The owners demanded almost double what the buyers would pay. Trades that classical economics predicted would happen did not happen, because the two sides of the market could not agree on a price for an identical mug whose only difference was who was holding it.
The experiment has been replicated with chocolate bars, pens, lottery tickets, and at considerably larger stakes. The size of the gap varies. The direction does not.
## Why do we overvalue what we own?Three mechanisms account for most of the effect, and they tend to compound rather than substitute for each other.
Loss aversion is the dominant driver. Once ownership is established, the prospect of parting with the object is coded as a loss rather than a forgone gain. Because losses register at roughly 2x the intensity of equivalent gains in prospect theory, the price required to make the seller whole is higher than the price the buyer needs to feel they got value.
Ownership signals identity and status. The thing on your desk is not just a mug. It is your mug. It has a place in your kitchen and a small role in your routine. Selling it requires you to renegotiate that small piece of your identity. Even brief, arbitrary ownership of a few minutes is enough to produce the effect in laboratory settings, but longer ownership and emotional attachment amplify it.
The reference point shifts when ownership transfers. Once you have something, your mental baseline updates to include it. Returning to the prior state, without the object, now feels like moving backward. Buyers do not have this problem - their reference point is the state without the object, so acquiring it is neutral or pleasant rather than a loss reversal.
## How does the endowment effect show up in real life?The lab effect generalises. Anywhere people own something and need to value it for sale or exchange, expect a gap between owner valuations and market reality.
Property sellers anchor on purchase price. Homeowners who bought during a peak market routinely refuse to sell below their original price even when comparable sales clearly show the market has moved. The original price acts as a psychological reference point, and selling for less feels like a loss the seller refuses to accept. Estate agents call these properties "sticky listings", and the academic literature on housing markets documents the pattern at scale.
Investors hold losing positions too long. A share bought at £100 and trading at £60 feels like a loss the investor needs to recover. The same investor would not buy fresh shares of the company at £60 today, but they will not sell either. The purchase price becomes the reference point and the position becomes endowed. Daniel Kahneman (the Princeton psychologist who won the 2002 Nobel Memorial Prize for prospect theory) covered this pattern in Thinking, Fast and Slow, where it overlaps with the sunk cost fallacy and disposition effect.
Workplaces overvalue their own decisions. Once a team has committed to a strategy, vendor, or project, they tend to defend it against superior alternatives. The strategy is now "ours", and abandoning it codes as a loss of effort and identity rather than a rational pivot. This is one reason pre-mortems and structured devil's-advocate exercises work - they force the team to evaluate the current path as if they did not yet own it.
Free trials are designed to trigger the effect. Streaming services, software vendors, and gyms offer free trials because giving you the product for a month creates ownership. By the time the trial ends, cancellation feels like a loss of access to something you already have. The conversion rates are far higher than they would be if the same vendor simply asked you to buy at the start.
## What is the IKEA effect and how is it related?The IKEA effect is a close cousin of the endowment effect, identified by Michael Norton, Daniel Mochon, and Dan Ariely in a 2012 paper. It says that people value objects they have partially built more highly than identical pre-built versions, in proportion to the effort they put in.
Norton's team had participants assemble IKEA boxes, fold origami, or build Lego sets, then asked them to bid against non-builders. Builders bid 63% more than buyers for the same finished object. The bid gap held up even when the builders were objectively worse at the assembly than experienced craftspeople. Effort, not skill, produced the overvaluation.
The IKEA effect adds a complication to the standard endowment story. Ownership matters, but earned ownership matters more. This is why custom-built kitchens get defended past their useful life, why founders cannot bring themselves to pivot away from a flagship product, and why home-renovation budgets always blow out - the partly-finished project is endowed by the effort already poured into it.
**Pro tip** If you cannot honestly say you would buy this position, asset, or feature at today's price with fresh money, the endowment effect is probably the only reason you still hold it. The fresh-money test is the simplest practical counter-bias you can run. ## How can you counter the endowment effect?You cannot disable a cognitive bias by knowing about it. The asymmetric pull between losses and gains is wired into the same neural circuitry that helped your ancestors avoid predators. What you can do is build a small set of decision rules that force you out of the owner frame and into the buyer frame before you commit.
- **Run the fresh-money test.** - Ask whether you would acquire this position, possession, or commitment at today's price using money or time you do not yet have allocated. If the answer is no, you are holding it because you own it, not because it is the best use of the resource. - **Separate the decision to hold from the decision to buy.** - Treat the choice to keep something as a fresh purchase decision. This is how rebalancing rules and stop-losses work in investing - they break the link between purchase price and current decision. - **Use a pre-commitment rule.** - Decide the conditions under which you will sell or exit before you take ownership. Pre-commitment ties the future you, who will be in the owner frame, to the past you, who could still think like a buyer. - **Get an outside view.** - Show the situation to someone who has no stake. Their valuation is closer to the unowned reference point. If they would not pay what you are demanding, you are anchored to ownership, not to value. - **Use blind comparison.** - Write the asking price for your item next to the prices of three comparable items currently available to buy. If your number is the outlier, the gap is the endowment effect, not the market. ## Where does the endowment effect end?The effect is not universal. Two consistent boundaries show up in the literature.
It does not apply, or applies much more weakly, to items held for resale. Traders, market makers, and shopkeepers value inventory closer to market price because their reference point was never "I own this", it was "this is for sale". List Vol's 2003 trading-card market experiments at sports memorabilia conventions showed experienced traders displaying no endowment effect on the trading items themselves, even as inexperienced consumers showed the standard gap.
It also fades when the item is fungible cash or near-cash. People do not become more reluctant to spend a specific £10 note than to spend any other £10. The bias hooks into objects with identity, not abstract value.
Both boundaries point at the same underlying mechanism: the endowment effect is a feature of how owners code identity and reference points around specific objects. Strip out the identity attachment, or trade so often that ownership is never your reference frame, and the gap shrinks toward zero.
## Frequently asked questions **Q: Is the endowment effect the same as loss aversion?** They are closely related but not identical. Loss aversion is the broader principle that losses feel about twice as bad as equivalent gains feel good. The endowment effect is what happens when you apply loss aversion to the moment of parting with something you own - the prospect of losing the item triggers the asymmetric weighting and inflates the owner's required price. **Q: Does the endowment effect apply to digital goods?** Yes, and with some surprising twists. Studies on virtual currencies, digital downloads, and game items find the same WTA-WTP gap. Subscription cancellation friction works in part because months of access have endowed users with the service, even though nothing physical changes hands. **Q: How big is the endowment effect in money terms?** In laboratory experiments, selling prices typically run 1.5x to 3x buying prices for the same low-value item. Field studies on property, cars, and used goods show smaller but persistent gaps. The effect scales nonlinearly - very high-value items often show smaller percentage gaps because real money concentrates the mind, while low-stakes items can show very wide gaps because the decision is mostly emotional. **Q: Do experts show the endowment effect?** Less so, but only within their domain. Professional traders show muted endowment effects on the items they trade. Outside their expertise they show the same biases as everyone else. Expertise reduces the bias by reframing ownership from identity to inventory, but it does not generalise. **Q: Is the endowment effect a problem or a feature?** It depends on the decision. For long-term commitments like marriages, careers, and homes, a small overvaluation of what you already have can support useful stability. For short-term economic decisions like selling a depreciating asset or cancelling a poorly-fitting subscription, the same bias becomes expensive. The decision rules in the previous section are designed for the second case, not the first. ## Related guides --- # Negative Expected Value: Why Lottery Tickets Always Lose Source: https://expectedvalue.co.uk/blog/negative-expected-value/ # Negative Expected Value: Why Lottery Tickets Always Lose Why every lottery ticket, roulette spin and slot pull is mathematically a loss in expectation. The maths behind why the house always wins. **Answer** A negative expected value (negative EV) bet is one where the average outcome, summed over a long series of identical bets, is a loss. A £2 UK Lotto ticket has an expected value of roughly £0.95 - every £1 spent returns ~47.5p on average. Every casino game (except advantage-play blackjack and a few exotic edge cases) is structurally negative EV by design - that mathematical edge is how casinos exist.Casinos, bookmakers and lottery operators are not in the business of luck. They are in the business of selling small chances of large wins at prices slightly above what those chances are mathematically worth. The gap between the price and the mathematical value is called the house edge - and across millions of customers it compounds into reliably enormous profits.
This post covers the maths behind every common gambling product: UK National Lottery, scratch cards, roulette, slot machines, and sports betting. It builds on our Expected Value Explained pillar, uses the same notation as our Expected Value Formula walkthrough, and gives you a calculator-free way to spot negative-EV products in everyday life.
## What does negative expected value mean?The expected value of a bet is the probability-weighted average of all possible outcomes. In symbols:
EV = Σ (probability × outcome)
A bet is positive EV when the sum of those probability-weighted outcomes is greater than what you paid to play it. It is negative EV when the sum is less than you paid. A fair bet - exactly zero expected profit - is zero EV.
Concrete example. Imagine a fair coin flip where heads pays £1 and tails costs you £1.
Now imagine the same coin flip, but heads pays only 90p and tails still costs £1.
That negative 5p per flip is the house edge. Play this game 1,000 times and you expect to lose £50. Play it 100,000 times and you expect to lose £5,000. The maths is silent at any individual flip; the law of large numbers makes it loud across the bookmaker's whole customer base.
## How does the UK National Lottery have negative EV?Camelot's UK Lotto returns approximately 47.5% of revenue to prizes. The Lotto Plus and EuroMillions products have similar shapes. That ratio tells you most of what you need.
A £2 Lotto ticket has an approximate expected value of £0.95 - meaning you can expect to lose 52.5p, on average, every ticket you buy. The split of the missing £1.05 typically breaks down to roughly:
That structure is non-negotiable for a non-rollover week. The jackpot probability for a single UK Lotto ticket is roughly 1 in 45 million. Multiplying that by the headline jackpot and adding the expected values of the lower tiers gets you back to that ~£0.95 figure. Rollovers shift the maths because uncollected jackpots compound - see the rollover section below.
EuroMillions has even longer odds (around 1 in 139.8 million for the main jackpot) and a similar overall return rate. Scratch cards typically return around 65% of their ticket price - better than Lotto on a per-ticket EV basis, but still negative.
## Why does roulette always favour the house?European roulette has 37 numbers - 18 red, 18 black, and a single green zero. Betting £1 on red pays £1 if a red number comes up. There are 18 winning slots and 19 losing slots.
That single green zero is the casino's structural edge: 2.70% on every spin on red, black, odd, even, and the column bets. American roulette has two green zeros (0 and 00), pushing the edge to 5.26%. Casinos make far more from American roulette than European roulette for the same nominal stake - which is why American wheels remain common in the United States despite the player-side maths being objectively worse.
The roulette edge is one of the most honest in gambling: the operator tells you exactly what it is, the odds are visible on the wheel itself, and no skill on the player's side can overcome the green zero. Across an evening of 100 spins at £10 per spin, you expect to lose £27 on European roulette and £52.60 on American roulette. Variance dominates any single session; the edge dominates across many sessions.
## How much do slot machines actually take?Slot machines disclose their return-to-player (RTP) percentage - the long-run share of total stakes returned to players in winnings. Typical RTPs in the UK in 2026:
RTP is calculated over the lifetime of the machine, not your individual session. A slot with 96% RTP returns £96 to players for every £100 staked in expectation, but the variance is enormous - most players walk away losing far more than 4%, while a small minority walk away with significant wins. The mean of those individual outcomes matches the published RTP. The UK Gambling Commission requires that licensed operators disclose RTP on online slots and that all UK casinos use random number generators tested to a specification published by the regulator.
The key arithmetic: a machine running at 96% RTP has an expected loss of 4% of total turnover, not 4% of your initial bankroll. If you wager £100 ten times in a session (re-betting winnings), your total turnover is £1,000, and your expected loss is £40 - not £4. This is why session bankroll figures and stake totals diverge so quickly on slot play.
## Why is sports betting also negative EV (the vig)?Sports bookmakers don't take a fixed percentage like a casino. They build their edge into the prices themselves - the technical term is vigorish (US shorthand: vig) or overround (UK).
Take a coin-flip-style market like Manchester United vs Chelsea where the true probabilities are 50/50. A fair price would be 2.00 on each side (decimal odds, equivalent to even money). What you actually see is more like:
Implied probabilities: 1/1.91 = 52.4% on each side. The sum is 104.8% - that 4.8% overround is the bookmaker's structural edge. On a fair market that ought to sum to 100%, you're being asked to pay for 104.8% worth of probability.
That 4.8% is the expected loss per pound of turnover, evenly distributed across the two sides. Whichever side wins, the bookmaker keeps roughly £4.80 of every £100 staked across all customers. High-overround markets (correct score, exotic bet types) often run 15-25% - those are the highest-EV bets for the bookmaker and the lowest-EV bets for you.
## Does positive EV gambling exist?Mathematically, yes - but it's rarer than gambling marketing suggests.
For ordinary consumers, lottery tickets, casino games and standard sports betting are all structurally negative EV by design. The 1% of profitable participants are specialists. The 99% are paying for entertainment, which is a defensible reason to play - as long as you're honest with yourself that the price is negative EV in expectation.
## When is buying a negative EV product still rational?Negative EV does not always mean irrational. Insurance is the cleanest example.
Insurance is structurally negative EV - premiums must exceed expected payouts for insurers to stay in business. A typical UK car insurance policy returns roughly 60-75 pence per pound of premium to policyholders in claim payouts. That makes it a worse EV product than a UK Lotto ticket in pure expectation.
And yet buying insurance is rational for most people most of the time. The reason is the shape of the loss. A £500 annual premium for car insurance protects against a £20,000 write-off that would devastate household finances. The utility (not just the expected value) of avoiding catastrophic losses is much greater than the utility lost from paying the premium. Our Insurance Companies Use Probability deep-dive walks through why insurance can be rationally bought even when its expected value is negative - risk aversion and the diminishing marginal utility of money are the cleanest formal explanations.
Lottery tickets occupy a different category. They are not protecting against catastrophic loss; they are buying the consumer experience of imagining a positive outcome. Some people get £2 worth of pre-draw daydreaming value from a Lotto ticket. That's a defensible purchase, even though its expected monetary value is −52.5p. The mistake comes when buyers convince themselves they are making a financial investment rather than a small entertainment purchase.
## How do I calculate the EV of any bet?The general workflow:
For complex markets where you can't estimate true probabilities directly, our EV calculator tool handles the arithmetic. Where you can estimate true probabilities (lottery jackpots, roulette spins, simple coin-flip markets), the back-of-envelope multiplication is enough. Either way the test for a positive-EV consumer gambling product is straightforward: EV ≥ cost. If you can't find an angle that satisfies that inequality, you are buying entertainment.
## Frequently asked questions **Q: Why is the UK National Lottery expected value approximately £0.95 on a £2 ticket?** Camelot returns approximately 47.5% of UK Lotto revenue to prizes. £2 × 47.5% = £0.95. The remaining 52.5p covers good causes, lottery duty, retailer commission and Camelot's margin. The same approximate ratio applies across UK Lotto, EuroMillions and Set For Life under the current operating licence. **Q: Does buying more lottery tickets improve my expected return?** No. Each additional ticket has the same approximate £0.95 expected value on a £2 stake. The probability of winning the jackpot scales linearly with the number of tickets, but so does your spend. You can never reach positive expected value by simply buying more tickets at the same draw - the only way is a positive-EV rollover situation, and those are rare and quickly arbitraged. **Q: Is European roulette better than American roulette?** Yes, structurally and unambiguously. European roulette has a single green zero (2.70% house edge), while American roulette has both a 0 and a 00 (5.26% house edge). For any nominal stake, your expected loss is roughly half on European wheels. If both are available in the same venue, the European wheel is the strictly dominant choice. **Q: What's a fair return-to-player percentage for a UK slot machine?** Online slots in the UK typically advertise 95-97% RTP, which is the best slot category by player return. Pub fruit machines (Cat-C) run 70-80% RTP - significantly worse. The UK Gambling Commission requires RTP disclosure on online slots; check it before playing. Note that RTP is calculated over the lifetime of the machine, not your session, so individual sessions can vary widely from the published number. **Q: Can matched betting really make money against the bookmakers?** Yes, but with caveats. Matched betting exploits bookmaker free-bet promotions and the price differences between a bookmaker and a betting exchange. The maths can lock in a positive EV per promotion. The constraints are practical: most accounts get restricted or banned within months for consistent value extraction (bookmakers identify the pattern), and the time investment per pound earned tends to scale poorly. It is a real positive-EV opportunity for some users, but not a scalable income source for most. --- # Probability vs Odds: The Difference and Why It Matters Source: https://expectedvalue.co.uk/blog/probability-vs-odds/ # Probability vs Odds: The Difference and Why It Matters Probability and odds describe the same uncertainty differently - and confusing them costs people money. How each works, conversions, and bookmaker tricks. ## What is the difference between probability and odds? **Answer** Probability and odds are two ways of expressing the same uncertainty. Probability is the fraction of outcomes where the event happens: 0.25 means 25% of cases. Odds are the ratio of favourable to unfavourable: 1:3 odds means one win for every three losses (the same 25% probability). The relationships: probability = favourable / (favourable + unfavourable); odds = probability / (1 - probability). Bookmakers quote odds, statisticians use probability, and logistic regression coefficients work in log-odds (sometimes called logits). ## Bookmaker margin calculator Convert a two-way market’s odds into the bookmaker’s built-in margin. A fair book would total 100% implied probability - anything above that is the cut that, on average, the bettor pays. Interactive calculator. Computes: The bookmaker’s margin (over-round); Implied probability total (a fair book = 100%).Probability is the share of all possible outcomes in which an event occurs. It is a number between 0 and 1, often expressed as a percentage. If a horse has a 25% probability of winning, that means in 25 out of 100 equally likely scenarios, this horse crosses the line first.
Odds are the ratio of the ways an event can happen to the ways it cannot. The same horse with a 25% probability has odds of 1-to-3 against winning - for every 1 scenario where it wins, there are 3 scenarios where it loses.
The two numbers describe identical uncertainty. They are simply different mappings of the same underlying truth. But the choice of language matters: probability is the natural unit for thinking about chance, while odds are the natural unit for pricing it. That is why mathematicians, scientists and forecasters speak in probabilities, and why bookmakers, prediction markets and poker players speak in odds.
Learning to translate between probability vs odds - quickly, in your head, across all three odds formats - is one of the highest-leverage skills in betting, investing and any decision involving uncertainty. This guide shows you exactly how.
## The Formula That Connects ThemThere are only two formulas you need to memorise, and they are inverses of each other.
Probability to odds (against):
Odds against = (1 − P) / P
Odds (against) to probability:
P = 1 / (1 + Odds against)
Let us run a quick check. Suppose a coin flip has P = 0.5. Plugging in: (1 − 0.5) / 0.5 = 1, so the odds are 1-to-1, often written as evens. Reverse: 1 / (1 + 1) = 0.5. The maths reconciles.
Now take a probability of 0.2 (a 20% chance). Odds against = (1 − 0.2) / 0.2 = 4. So the odds are 4-to-1 against. Reverse: 1 / (1 + 4) = 0.2. Confirmed.
A subtle but important point: people use odds against and odds in favour almost interchangeably, but they are reciprocals. Odds in favour of an event are P / (1 − P). For our 20% horse: 0.2 / 0.8 = 0.25, or 1-to-4 in favour - the same uncertainty seen from the opposite direction. In gambling, odds almost always means odds against unless explicitly stated otherwise. We will follow that convention for the rest of this guide.
## The Three Odds Formats You Will Encounter Walk into a betting shop in London, open an Australian sportsbook, and visit a US-facing site, and you will see three completely different ways of writing the same odds. The format you see depends almost entirely on where the operator is based and which audience they serve. Understanding all three is essential because mixing them up leads to expensive mistakes. Each format encodes two pieces of information: the implied probability of the outcome, and the payout you receive if you win. They differ only in how that information is presented. - **Decimal odds (Europe, Australia)**: Total return per £1 staked, including the stake. 2.00 means you get £2 back from a £1 bet. - **Fractional odds (UK, Ireland)**: Profit-to-stake ratio. 5/1 means £5 profit for every £1 staked, plus your £1 back. - **American odds (USA)**: Positive numbers show profit on £100 stake. Negative numbers show stake required for £100 profit. +200 = £200 profit on £100. -150 = £150 stake to win £100. ## The Master Conversion TableBookmark this table. It contains the most common reference points you will encounter and is the single fastest way to build intuition for probability vs odds across formats. The implied probability column is what each set of odds is really saying about the event.
- **90% probability**: Decimal 1.11 · Fractional 1/9 · American -900 - heavy favourite (e.g. top seed in early-round match) - **80% probability**: Decimal 1.25 · Fractional 1/4 · American -400 - strong favourite - **75% probability**: Decimal 1.33 · Fractional 1/3 · American -300 - clear favourite - **66.7% probability**: Decimal 1.50 · Fractional 1/2 · American -200 - solid favourite - **60% probability**: Decimal 1.67 · Fractional 4/6 · American -150 - moderate favourite - **50% probability**: Decimal 2.00 · Fractional 1/1 (evens) · American +100 - true coin flip - **40% probability**: Decimal 2.50 · Fractional 6/4 · American +150 - moderate underdog - **33.3% probability**: Decimal 3.00 · Fractional 2/1 · American +200 - clear underdog - **25% probability**: Decimal 4.00 · Fractional 3/1 · American +300 - outsider - **20% probability**: Decimal 5.00 · Fractional 4/1 · American +400 - long shot - **10% probability**: Decimal 10.00 · Fractional 9/1 · American +900 - heavy long shot - **5% probability**: Decimal 20.00 · Fractional 19/1 · American +1900 - very heavy long shot ## How to Convert Each Format Into Implied ProbabilityImplied probability is the probability of the outcome that the odds are claiming. If the implied probability of all outcomes in a market sums to more than 100%, the bookmaker has built in a margin (the overround or vig). Spotting that gap is the first step to identifying value bets.
From decimal odds:
Implied probability = 1 / decimal odds
Decimal odds of 2.50 imply a probability of 1 / 2.50 = 0.40, or 40%.
From fractional odds (a/b):
Implied probability = b / (a + b)
Fractional odds of 6/4 imply 4 / (6 + 4) = 4 / 10 = 0.40, or 40%. Same outcome, same probability - just written differently.
From American odds:
For positive American odds: Implied probability = 100 / (American odds + 100). So +150 implies 100 / 250 = 0.40, or 40%.
For negative American odds: Implied probability = |American odds| / (|American odds| + 100). So -200 implies 200 / 300 = 0.667, or 66.7%.
The rule of thumb: positive American odds belong to underdogs (probability under 50%), negative belong to favourites (probability above 50%), and the size of the number reflects how strongly the market believes in the outcome.
**Quick mental shortcut** For decimal odds, just divide 1 by the number. For American odds, the bigger the number (in either direction), the more lopsided the market is. -1000 means the market is 91% sure. +1000 means it is only 9% sure. Anything near +100 / -100 is close to a coin flip. ## Why Bookmakers Use Odds, Not ProbabilitiesIf probabilities are the cleaner mathematical object, why does the entire gambling industry - and most prediction markets - quote odds instead? Three reasons.
1. Odds make the payout obvious. When a bookmaker quotes 5/1, you instantly know that £10 wins £50. If they quoted 16.7%, you would have to do extra arithmetic at the till. Odds bake the payout into the price. They are a price, not a probability - and that distinction matters because customers are buying a contract, not a forecast.
2. Odds let bookmakers build in a margin without it being obvious. Imagine a fair coin flip. The true probabilities are 50% / 50%. A bookmaker offering fair odds would price both sides at 2.00 (decimal). Instead they offer 1.91 / 1.91. Each side has an implied probability of 52.4%. Add them up: 104.8%. That extra 4.8% is the bookmaker's edge, called the overround. Customers see odds that look reasonable; the margin is hidden in the cumulative gap from 100%. If books quoted probabilities, every customer would see the overround instantly.
3. Odds scale better at the extremes. A 0.5% probability is awkward to think about. The equivalent +19900 American odds, or 199/1 fractional, communicates 'long shot' more visually. At the other end, the difference between 99% and 99.5% probability sounds tiny but represents a doubling of risk for the bookmaker - clearer in the odds language as -9900 vs -19900.
For punters and forecasters, the takeaway is that every set of odds you see has been priced with a margin. Your job is to convert those odds back into the implied probabilities, then ask whether the bookmaker's probability is higher or lower than your own honest estimate. If your probability is meaningfully higher, you have a positive expected value bet. This is the entire mathematical foundation of profitable betting and is covered in detail in our expected value guide.
## The Overround: Why Implied Probabilities Sum to More Than 100%Take a typical Premier League match. The bookmaker offers:
Add them up: 47.6% + 29.4% + 27.8% = 104.8%.
A real probability distribution must sum to 100% - there are only three possible outcomes and one of them must occur. The extra 4.8% is the overround, and it is the bookmaker's gross margin. To recover the true implied probabilities the bookmaker is forecasting (rather than the prices they are charging), divide each implied probability by the overround:
Now they sum to 100%. These are the bookmaker's fair-market probabilities. To find a value bet, you need a personal probability estimate that is meaningfully higher than the fair-market figure for that outcome - not just higher than the implied price. This is one of the most common mistakes in amateur sports betting: treating implied probability as the bookmaker's true forecast, when in fact it includes a hidden margin.
Different markets carry very different overrounds. Football match-winner markets typically run at 104–107%. In-play markets can be 110%+. Horse racing can hit 115% or higher in large fields. Pinnacle and other sharp books run as low as 102%. Prediction markets like Polymarket and Manifold can run at or even slightly below 100% because they are peer-to-peer rather than sportsbook-to-customer. Lower overround means more of your edge survives.
## Common Mistakes With Probability and Odds1. Treating implied probability as the bookmaker's true forecast. As the overround example shows, the implied probability you read off a price is inflated by the margin. Always normalise before comparing it to your own estimate.
2. Reading 5/1 as '20%' when it is actually about 16.7%. People often confuse '5 to 1' with '1 in 5'. They are not the same. 5/1 is 1 in 6 - five losing scenarios for every winning one, totalling six. The probability is 1 / 6 = 16.7%, not 20%. The off-by-one error in how we count outcomes is one of the most persistent mistakes in betting.
3. Thinking past results change current odds. A roulette wheel landing on red ten times in a row does not raise the probability of black on the next spin. The wheel has no memory. Confusing odds (a property of the next event) with frequencies (a property of past events) is the gambler's fallacy in action.
4. Ignoring base rates when interpreting odds. A diagnostic test reporting 95% accuracy sounds compelling, but if the underlying condition has a base rate of 1 in 1,000, the post-test probability of actually having the disease is far lower than most people guess. This trap deserves its own deep dive - see our guides to base rate neglect and the false positive paradox.
5. Treating odds as fixed when you should be updating. New information should change your probability estimates and therefore the odds you would offer. The discipline of revising your beliefs in response to evidence is the core of Bayesian thinking - and it is what separates calibrated forecasters from people who anchor to their first guess.
6. Confusing odds with returns. Decimal odds of 2.00 do not mean a '100% return'. They mean a 100% profit on the staked amount, contingent on winning. The expected return is the probability-weighted average of all outcomes, which depends on whether the odds are fair. If the true probability is 50% and the odds are 2.00, the expected return is zero. Every set of odds combined with a probability gives an expected value - see expected value explained for the full framework.
7. Forgetting that small probabilities are hard to estimate. The difference between 1% and 0.5% might feel trivial, but it is a 2x change in the implied odds and the expected value of a long-shot bet. Humans are notoriously bad at calibrating small probabilities - there is good evidence we systematically overweight tiny chances (which is why lotteries thrive) and underweight near-certainties. This is covered in thinking in probabilities.
## Why This Matters Beyond BettingThe probability/odds distinction shows up everywhere uncertainty is priced.
In investing, options markets effectively quote odds. The implied volatility of an option is the market's probability distribution over future prices, and the option's premium is its price. A trader who can convert implied volatility into a personal probability estimate - and identify cases where the market is mispricing - has the same edge as a sharp sports bettor.
In medicine, doctors increasingly use odds ratios to communicate the relative risk of conditions and the effect of treatments. An odds ratio of 2.0 means the odds of the outcome are doubled by exposure - but if the baseline probability is small, the absolute risk increase might still be tiny. Confusing odds ratios with risk ratios leads to genuinely bad health decisions.
In forecasting, calibrated forecasters like the Good Judgment Project superforecasters quote probabilities directly (often to two decimal places), but they think in odds when checking themselves. A 70% probability is 7-to-3 in favour. A forecaster asking 'would I take this bet at 7-to-3?' often catches themselves overstating confidence.
In prediction markets, contracts trade between zero and one (with one representing a £1 payoff if the event happens). The price is the probability. This is the cleanest mapping you will ever see between probability and odds - and it is why prediction markets are an excellent training ground for probabilistic thinking. Our piece on how prediction markets work walks through the mechanics.
The meta-skill is fluency - being equally comfortable with probability, decimal, fractional and American odds, and being able to translate between them faster than the situation requires you to. Most of the value in this skill is defensive: it stops you from being misled by the way numbers are presented. The offensive side - using fluency to find mispriced bets - only opens up once the defensive side is automatic.
**Practice routine** For one week, every time you see odds quoted (in a sports app, financial news, weather forecast, or political article), do a five-second mental conversion to implied probability. After 50 reps the conversion becomes automatic. This single drill closes the gap between knowing the formulas and using them. ## Summary: When to Think in EachUse probability when you are forecasting, comparing your estimate to someone else's, doing expected value calculations, or communicating uncertainty in plain English. Probabilities are additive (P(A or B) = P(A) + P(B) for mutually exclusive events) and feel intuitive once you commit to the 0–1 scale.
Use odds when you are pricing a bet, communicating with a sportsbook, working with prediction markets, or trying to feel the asymmetry of an outcome. A 95% probability and a 5% probability look symmetric; the equivalent odds (-1900 vs +1900) make the asymmetry obvious.
The best operators move between the two languages constantly. They forecast in probability, price in odds, and check their own work in both. Once you can do the conversions in your head - fluently, fast, across formats - uncertainty stops being intimidating and starts being something you can negotiate with.
**Q: What is the difference between probability and odds?** Probability is the share of all possible outcomes in which an event occurs, expressed as a number between 0 and 1 (or as a percentage). Odds are the ratio of the ways an event can happen to the ways it cannot. A 25% probability is the same uncertainty as 3-to-1 odds against - three losing scenarios for every one winning scenario, totalling four equally likely scenarios. **Q: How do I convert decimal odds to probability?** Divide 1 by the decimal odds. Decimal odds of 2.50 imply a probability of 1 / 2.50 = 0.40, or 40%. This gives you the implied probability - the probability the bookmaker's price is claiming. Note that this includes the bookmaker's margin, so it is slightly higher than their actual forecast. **Q: How do I convert American odds to probability?** For positive American odds (underdogs): probability = 100 / (odds + 100). So +200 implies 100 / 300 = 33.3%. For negative American odds (favourites): probability = |odds| / (|odds| + 100). So -150 implies 150 / 250 = 60%. The bigger the number in either direction, the more lopsided the market believes the outcome to be. **Q: Why do bookmakers' implied probabilities sum to more than 100%?** The excess over 100% is the bookmaker's margin, called the overround or vig. It is how the bookmaker makes money. To recover the bookmaker's true forecast probabilities, divide each outcome's implied probability by the total overround. Sharp books like Pinnacle run overrounds as low as 102%; recreational books often run 105–110%; in-play and horse racing markets can run 115% or higher. **Q: Are 5/1 odds the same as a 20% probability?** No, this is one of the most common mistakes. 5/1 means five losing scenarios for every one winning scenario - six total scenario …[truncated for length] --- # Fractional Kelly: Why Half Kelly Is the Smart Default Source: https://expectedvalue.co.uk/blog/fractional-kelly/ # Fractional Kelly: Why Half Kelly Is the Smart Default Full Kelly is mathematically optimal but emotionally brutal. Why half Kelly is the practical default, when to scale up or down, and how the pros use it.The fractional Kelly criterion is a position-sizing rule where you bet a fixed fraction - usually a half or a quarter - of what the full Kelly formula recommends. It exists for a simple reason: full Kelly is mathematically optimal under perfect information, and almost nobody has perfect information. Half Kelly trades a small amount of long-run growth for a large reduction in drawdowns, and it is what serious practitioners - from Edward Thorp's hedge fund to Bill Benter's horse-betting syndicate - actually used.
**TL;DR** If you're using Kelly and you don't already know exactly which fraction you should use, start with half Kelly. It cuts your expected drawdown by 75% while only sacrificing 25% of expected growth - the best risk-adjusted trade-off in the family. ## What Fractional Kelly Actually MeansFull Kelly tells you the bet size that maximises the long-run geometric growth rate of your bankroll, assuming you know your true edge. For a simple binary bet, the formula is f* = (bp - q) / b, where b is the decimal odds minus one, p is your true probability of winning, and q = 1 - p.
Fractional Kelly is just k · f*, where k is some fraction less than 1. Half Kelly (k = 0.5) is by far the most common, followed by quarter Kelly (k = 0.25) and, occasionally, three-quarter Kelly. Some funds run as low as eighth Kelly when their edge is highly uncertain.
The reason this rule even exists is that the math behind Kelly is fragile in a specific way. If your real edge is slightly smaller than you think - which it almost always is - full Kelly stops being optimal and becomes catastrophic. Fractional Kelly is a robustness adjustment, not a correction to a mistake in the formula.
## Why Full Kelly Is Brutally VolatileThe Kelly fraction maximises log-utility - the expected logarithm of terminal wealth. This is exactly the right objective if you're an immortal bettor with perfect information and no path-dependent constraints. In the real world, three properties of full Kelly make it intolerable.
First, the variance is huge. Under full Kelly, the standard deviation of your bankroll growth is equal to your edge. That means a strategy with a 2% edge per bet will swing wildly between gains and losses on any short horizon, even though it's profitable in expectation.
Second, drawdowns are extreme. A well-known result is that under full Kelly, the probability of seeing your bankroll cut in half at some point before doubling is exactly 50%. The probability of an 80% drawdown is 20%. These aren't tail events - they are the typical experience of a full-Kelly bettor.
Third, the strategy is acutely sensitive to estimation error. If you overestimate your edge by even a small amount, the optimal-feeling bet size becomes systematically too large, and the strategy can be worse than not betting at all. This connects to a deeper idea in ergodicity: a time-average outcome can diverge brutally from an ensemble-average outcome when variance is high, and full Kelly maximises a quantity that only makes sense in the ensemble.
## The Drawdown Math: Half vs Full vs QuarterThere's an elegant closed-form result for fractional Kelly drawdowns. Under continuous-time Kelly betting with fraction k, the probability of drawing down by a factor of x before doubling is approximately x^(2(1-k)/k) for k between 0 and 1.
That formula does a lot of work. It says that halving the Kelly fraction doesn't just halve the drawdown probability - it transforms the exponent. Below is what that looks like for common drawdown thresholds, assuming a stable edge:
- **50% drawdown - Full Kelly**: 50% - **50% drawdown - Half Kelly**: 25% - **50% drawdown - Quarter Kelly**: 6.25% - **80% drawdown - Full Kelly**: 20% - **80% drawdown - Half Kelly**: 4% - **80% drawdown - Quarter Kelly**: 0.16%The cost of that risk reduction is surprisingly small. Geometric growth rate under fractional Kelly scales as k(2-k) times the full-Kelly growth rate. Half Kelly captures 75% of full Kelly's growth, and quarter Kelly captures 43.75%. So half Kelly gives up a quarter of the upside to slash the chance of a 50% drawdown from a coin-flip to a one-in-four event, and the chance of an 80% drawdown from one-in-five to one-in-twenty-five.
That asymmetric trade-off - small reduction in growth for a large reduction in volatility - is why half Kelly is the default among practitioners who can do the math.
## Half Kelly: The Practical DefaultHalf Kelly is recommended whenever three conditions hold. The edge is real but estimated with normal uncertainty. The bettor will keep betting for a long time and cares about geometric growth rather than a single-period payoff. And there is no external constraint (margin call, withdrawal, deadline) that would make a transient drawdown ruinous.
Those conditions describe most serious sports betting, value investing, and quantitative trading. They are also approximately what a poker bankroll-management framework looks like - see expected value in poker for the parallel logic about how aggressively to play with a given edge.
The reason half Kelly works as a default is that it's robust to the most common mistake: overconfidence in your own edge. If your true edge is half of what you think it is - which is approximately what you should assume for any model you didn't already calibrate against years of out-of-sample data - then half Kelly on your estimated edge equals full Kelly on your true edge. You're back at the theoretical optimum, by accident, by being humble.
## Quarter Kelly: For Less Certain EdgesQuarter Kelly is appropriate when your edge estimate is itself a wide distribution rather than a point estimate. This is the norm in two situations.
The first is any new strategy that hasn't yet accumulated thousands of bets of out-of-sample evidence. A backtested model on historical data is notorious for overstating its real-world edge - a backtest with a 4% edge often has a real edge closer to 1–2%, and sometimes none. Quarter Kelly buys you survival during the period when you're discovering this.
The second is bets in regimes with structural shifts. Sports betting markets sharpen rapidly, financial markets reprice persistent inefficiencies, and what was a 3% edge last season can be a 0.5% edge this season. Quarter Kelly lets a strategy degrade gracefully rather than vaporise.
Below is how the trade-offs stack up for the three most common fractions, assuming you're working with a true 2% edge per bet on even-money outcomes:
- **Full Kelly - growth rate**: 100% of theoretical max - **Full Kelly - typical 12-month worst drawdown**: 40-60% - **Half Kelly - growth rate**: 75% of theoretical max - **Half Kelly - typical 12-month worst drawdown**: 20-30% - **Quarter Kelly - growth rate**: 44% of theoretical max - **Quarter Kelly - typical 12-month worst drawdown**: 10-15% ## When to Scale Up - and When Not ToThree-quarter Kelly is occasionally defensible, but only in a narrow set of circumstances. The edge must be quantitatively measured against a deep out-of-sample dataset. The portfolio must be diversified across many uncorrelated bets, so that any individual loss has a small effect on total bankroll. And the bettor must have a documented psychological tolerance for drawdowns of 30-40%, because they will happen.
Full Kelly is almost never the right choice. Even Edward Thorp, who first applied Kelly to blackjack card-counting in the early 1960s, recommended in his retrospective writings that practitioners use fractional Kelly. His own hedge fund, Princeton Newport Partners, used a heavily fractional approach with strong diversification.
The signs that you should scale down from half Kelly are equally important. If your edge depends on a market condition that could change (a soft sportsbook, an underpriced volatility regime, a competitor not yet in the market), scale down. If you've had two consecutive losing months and are starting to doubt the model, scale down. If you've taken on outside capital with redemption rights, scale down. Drawdowns interact with your own behaviour in ways that the math doesn't capture - and the math doesn't have to face redemption calls.
## Thorp, Benter, and the Empirical CaseThe two most famous case studies of professional Kelly-style betting both used fractional, not full, Kelly.
Edward Thorp's blackjack work, published in Beat the Dealer in 1962, established the original real-world application of Kelly. By his own later accounts, his team used roughly half-Kelly sizing because the variance of full Kelly was unacceptable for a team operating with shared capital and limited tolerance for visible swings at the table. Thorp's subsequent hedge fund, Princeton Newport Partners, ran for nearly twenty years with returns of around 15% annually and very low volatility - possible only because of fractional sizing combined with broad diversification.
Bill Benter ran a quantitative horse-betting syndicate in Hong Kong from the mid-1980s and reportedly earned hundreds of millions of dollars over several decades. Public interviews and conference talks attribute his sizing approach to a Kelly framework, but applied conservatively - typically well below full Kelly, and dynamically reduced when the model's confidence was lower. Benter's writing on the topic emphasises that the marginal cost of being too conservative is far smaller than the marginal cost of being too aggressive, which is the core intuition behind fractional Kelly.
The pattern across both is consistent. Professional bettors with a measured edge, working over long horizons, with their own capital at stake, choose growth-volatility trade-offs much closer to half Kelly than full Kelly. The theoretical optimum is almost never the practical optimum.
## Practical ImplementationThe mechanical steps for using fractional Kelly look like this. First, calculate the full Kelly stake for the bet, using your honest estimate of edge and the offered odds. The Kelly criterion calculator handles the arithmetic for binary bets and multiple-outcome scenarios.
Second, multiply by your chosen fraction. If you're using half Kelly, halve it. Quarter Kelly, quarter it. Don't deviate session-by-session - the discipline of a fixed fraction is half the point.
Third, recompute your bankroll between bets. Kelly is a proportional rule, so the stake should be a fraction of the current bankroll, not the starting bankroll. This automatically scales up bets after wins and down after losses, which is part of why Kelly survives long drawdowns: the bet sizes shrink with the bankroll, so the percentage drawdown stays bounded.
Fourth, cap individual bets. Even half Kelly can recommend a 20% bet when the edge is high and the odds are short. Most practitioners impose an additional cap - typically 5% to 10% of bankroll per bet - to defend against model errors on outlier opportunities. For the systematic relationship between bet size, bankroll, and risk of ruin, see position sizing with the Kelly criterion.
**Practical tip** Keep a written record of every bet's recommended full-Kelly size, your applied fraction, and your actual stake. Reviewing this monthly is the single best way to catch creeping over-confidence - the most common failure mode of Kelly bettors. ## Common Mistakes With Fractional KellyThree errors recur in almost every retail attempt at Kelly betting.
Mixing fractions in the same portfolio. Using half Kelly on bets you feel confident about and quarter Kelly on speculative ones sounds prudent, but it interacts badly with the fact that your confidence is itself the most error-prone input. The bets you feel most confident about are statistically the ones most likely to embed an overestimation bias. A fixed fraction across the whole portfolio is more robust than a self-reported confidence-weighted one. This is closely related to the lesson of overconfidence bias: the bets you'd label highest-confidence are exactly where your edge estimate is most likely inflated.
Ignoring correlation between bets. Kelly assumes independent outcomes. If you have ten bets on related markets - say, ten football matches with overlapping injury and weather effects - fractional Kelly on each one ignores that they could all lose together. The fix is to compute Kelly on the portfolio joint distribution, or to use a smaller fraction (eighth Kelly is common in highly correlated books) as a practical proxy.
Treating Kelly as a precise number. The output of a Kelly formula carries the precision of its inputs, which usually aren't precise. Quoting a 3.247% stake suggests a precision you don't have. Round to the nearest meaningful unit - typically the nearest 0.5% of bankroll - and accept the residual error as below the noise of your edge estimate. For more on the gap between the precision a model offers and the precision an estimate deserves, see risk vs uncertainty.
## Frequently asked questions **Q: Why not just use the Kelly fraction that maximises my expected return?** Because expected return is the wrong objective for repeated betting. The Kelly fraction that maximises geometric growth - which is what you actually experience over many bets - is much smaller than the one that maximises arithmetic expected value. And the optimal geometric-growth fraction itself depends on knowing your edge exactly, which you don't, which is why fractional Kelly exists. **Q: Is half Kelly always better than full Kelly?** Not in expectation - full Kelly has a higher expected geometric growth rate when your edge is known with certainty. But it has dramatically higher variance and drawdown risk. For any human bettor with imperfect information and a finite horizon, the lower variance of half Kelly is worth the smaller expected growth in nearly all realistic settings. **Q: Can I use fractional Kelly for stock investing?** Yes, with caveats. Kelly works best when you have a defined edge and discrete betting opportunities. For diversified passive investing, the model doesn't really apply - you're not making isolated bets with measurable edges. For active strategies (concentrated positions, options, event-driven trades), fractional Kelly is a reasonable position-sizing framework, but use quarter or eighth Kelly because financial-market edges are notoriously hard to estimate accurately. **Q: What fraction did Edward Thorp use?** Thorp has stated in various retrospective writings that his blackjack teams used approximately half Kelly, and his hedge fund used a heavily fractional approach combined with broad diversification. He has cautioned against full Kelly for individual practitioners in nearly every public discussion of the topic. **Q: How do I know if my edge estimate is reliable enough for half Kelly?** A useful rule of thumb: if your edge is based on fewer than 500 out-of-sample observations, use quarter Kelly. If it's based on 500–2,000, half Kelly is defensible. Beyond that, the edge estimate is stable enough that the choice becomes a question of personal volatility tolerance, not statistical confidence. And if your edge is based purely on a backtest, assume the real edge is one-third to one-half of the backtested figure. **Q: Does fractional Kelly still beat flat betting?** Yes, dramatically. Flat betting (constant stake) doesn't compound and ignores the size of your edge. Even quarter Kelly produces meaningfully higher long-run growth than flat betting on the same edge, because it scales with bankroll. The right comparison isn't full vs half Kelly - it's any fractional Kelly vs flat betting, and any fractional Kelly wins. --- # 12 Best Books on Probabilistic Thinking and Decision-Making Source: https://expectedvalue.co.uk/blog/best-books-probabilistic-thinking-decision-making/ # 12 Best Books on Probabilistic Thinking and Decision-Making A curated reading list on probability, decision-making under uncertainty, and rational thinking - from Kahneman to practical guides for forecasters. ## Why Read About Probabilistic Thinking? **Answer**The shortest path into probabilistic thinking is Daniel Kahneman's Thinking, Fast and Slow plus Annie Duke's Thinking in Bets. From there, branch out with Nassim Taleb (Fooled by Randomness, The Black Swan), Philip Tetlock (Superforecasting), and Howard Marks (The Most Important Thing). The other seven are useful follow-ups once you have the basics.
Most important decisions are made under uncertainty. Whether you're investing, hiring, diagnosing, negotiating, or planning a project, you're working with incomplete information and uncertain outcomes. The quality of your decisions depends on how well you handle that uncertainty. The books below are the best resources for developing this skill. They range from accessible introductions to technical deep dives, covering probability theory, cognitive biases, forecasting, risk management, and applied decision-making. I've organised them roughly from most accessible to most technical, with a note on who will get the most from each one.This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
## 1. Thinking, Fast and Slow - Daniel KahnemanKahneman's magnum opus distils decades of Nobel Prize-winning research on human judgement and decision-making. The central framework - System 1 (fast, intuitive, error-prone) versus System 2 (slow, deliberate, effortful) - has become the standard lens for understanding why we make systematic mistakes.
The book covers loss aversion, anchoring, the availability heuristic, base rate neglect, overconfidence, and dozens of other biases that distort our probabilistic reasoning. It's dense (nearly 500 pages) and occasionally academic, but the insights are transformative.
Best for: Everyone. If you read one book from this list, make it this one. It provides the vocabulary and frameworks that every other book on this list builds upon.
Key insight: We don't just make random errors - we make predictable errors. Understanding these systematic biases is the first step to correcting them.
## 2. Superforecasting: The Art and Science of Prediction - Philip Tetlock & Dan GardnerBased on the Good Judgment Project (a massive forecasting tournament funded by IARPA), this book identifies what makes some people dramatically better at predicting future events. "Superforecasters" aren't smarter or more knowledgeable - they think differently.
The key habits: thinking in probabilities rather than certainties, breaking complex questions into components, updating beliefs incrementally when new evidence arrives, and actively seeking disconfirming evidence. These are learnable skills, and this book is the best practical guide to developing them.
Best for: Anyone who makes predictions professionally - investors, strategists, analysts, project managers. Also excellent for anyone who wants to be less wrong about the future in general.
Key insight: The best forecasters treat their beliefs as hypotheses to be tested, not positions to be defended. They update constantly and in small increments - exactly the Bayesian ideal.
## 3. Thinking in Bets - Annie DukeAnnie Duke, a former World Series of Poker champion turned decision scientist, makes the case that all decisions are bets - and that we should evaluate them by the quality of our reasoning process, not by the outcome.
The book's strongest contribution is the concept of "resulting" - judging decisions by their results rather than their logic. A good decision can produce a bad outcome (and vice versa) because of randomness. Duke argues we should build "decision groups" to help us distinguish between skill and luck in our past choices.
Shorter and more practical than Kahneman, this is an excellent entry point for people who want actionable frameworks rather than academic theory.
Best for: Business leaders, poker players, and anyone who struggles with hindsight bias. Also useful for teams that need to build a culture of better decision-making.
Key insight: Don't say "I'm sure" or "I have no idea." Every belief has a probability. Saying "I'm 70% confident" forces you to acknowledge uncertainty and creates space for new information to shift your view.
## 4. The Signal and the Noise - Nate SilverSilver, the statistician behind FiveThirtyEight, surveys prediction across domains: weather forecasting, baseball scouting, earthquake prediction, chess, poker, political polling, epidemiology, and finance. Some fields have gotten dramatically better at prediction; others remain terrible.
The book's central argument is about signal versus noise: most people and models are bad at predictions because they overfit to noise (random variation) and miss genuine signals. The chapters on weather forecasting (remarkably accurate) and earthquake prediction (essentially impossible) illustrate why some phenomena are inherently more predictable than others.
Best for: Data analysts, aspiring forecasters, and anyone interested in why prediction works in some domains but not others. The chapter on Bayesian reasoning is one of the best popular introductions available.
Key insight: The confidence of a prediction should scale with the predictability of the domain. Weather 3 days out is highly predictable; stock prices tomorrow are not. Knowing this distinction is worth more than any model.
## 5. Fooled by Randomness - Nassim Nicholas TalebTaleb's first and most readable book (before The Black Swan and Antifragile) explores how we consistently underestimate the role of randomness in success and failure. Successful traders may just be lucky; failed entrepreneurs may have made excellent decisions in the face of unavoidable variance.
The book is discursive, opinionated, and occasionally combative - Taleb's personality is part of the package. But the core ideas are profound: survivorship bias means we study winners and ignore the silent majority of losers; our brains are wired to see patterns in randomness; and rare events (black swans) dominate outcomes in many domains far more than we acknowledge.
Best for: Investors, entrepreneurs, and anyone whose self-assessment might be contaminated by survivorship bias. A useful antidote to the "great man" theory of success.
Key insight: Alternative histories matter. Before crediting someone's success to skill, ask: "In how many parallel universes would this person have failed doing exactly the same thing?"
## 6. How to Decide - Annie DukeDuke's follow-up to Thinking in Bets is more structured and practical - almost a workbook. It introduces tools like the "decision matrix," "pre-mortem," and "knowledge tracking" that you can apply immediately.
The book excels at making abstract decision theory tangible. Each chapter ends with exercises. If Thinking in Bets was the theory, How to Decide is the practice.
Best for: People who liked Thinking in Bets and want concrete tools. Also good as a standalone for someone who prefers hands-on exercises to narrative non-fiction.
## 7. Fortune's Formula - William PoundstoneA gripping narrative history of the Kelly Criterion - the mathematical formula for optimal bet sizing. The story connects Claude Shannon (information theory), Ed Thorp (card counting and hedge funds), John Kelly (Bell Labs), and the mob-connected world of early Las Vegas.
Beyond the history, the book clearly explains why bet sizing matters as much as edge, why geometric growth beats arithmetic growth over time, and why most investors over-bet or under-bet. It's the best popular-science treatment of the Kelly Criterion and its real-world applications.
Best for: Investors, traders, and gambling enthusiasts. Anyone interested in the history of quantitative finance. Reads like a thriller despite being about mathematics.
Key insight: Having an edge isn't enough - you need to size your bets correctly. Overbetting with a genuine edge can still lead to ruin. This is the central lesson of Kelly, and Fortune's Formula makes it visceral.
## 8. The Drunkard's Walk - Leonard MlodinowMlodinow, a physicist and screenwriter, provides an accessible tour of probability theory and its application to everyday life. From the Monty Hall problem to regression to the mean, the book explains why our intuitions about probability are so consistently wrong.
Lighter than Kahneman and more mathematically grounded than Taleb, this is an excellent introduction for readers who want to understand the mechanics of probability without heavy formalism.
Best for: General readers new to probability. A good starting point before tackling Kahneman or Silver.
## 9. Against the Gods: The Remarkable Story of Risk - Peter L. BernsteinBernstein traces the intellectual history of risk from ancient Greece to modern portfolio theory. The book covers Pascal's wager, the development of probability theory, the invention of insurance, the birth of statistics, and the evolution of financial risk management.
It's more historical than practical, but understanding how we arrived at modern risk thinking - and how recently we got there - gives valuable perspective. The ideas we now take for granted (diversification, expected value, standard deviation) were revolutionary when first proposed.
Best for: Investors and finance professionals who want deeper context. History enthusiasts. Anyone interested in how mathematical ideas shaped the modern world.
## 10. Noise: A Flaw in Human Judgment - Daniel Kahneman, Olivier Sibony & Cass SunsteinWhile Thinking, Fast and Slow focused on bias (systematic errors), Noise tackles the other half of the problem: noise (unwanted variability in judgments). Two judges sentencing the same case give wildly different sentences. Two doctors reading the same X-ray give different diagnoses. Two underwriters pricing the same risk quote different premiums.
The book argues that noise is as costly as bias but receives far less attention because it's invisible - you need to compare multiple judgments to see it. The proposed remedies (decision hygiene, structured approaches, algorithms) are practical and applicable across organisations.
Best for: Managers, HR professionals, hiring managers, legal professionals - anyone whose organisation makes repeated judgments that should be consistent but aren't.
Key insight: Bias is like shooting consistently to the left of the target. Noise is like shooting all over the place. Both are problems, but noise is harder to detect because the errors don't have a visible pattern.
## 11. The Black Swan - Nassim Nicholas TalebTaleb's most famous work argues that rare, unpredictable events ("black swans") have disproportionate impact - and that most of our models, forecasts, and plans are dangerously fragile because they ignore this reality.
The book challenges the Gaussian (bell curve) assumptions that underpin most of finance and risk management. In "Extremistan" (domains like financial markets, book sales, and technology), outcomes follow power laws, not normal distributions, and our standard statistical tools break down.
More philosophical and provocative than Fooled by Randomness. Some readers find Taleb's style grating; others find it exhilarating. The ideas are genuinely important regardless.
Best for: Anyone in finance, risk management, or strategic planning. Essential reading for understanding tail risk and why "once in a century" events happen every few years.
## 12. How to Measure Anything - Douglas HubbardHubbard's thesis is radical: if something matters, it's observable; if it's observable, it can be measured; and you probably need less data than you think to measure it usefully. The book provides practical methods for quantifying things that seem unquantifiable - customer satisfaction, security risk, the value of information.
The core technique is calibrated probability estimation: learning to assign accurate probabilities to uncertain quantities. Combined with simple Monte Carlo simulations, this approach lets you model uncertainty in any business decision.
Best for: Business analysts, project managers, security professionals, and anyone who's been told "you can't measure that." The most practically useful book on this list for day-to-day business decisions.
Key insight: The first measurement that reduces your uncertainty is almost always worth far more than the cost of obtaining it. People routinely over-invest in precision and under-invest in breadth.
## Where to StartIf you're new to probabilistic thinking, start with Thinking in Bets (shortest, most practical) or The Drunkard's Walk (most accessible introduction to probability). Then read Thinking, Fast and Slow for the complete framework.
If you're an investor, prioritise Fortune's Formula, Fooled by Randomness, and The Black Swan.
If you make forecasts or predictions professionally, Superforecasting and The Signal and the Noise are essential.
If you want to improve decision-making in your organisation, start with Noise and How to Decide.
And if you haven't already, start with our foundational articles: Expected Value Explained, The Kelly Criterion, and Thinking in Probabilities. They cover the core concepts that these books expand upon.
**Q: What is the best beginner book on probability?** The Drunkard's Walk by Leonard Mlodinow is the most accessible introduction to probability for readers with no mathematical background. For a more decision-focused starting point, Thinking in Bets by Annie Duke is shorter and immediately practical. **Q: What is the single most important book on decision-making?** Thinking, Fast and Slow by Daniel Kahneman. It provides the foundational framework (System 1 vs System 2, cognitive biases, prospect theory) that nearly every other book on decision-making builds upon. It's dense but essential. **Q: Are there good books on probability for investors specifically?** Fortune's Formula (optimal bet sizing via the Kelly Criterion), Fooled by Randomness (the role of luck in investing success), and The Black Swan (why extreme events dominate financial markets) form an excellent trilogy for investors. **Q: What book teaches practical forecasting skills?** Superforecasting by Philip Tetlock is the definitive practical guide. It identifies the specific habits and techniques that separate excellent forecasters from poor ones, based on the largest forecasting study ever conducted. --- # 15 Best Decision Making Books to Change How You Think Source: https://expectedvalue.co.uk/blog/best-decision-making-books/ # 15 Best Decision Making Books to Change How You Think 15 best decision-making books - choice architecture, behavioural economics, frameworks, and habits. Mini-reviews, who each is for, and reading order.This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
**Answer** Start with Daniel Kahneman's Thinking, Fast and Slow for the cognitive-bias map, Annie Duke's Thinking in Bets for the everyday probability mindset, and Chip and Dan Heath's Decisive for the four-step decision framework. The other twelve build out specialised lenses: superforecasting, signal vs noise, value investing, and behavioural finance.The best decision-making books don't teach you a formula - they retrain the way you frame choices in the first place. This list pulls together 15 of them, drawn from behavioural economics, choice architecture, cognitive psychology, and practical decision frameworks. Each entry includes a short verdict on what the book actually teaches, who it's for, and how it sits in a wider reading order.
**How this list is different**We've deliberately kept this complementary to our existing list of probability-focused books - Thinking Fast and Slow, Superforecasting, Fooled by Randomness, and the rest are covered in our separate guide on the 12 best books on probabilistic thinking. The list below focuses on the wider decision-science literature: behavioural economics, choice architecture, decision frameworks, habit formation, and the psychology of choice. Together the two lists give you a complete shelf.
Decision-making is a sprawling field. It draws from psychology, economics, neuroscience, statistics, and operations research. A good reading list won't pretend the field has one canonical answer; it'll give you the half-dozen genuinely useful frames and let you pick the ones that match your own decisions. That's the goal here.
Most lists of this kind are a name-check of bestsellers. We've tried to do something more useful: each book gets a 100-200 word entry that tells you the actual idea, the situation it changes how you think about, and the readers who'll get the most from it. Where a book has been overtaken or partially refuted by later research, we've said so. Where a book is genuinely worth the time, we've said why.
If you're new to the topic, jump straight to the "Where to start" section at the end - it walks you through the order in which to read these books for the steepest learning curve. If you've already read a few, the list groups books loosely by theme so you can fill in the gaps.
For background on why this matters, our guide to thinking in probabilities covers the underlying mental shift, and expected value explained walks through the single most useful formula in the entire field.
## Which behavioural-economics books should you read first?1. Nudge - Richard Thaler and Cass Sunstein
The book that introduced the term "choice architecture" and arguably launched modern behavioural-economics policy. Thaler and Sunstein argue that small changes in how options are presented (defaults, ordering, the way information is framed) can dramatically shift the choices people make, without restricting freedom. Pension auto-enrolment, organ-donor opt-out schemes, and the way menus get redesigned in canteens all trace back to ideas in this book. The best chapters are on retirement saving, school choice, and organ donation - concrete, evidence-led, and immediately useful for anyone designing a system other people have to navigate. The 2021 "final edition" updates the original with a decade of follow-up research.
Who it's for: anyone who designs systems other people use - managers, policymakers, founders, parents. Also a strong second behavioural-economics book if you've already read Thinking, Fast and Slow.
2. Predictably Irrational - Dan Ariely
Ariely's catalogue of the small, repeatable ways human decision-making goes off the rails. The chapter on the power of free ("FREE!" pricing distorts choice in ways that pure economics can't explain) is genuinely revelatory, and the analyses of social vs market norms, the cost of zero-cost decisions, and the influence of expectations on perceived experience are sharp and well-evidenced. Ariely's writing is conversational and the experiments are memorable, which makes the book genuinely re-readable.
Who it's for: readers who want behavioural-economics ideas illustrated through experiments rather than equations. A great gateway book.
3. Misbehaving: The Making of Behavioural Economics - Richard Thaler
Half memoir, half history of the field. Thaler walks through the origins of behavioural economics from his own perspective - the early scepticism of mainstream economists, the slow accumulation of empirical results, the fights over loss aversion, mental accounting, and the endowment effect. It's funnier than it sounds, and gives you a far better sense of why the field developed the way it did than any straight textbook would. If you've already read Nudge, this is the natural next step.
Who it's for: readers who want both the ideas and the institutional history of behavioural economics in one book.
4. The Paradox of Choice - Barry Schwartz
Schwartz's thesis is that more choice often produces worse decisions and lower satisfaction, not better ones. He distinguishes between "maximisers" (who want the best possible option and end up exhausted by comparison) and "satisficers" (who pick the first option that meets their criteria and move on), and argues that satisficing is generally a better strategy. Some of the original empirical claims have been challenged in follow-up research, but the core idea - that choice has a real cost, and most environments are over-optimised for adding more - has aged well.
Who it's for: anyone who agonises over decisions, anyone designing customer experiences, and anyone who's noticed that more options didn't lead to better outcomes.
## Which books cover cognitive biases best?5. The Art of Thinking Clearly - Rolf Dobelli
Ninety-nine cognitive errors in three-page chapters. Dobelli's book is a genuinely useful field guide rather than a deep theoretical work - each chapter names a bias, gives a memorable example, and tells you how to spot it in your own thinking. It's the kind of book you keep on a shelf and consult before big decisions. The format makes it easy to dip into and re-read; the trade-off is that the depth is shallower than Kahneman or Ariely.
Who it's for: readers who want a fast, practical reference rather than a theoretical treatise. Excellent third behavioural book after Predictably Irrational and Thinking, Fast and Slow.
6. Influence: The Psychology of Persuasion - Robert Cialdini
Cialdini's six principles of influence - reciprocity, commitment and consistency, social proof, liking, authority, and scarcity - are one of the most cited frameworks in applied psychology. The book is partly a how-to-spot-manipulation guide and partly a how-to-influence-ethically manual, and it works as both. Twenty years of follow-up research have largely held up the original framework, with a seventh principle (unity, the sense of shared identity) added in the 2021 edition. The book changes how you read advertising, sales pitches, and political messaging.
Who it's for: anyone in sales, marketing, negotiation, or hiring; anyone who suspects they're being persuaded and wants to understand the mechanism.
7. Sources of Power - Gary Klein
While Kahneman and Tversky focused on the failures of intuition, Gary Klein spent decades studying when intuition genuinely works - and why. His research on firefighters, paramedics, and military commanders shows that experts often make excellent decisions in seconds, drawing on pattern recognition built up over thousands of cases. The book introduces the recognition-primed decision model, which is now standard in naturalistic decision-making research. It's a useful corrective to a behavioural-economics literature that sometimes implies all intuition is suspect.
Who it's for: readers who want a more nuanced view of intuition vs analytical decision-making. Particularly useful for anyone in operational, time-pressured roles where slow analysis isn't an option.
## Which books give you usable decision frameworks?8. Decisive - Chip and Dan Heath
The Heath brothers' WRAP framework (Widen options, Reality-test assumptions, Attain distance, Prepare to be wrong) is one of the most useful practical decision tools we know of. Each step has a chapter of its own with concrete techniques: the "Vanishing Options Test" for breaking out of binary either/or thinking, pre-mortems for stress-testing decisions before committing, and the 10/10/10 rule for getting emotional distance. The case studies are largely from the business world but the framework generalises to any consequential personal decision. We use parts of WRAP at work weekly.
Who it's for: anyone who wants a framework rather than a description of what goes wrong. The single most actionable book on this list.
9. Smart Choices - Hammond, Keeney and Raiffa
An older book (1999) but the foundational guide to structured decision analysis. The PrOACT framework (Problem, Objectives, Alternatives, Consequences, Trade-offs) is a workhorse for any decision with multiple competing criteria - career moves, big purchases, organisational choices. The chapters on building consequence tables and explicit trade-off analysis are excellent and the methodology is genuinely practical, not theoretical. The writing is dry compared to Decisive, but the depth is greater.
Who it's for: readers who want a more rigorous structured approach. Particularly useful for high-stakes decisions where you have time to think.
10. Algorithms to Live By - Brian Christian and Tom Griffiths
Christian and Griffiths translate computer-science decision algorithms into everyday rules of thumb. The chapter on optimal stopping (when to commit to a partner, employee, or apartment after a search) gives you the 37% rule. The exploration-vs-exploitation chapter explains when to keep trying new restaurants and when to settle into the favourites. Sorting, scheduling, caching, and game theory each get their own translation. It's intellectually playful, but the underlying advice is genuinely sharp - most of these algorithms are provably optimal under their assumptions.
Who it's for: readers who enjoy the mathematical side of decision-making. Probably the most fun book on the list.
11. The Checklist Manifesto - Atul Gawande
Gawande, a surgeon, makes the case that simple checklists dramatically reduce errors in complex domains where expertise alone isn't enough. The case studies (aviation, surgery, building construction) are compelling, and the framework for designing a useful checklist - distinguishing pause points from continuous tasks, reading-do vs do-confirm, keeping it short - is concrete and reusable. It's not a decision-making book in the cognitive-bias sense, but it's a brilliant book on decision execution: turning what you've decided into what you actually do.
Who it's for: anyone in a complex operational role, anyone who repeatedly makes the same kind of decision and wants to professionalise the process.
## Which books cover habits and long-run choices?12. The Power of Habit - Charles Duhigg
Most of your daily choices aren't really choices - they're habits triggered by cues and sustained by rewards. Duhigg's habit loop (cue, routine, reward) and his concept of "keystone habits" (habits whose change cascades into other behavioural changes) have become standard vocabulary. The chapters on individual habits (Michael Phelps's pre-race routine, Starbucks' employee training) and organisational habits (Alcoa's safety transformation) are well-told and well-evidenced. Some of the neuroscience claims are simplified, but the practical framework is sound.
Who it's for: anyone who wants to build or change a recurring behaviour, including the daily decisions they don't think of as decisions.
13. Switch - Chip and Dan Heath
The Heath brothers' earlier book on how change happens - including how you persuade yourself or others to make a difficult choice and stick to it. The central metaphor is the rider (rational mind), the elephant (emotional mind), and the path (the environment). Effective change requires directing the rider, motivating the elephant, and shaping the path simultaneously; failures of change usually neglect one of the three. It pairs naturally with The Power of Habit and Decisive.
Who it's for: anyone trying to change their own behaviour or persuade a team / family / organisation to change theirs.
14. Drive - Daniel Pink
Pink's synthesis of motivation research argues that for any task more complex than a routine procedure, intrinsic motivation (autonomy, mastery, purpose) beats extrinsic incentives (carrots and sticks). The book draws heavily on Edward Deci and Richard Ryan's self-determination theory, and the practical implications for how to structure your own work life - and how to incentivise others - are immediate. Some of the empirical claims have been disputed, but the core distinction between routine and creative work and the appropriate motivational structures for each has aged well.
Who it's for: anyone making career decisions, designing roles, or deciding how to incentivise themselves on a difficult project.
15. Range - David Epstein
Epstein's argument is that for many domains, breadth of experience beats narrow specialisation - particularly in fields where the problems are ill-defined, the rules change, or feedback is delayed. He draws on case studies from sports, music, science, and medicine to show that the most effective experts often took winding paths and resisted early specialisation. As a decision-making book, it's a useful corrective to the 10,000-hours-of-deliberate-practice narrative - and a reminder that the right strategy for a wicked problem (uncertain rules, delayed feedback) is very different from the right strategy for a kind problem (clear rules, immediate feedback).
Who it's for: anyone making career or skill-development decisions; anyone evaluating the depth-vs-breadth question for themselves or their team.
## Where should you start? A 5-book reading order1. Predictably Irrational (Dan Ariely) - gateway to behavioural economics, fast and memorable. Sets up the why.
2. The Art of Thinking Clearly (Rolf Dobelli) - fast field guide to 99 cognitive errors. Lets you start spotting the patterns in your own thinking.
3. Decisive (Chip and Dan Heath) - the most practical decision framework on the list. WRAP gives you a tool to use immediately.
4. Nudge (Thaler and Sunstein) - choice architecture and the systems-level view. Once you've seen biases in yourself, you start to notice them in environments.
5. Algorithms to Live By (Christian and Griffiths) - the playful, mathematical complement. Gives you decision rules from computer science you'll actually use.
These five cover the landscape: descriptive (what goes wrong), prescriptive (what to do instead), structural (how environments shape choice), and algorithmic (provably optimal rules for specific problems). After these, our 12 best books on probabilistic thinking covers the more mathematical and probability-leaning side of the same field - and if you only read one book from that list, make it Annie Duke's Thinking in Bets.
## Which books did we deliberately leave off?A few well-known titles aren't on this list, and it's worth saying why.
The 7 Habits of Highly Effective People is more a self-help / character book than a decision-making one. Useful, but a different genre.
Atomic Habits (James Clear) overlaps heavily with The Power of Habit; we've gone with Duhigg as the older and more thoroughly research-led of the two. If you've already read The Power of Habit, Atomic Habits is largely redundant.
Blink (Malcolm Gladwell) makes interesting claims about rapid intuition but several of its central case studies have been undermined by later research. Sources of Power covers similar ground more rigorously.
Freakonomics is great pop economics but isn't really about decision-making methodology - it's about applying empirical analysis to surprising questions.
## Frequently asked questions **Q: What's the single best decision-making book to start with?** Decisive by Chip and Dan Heath, if you want a practical framework you can use immediately. Predictably Irrational by Dan Ariely, if you'd rather start with what goes wrong before learning what to do instead. **Q: How is this list different from your probabilistic-thinking list?** Our other list focuses on books that teach probability, expected value, and statistical thinking - Thinking, Fast and Slow, Superforecasting, Fooled by Randomness, and similar. This list covers behavioural economics, choice architecture, decision frameworks, habits, and motivation. The two lists are designed to be complementary rather than overlapping. **Q: Is Atomic Habits worth reading if I've read The Power of Habit?** Marginally. The core ideas are very similar, with Atomic Habits adding more concrete implementation tactics (habit stacking, environment design, …[truncated for length] --- # Expected Value Formula: Derivation and 5 Worked Examples Source: https://expectedvalue.co.uk/blog/expected-value-formula/ # Expected Value Formula: Derivation and 5 Worked Examples The expected value formula is E[X] = Σ(p × x). Full derivation and five worked examples - coin flip, lottery, insurance, poker, and stock investment. **Answer** The expected value formula is EV = sum of (probability x outcome) across all possible outcomes. For a single fair die paying out £1 per dot rolled, EV = (1/6)(£1) + (1/6)(£2) + ... + (1/6)(£6) = £3.50 per roll. The 5 worked examples below extend the formula to weighted dice, multi-stage bets, and decisions with both positive and negative outcomes.The expected value formula for a discrete random variable is E[X] = Σ(p × x) - the sum, over every possible outcome, of that outcome's value multiplied by its probability. It is the most-used formula in probability theory and the foundation for every rational decision under uncertainty. This page derives the formula from first principles, then works through five examples - a coin flip, a UK lottery ticket, a home-insurance policy, a poker hand, and a stock investment - so you can see exactly how it applies in each domain.
Discrete random variable: E[X] = Σ pᵢ × xᵢ - sum of each outcome's value weighted by its probability.
Continuous random variable: E[X] = ∫ x × f(x) dx - integral of the value times the probability density function over its full range.
For all practical decision-making, the discrete form is what you need. The continuous form is for derivatives pricing, statistical inference, and physics. The two are the same idea - "average over the distribution, weighted by probability" - written for different kinds of outcome space.
## What does the expected value formula actually say?Expected value is the long-run average of a random outcome. If you repeated the same uncertain decision an infinite number of times, the average of your results would converge to E[X]. That convergence is guaranteed by the law of large numbers - provided the variance of X is finite.
The formula encodes a simple idea: outcomes that are more likely should count for more, and outcomes that are less likely should count for less. The probability is the weight; the outcome value is what gets weighted. Sum them all up and you have the average you would expect over many repetitions.
Three things to notice. First, the sum runs over every possible outcome - miss one and the formula gives the wrong answer. Second, the probabilities must add to 1, otherwise the calculation is meaningless. Third, expected value can equal a value the variable never actually takes - the expected value of a fair six-sided die is 3.5, which is not a face on the die.
## How is the EV formula derived from first principles?Imagine a game with three possible outcomes: you win £10, you win £0, or you lose £5. Suppose the probabilities are 30%, 50%, and 20% respectively (they sum to 100%, as required). What is the average outcome?
Play the game 100 times. Statistically, you would expect roughly 30 wins of £10 (= £300), 50 nothings (= £0), and 20 losses of £5 (= −£100). Total: £200. Average per game: £2.
The shortcut to that calculation is exactly the expected value formula. The 30 wins of £10 sum to £300, which is the same as 100 × 0.30 × £10. The 20 losses of £5 sum to −£100, which is 100 × 0.20 × (−£5). Add them together, divide by 100, and you get:
E[X] = 0.30 × £10 + 0.50 × £0 + 0.20 × (−£5) = £3 + £0 − £1 = £2
The formula falls out of the long-run average by definition. There is nothing magical about it - it is what "average" means when outcomes happen with different probabilities.
## Example: what is the EV of a fair coin flip bet?A friend offers you a bet: flip a fair coin. Heads, you win £15. Tails, you lose £10. Should you take it?
Compute the expected value. There are two outcomes, each with probability 0.5.
E[X] = 0.5 × £15 + 0.5 × (−£10) = £7.50 − £5.00 = £2.50
Every time you play, you expect to make £2.50 on average. Over a single flip, you will either gain £15 or lose £10 - those are the only two possibilities. But over a thousand flips, your total winnings should be roughly £2,500. Take the bet.
This is the cleanest illustration of why expected value matters. The bet pays out an asymmetric amount because the probabilities of both outcomes are equal, so the higher payoff dominates. The same expected value calculation generalises to any odds and any payout structure.
## Example: what is the EV of a UK lottery ticket?The UK National Lottery's Lotto draw requires you to match six numbers from 1 to 59. The probability of matching all six is 1 / C(59, 6) = 1 / 45,057,474. A ticket costs £2 and the jackpot averages around £4 million.
For simplicity, consider only the jackpot outcome and the no-win outcome. The expected value of one ticket is:
E[X] = (1 / 45,057,474) × £4,000,000 + (45,057,473 / 45,057,474) × £0 − £2
E[X] ≈ £0.089 − £2 = −£1.91
Every ticket has an expected value of around negative £1.91. You are paying £2 for an asset worth about 9 pence. Over many years, lottery players lose 96% of what they spend. The other prize tiers (matching five or four numbers) lift the figure slightly but not enough to make the bet positive - under almost any jackpot, the net expected value stays well below the ticket price.
This is one of the cleanest demonstrations that humans are not natural expected-value calculators. We respond to the size of the prize rather than its probability - see the availability heuristic for the psychology behind why lottery wins feel more vivid than the millions of losing tickets that funded them.
## Example: what is the EV of a home insurance policy?You pay £350 a year for buildings insurance on a £300,000 home. The probability of a total loss (fire, flood, structural collapse) in any given year is roughly 0.001 - one in a thousand homes. Smaller claims (storm damage, burst pipes) have higher probability, but assume the policy primarily exists to cover catastrophic loss.
The expected payout to you in any year is:
E[payout] = 0.001 × £300,000 + 0.999 × £0 = £300
Your expected value from buying the insurance is £300 − £350 = −£50. Every year you expect to lose £50 on average. Yet buying the policy is the rational choice - because losing your £300,000 home would be financially ruinous in a way that paying £350 a year is not.
This is the most important practical insight in expected-value thinking. EV is not the whole story when outcomes are non-linear in value - when a £300,000 loss hurts you more than 1,000 times as much as a £300 loss. The framework that handles this properly is expected utility, which weights outcomes by their personal impact rather than their monetary face value.
## Example: what is the EV of a poker hand decision?You are playing No-Limit Hold'em. The pot is £80 and your opponent has bet £40, making the total pot £120. You must call £40 to continue. You estimate (using conditional probability) that your hand has a 35% chance of winning at showdown.
There are two outcomes from calling: you win the £120 currently in the pot, or you lose your £40 call. Apply the formula:
E[call] = 0.35 × £120 + 0.65 × (−£40) = £42 − £26 = £16
The call has positive expected value of £16. Make the call. Every time you face this exact situation, you will sometimes win and sometimes lose, but the average across many such hands will be a £16 gain per call.
This is the calculation behind every poker decision worth taking seriously. The crucial input is your equity estimate - the 35% above. Get that wrong and the formula gives the wrong answer. The skill in poker is partly about working out your equity quickly and honestly; once you have it, the expected-value arithmetic is mechanical. See expected value in poker for more applications.
## Example: what is the EV of a stock investment?You are considering a £10,000 investment in a single stock. Three scenarios:
- **Bull case (probability 30%)**: Stock returns +40% → £14,000 - **Base case (probability 50%)**: Stock returns +8% → £10,800 - **Bear case (probability 20%)**: Stock returns −30% → £7,000Apply the expected value formula:
E[final value] = 0.30 × £14,000 + 0.50 × £10,800 + 0.20 × £7,000
E[final value] = £4,200 + £5,400 + £1,400 = £11,000
The expected final value is £11,000, so the expected return is (£11,000 − £10,000) / £10,000 = 10%. On expected-value grounds, the investment looks attractive - provided the input probabilities and outcome values are honest estimates.
But expected value alone does not capture risk. The same expected return is available from a fixed-income product with no bear-case loss, and most rational investors would prefer that. The full framework needs to account for the distribution around the expected value, which is what variance and Sharpe ratio capture. For position sizing on an investment like this, see the Kelly Criterion calculator.
## When does the EV formula give the wrong answer?Expected value is the right number to maximise across a long sequence of independent, repeatable bets where the outcomes are linear in their personal impact. Three situations break those assumptions and require a different framework.
One-shot decisions with ruinous downside. If a single outcome would wipe you out (financially, professionally, medically), you should not maximise EV. The insurance example above is the classic case - the expected value of buying the policy is negative, but the policy is rational because losing the house is not 1,000 times as bad as losing the premium, it is infinitely worse. Expected utility handles this case.
Non-ergodic systems. Some processes have a time-average that diverges from the ensemble-average. The classic example is a 50% gain followed by a 50% loss: on average across many traders the wealth stays flat, but a single trader playing repeatedly trends toward zero. See ergodicity explained for why this matters for compounding investments.
Adversarial settings with shared information. If you are taking a market price as input to your EV calculation, the price already encodes other people's EV estimates. Acting on EV alone in markets means systematically betting against people with better information than you - see risk versus uncertainty for the distinction Knight drew between the two in 1921.
## Frequently asked questions **Q: What does Σ mean in the expected value formula?** Σ is the Greek capital sigma - the mathematical notation for "sum of".Σ pᵢ × xᵢ means "add up the product of probability and value for every outcome i". If the variable can take values x₁, x₂, x₃ with probabilities p₁, p₂, p₃, the expected value is p₁x₁ + p₂x₂ + p₃x₃.
**Q: How is expected value different from the mean?**
For a probability distribution, they are the same number. The mean of a random variable is its expected value. The distinction matters when you have a sample of observations: the sample mean is your estimate of the population mean, which is the expected value. With infinite data, sample mean converges to expected value (the law of large numbers).
**Q: Can expected value be negative?**
Yes - and that is often the most useful answer the formula gives you. A negative expected value tells you the bet is unfavourable. Every lottery ticket, slot machine, and casino game has a negative expected value for the player. Recognising this is the first step to making rational decisions about whether to play.
**Q: What if I don't know the probabilities?**
You almost never do, exactly. Expected value calculations in real life use estimates - your best guess at the probabilities. The framework is still useful if your estimates are calibrated, even if they are imprecise. See probability calibration training for how to test whether your estimates are honest.
**Q: How does the formula change for continuous outcomes?**
The sum becomes an integral and the probabilities become a probability density function: E[X] = ∫ x × f(x) dx, integrated over the full range of possible outcomes. The intuition is identical - average over the distribution, weighted by probability density. In practice you rarely calculate the integral by hand; you use the known formula for the relevant distribution (normal, log-normal, exponential, etc.).
**Q: Why does the formula multiply rather than add probability and value?**
Because the probability is the weight applied to that outcome's value. "30% chance of £10" contributes £3 to the long-run average, not £10.30. Multiplication encodes the idea that an outcome that happens 30% of the time should count for 30% of its full value when you compute the average across many trials.
## Related reading
A Kelly Criterion calculator converts three inputs - your edge, the odds on offer, and your bankroll - into a single number: the percentage you should stake on this bet. The formula was published by Bell Labs researcher John Kelly in 1956 and is the mathematically optimal sizing rule for any positive-expected-value wager you can repeat. This guide gives you the formula, then walks through three calculators side-by-side: one for sports betting with decimal odds, one for investing using Sharpe-style inputs, and one for poker bankroll sizing. Worked examples throughout - plug your own numbers into each.
**TL;DR**For a simple win/lose bet at decimal odds d with true win probability p, the Kelly fraction is f* = (p × (d − 1) − (1 − p)) / (d − 1). For a continuous investment with excess return μ and variance σ², it is f* = μ / σ². In both cases, f* is the share of your bankroll to risk on each bet. Most practitioners use half or quarter of this number to reduce volatility - see fractional Kelly below for why.
The Kelly Criterion answers a precise question: given a sequence of bets with a positive expected value, what fraction of your bankroll should you stake on each one to maximise the long-run growth rate of that bankroll?
This is a different question from "what maximises expected profit?" Expected profit is maximised by betting everything on the highest-EV opportunity available. Long-run growth is maximised by Kelly - because growth is multiplicative across bets, and a 100% loss takes you out of the game forever regardless of your edge.
The intuition is straightforward. Each bet multiplies your bankroll by some random factor. Over many bets, your final bankroll is the product of all those factors. Maximising that product is equivalent to maximising the sum of their logarithms - and the Kelly fraction is exactly the stake size that maximises expected log-growth per bet.
If this is your first exposure to the framework, the full Kelly Criterion explainer covers the derivation and the philosophical case for it. This page is the practical calculator.
## What is the Kelly formula and what do its variables mean?The discrete Kelly formula - for a single bet that either wins or loses - has several equivalent forms, depending on whether you express the odds as decimal, fractional, or as a payout ratio. They all reduce to the same number.
The most general form, using a payout ratio b (net profit per unit staked if you win):
f* = (b × p − q) / b
Where:
- **f***: Optimal fraction of bankroll to stake (0 to 1) - **p**: True probability of winning, your honest estimate (0 to 1) - **q**: Probability of losing, equal to 1 − p - **b**: Net odds received on the wager (decimal odds − 1)If f* comes out negative, the bet has negative expected value and Kelly says do not bet at all. If it comes out above 1, the formula is telling you to stake more than your bankroll - meaning the edge is so large that capital constraints, not Kelly, are the binding limit.
For continuous outcomes (investing in an asset with a distribution of returns rather than a binary win/lose), the formula generalises to f* = (μ − r) / σ², where μ is the asset's expected return, r is the risk-free rate, and σ² is the return variance. This is the version most relevant for portfolio sizing.
Bookmakers quote decimal odds: 2.20 means a £1 stake returns £2.20 (your stake plus £1.20 profit). To plug decimal odds d into the Kelly formula, set b = d − 1. The simplified form is:
f* = (p × (d − 1) − (1 − p)) / (d − 1)
This is the calculator most matched-betting and value-betting communities use. The hard part is estimating p honestly - see the overconfidence bias piece for why most punters overstate their edge.
Bookmaker prices Liverpool to win at decimal odds of 2.20. After modelling, you believe the true win probability is 0.52 (a 4 percentage-point edge over the implied 45.5%).
Step 1: b = 2.20 − 1 = 1.20
Step 2: p = 0.52, q = 0.48
Step 3: f* = (1.20 × 0.52 − 0.48) / 1.20 = (0.624 − 0.48) / 1.20 = 0.144 / 1.20 = 0.12
Result: stake 12% of your bankroll on this bet. On a £1,000 bankroll that is a £120 stake.
If your true probability estimate were only 0.46 instead of 0.52, the calculation becomes (1.20 × 0.46 − 0.54) / 1.20 = 0.0072 / 1.20 = 0.6% - six tenths of a percent. A small change in your edge produces a large change in the optimal stake.
Two extensions are worth knowing. First, if you are betting into a market where the bookmaker takes a commission (a betting exchange like Smarkets or Betfair takes 2–5% on winnings), reduce b accordingly: b = (d − 1) × (1 − commission). Second, for multi-outcome markets such as football match results (home / draw / away), the standard Kelly does not apply directly - use the simultaneous-Kelly formulation, which adjusts each stake for the correlation between outcomes.
For investments where returns are continuously distributed rather than binary, the discrete Kelly formula does not apply. The continuous version uses expected excess return and variance:
f* = (μ − r) / σ²
Where μ is the expected annual return, r is the risk-free rate (UK gilts, US Treasuries, your savings rate), and σ is the standard deviation of returns. σ² is variance - square the volatility figure.
This is mathematically equivalent to f* = Sharpe / σ, which is how some portfolio managers prefer to think about it: position size scales with the Sharpe ratio and inversely with volatility.
Long-run historical figures for the S&P 500 are roughly: expected real return of around 7%, real risk-free rate near 1%, annualised volatility around 16%. So μ − r ≈ 0.06 and σ² ≈ 0.0256.
Step 1: f* = 0.06 / 0.0256 ≈ 2.34
Full Kelly is telling you to take a 234% position - meaning lever the portfolio 2.34× into equities. That is dramatic. In practice almost no individual investor does this - see the next section on why fractional Kelly is the working answer.
At half Kelly (1.17×), the recommendation drops to roughly a 117% equity position. At quarter Kelly (0.58×), around 58% equities and 42% bonds or cash - which lines up neatly with the 60/40 portfolio most pension trustees default to. The 60/40 is, in effect, a quarter-Kelly bet on equities under historical S&P inputs.
Two cautions on this version. The continuous Kelly assumes returns are roughly log-normal and the inputs are stable. Both assumptions break in crises - see ergodicity for why long-run averages can lie about your actual experience. And the formula assumes you know μ and σ². You do not. Both are estimated from historical data, with sampling error that swamps any apparent precision in the calculator output.
For deeper treatment of the investing application, see position sizing with Kelly, which works through stock-specific and portfolio-level versions.
## Calculator: Kelly bankroll fraction at the poker tablePoker is a multi-outcome game with very high variance relative to expected return, which is why most poker bankroll advice ends up well below full Kelly without explicitly invoking the formula.
For a cash-game session with winrate w in big blinds per 100 hands and standard deviation σ (also in bb/100), the per-hand variance is σ²/100 and per-hand EV is w/100. The Kelly fraction per hand is:
f* = w / σ²
This gives the fraction of bankroll to risk on each hand. To convert to buy-ins, divide a typical 100bb buy-in by the bankroll at which one buy-in equals that Kelly fraction.
**Worked example - NL cash game**A solid mid-stakes player runs at 2bb/100 winrate with a typical standard deviation of 100bb/100. So w = 2, σ² = 10,000.
Per hand: f* = 2 / 10,000 = 0.0002 (0.02% of bankroll per hand).
If a buy-in is 100bb and represents 0.02% × N hands of risk, full Kelly works out to roughly 5 buy-ins as a minimum bankroll. Almost no professional plays that thin - the standard recommendation is 20 to 40 buy-ins for cash games, which corresponds to running at quarter to one-eighth Kelly.
Why? Because in poker, your winrate estimate is unreliable (sample sizes need to be in the hundreds of thousands of hands for statistical confidence) and variance is high. The downside of overestimating w is catastrophic; the downside of being too conservative is just slower growth. The traditional buy-in conventions encode this asymmetry.
Tournament poker is harder still because the variance is dominated by rare large outcomes (top-three finishes) rather than per-hand swings. Most professional tournament players target 100+ buy-ins for the level they play - closer to one-tenth Kelly. The general principle: as variance rises relative to edge, the fraction of Kelly you should use falls. For the expected-value calculations behind poker decisions themselves, see expected value in poker.
## Why should most people use fractional Kelly?Full Kelly maximises long-run growth in theory. In practice almost every professional uses some fraction of it - typically half ("half Kelly") or quarter ("quarter Kelly"). There are two reasons.
The first is estimation error. Kelly assumes you know p exactly. You do not - you have an estimate of p with some uncertainty around it. Because the cost of overestimating p is dramatically worse than the cost of underestimating it (the formula is asymmetric around the true value), the rational response to estimation uncertainty is to bet less than the formula suggests.
The second is variance tolerance. Full Kelly produces extraordinary drawdowns. Edward Thorp's analysis showed that under full Kelly, the probability of seeing your bankroll halve at some point before doubling is roughly 50%. Most people, even professional investors and bettors, find that level of volatility psychologically untenable - and the worst time to abandon a strategy is during a drawdown, which converts paper losses into permanent ones.
Half Kelly is the working compromise. It captures around 75% of full Kelly's growth rate with roughly half the variance. Quarter Kelly drops growth to about 44% of full but cuts variance another half. For most practical applications - sports betting, portfolio management, poker bankroll - half or quarter is where you should set the calculator output.
**Common calculator mistakes**1. Overstating your edge. The most common error. Be honest about p. If you have not back-tested or independently calibrated your probability estimates, assume your edge is half what you think it is. See probability calibration training for how to test this.
2. Treating implied probability as true probability. Bookmaker odds embed both an estimate of the true probability and a margin (the overround). Implied probability from odds is not your edge - your edge is the gap between your estimate and the bookmaker's true estimate (which is closer to the closing line than the opening price).
3. Forgetting correlation in portfolios. If you are sizing five positions and they all move with the market, the Kelly fraction for each in isolation overstates the total risk. Treat correlated bets as one larger bet for sizing purposes.
4. Applying full Kelly to one-off decisions. Kelly maximises long-run growth across many bets. For a one-shot decision, expected utility (with a more concave utility function than log) is usually the better framework. See expected value vs expected utility.
## Frequently asked questions **Q: Is the Kelly Criterion calculator the same for fractional and decimal odds?** Yes - they are different notations for the same number. Convert fractional oddsa/b to decimal via (a/b) + 1, then use the decimal formula. So 11/10 fractional becomes 2.10 decimal, and the calculator input is b = 1.10.
**Q: What does Kelly tell me to do when my edge is negative?**
If f* comes out below zero, do not bet. The market is offering you worse odds than your true probability estimate. Negative-Kelly results are how the formula tells you the bet has negative expected value.
**Q: Why does the formula say I should bet more than 100% of my bankroll?**
Because the edge is so large that capital, not Kelly sizing, is the binding constraint. The formula does not know about leverage limits or wagering minimums - it just outputs the optimal fraction. In practice, cap any Kelly recommendation at 25% for single bets and 100% for total portfolio exposure unless you are explicitly using leverage.
**Q: Should I use Kelly for index-fund investing?**
The continuous Kelly formula applied to a broad index typically recommends well over 100% equities (i.e. leverage). Almost no individual investor implements this - partly for psychological reasons (the volatility is severe), partly because the input estimates (μ and σ²) are unstable. A 60/40 portfolio corresponds to roughly quarter-Kelly under historical S&P inputs, which is a defensible interpretation of the framework.
**Q: How does Kelly handle multiple simultaneous bets?**
If the bets are independent, you can use Kelly on each one separately and the total stake naturally adds up. If they are correlated - say, three bets on the same football match - you need the simultaneous-Kelly formulation, which solves the joint optimisation. As a heuristic, treat strongly correlated bets as a single combined wager and size the total accordingly.
**Q: Who invented the Kelly Criterion and where can I read the original?**
John L. Kelly Jr., a researcher at Bell Labs, published "A New Interpretation of Information Rate" in the Bell System Technical Journal in 1956. The framing was about information transmission over noisy channels; the bet-sizing application was a worked example. Edward Thorp later popularised it as a betting and investing tool - his book Beat the Dealer (1962) was the first commercial application.
## Related reading
- **The Kelly Criterion explainer** - The full derivation and intuition behind the formula
- **Position sizing with Kelly** - Stock-specific and portfolio-level applications
- **Expected value calculator** - Compute the EV of a bet before sizing it
- **Expected value vs expected utility** - When Kelly is the wrong framework - one-off decisions
- **Probability calibration training** - How to test whether your edge estimates are honest
Bayesian and frequentist statistics are two answers to the same question: what does "probability" actually mean? Frequentists treat it as the long-run frequency of an event in repeated trials. Bayesians treat it as a degree of belief that can be updated with evidence. From that single philosophical split fall almost every practical difference between the two - p-values versus posterior distributions, confidence intervals versus credible intervals, null-hypothesis tests versus Bayes factors.
The split matters because it changes what your numbers mean. A 95% confidence interval and a 95% credible interval look identical on a chart and answer subtly different questions. Picking the wrong framework for the problem you have produces results that are technically correct and practically misleading.
**TL;DR**Frequentist statistics treats parameters as fixed unknowns and data as random - it asks how surprising would this data be if the null were true? Bayesian statistics treats parameters as uncertain quantities with their own distributions and data as fixed once observed - it asks given this data, what is the probability of each possible parameter value? Use frequentist for repeatable experiments with clean designs; use Bayesian for one-off decisions where prior information matters.
## The Philosophical SplitThe disagreement starts with the definition of probability itself.
The frequentist view: probability is the long-run relative frequency of an event in an infinitely repeated experiment. Saying "this coin has a 50% probability of heads" means "if you flip it forever, the fraction of heads converges to 0.5." Under this definition, statements like "there is a 70% probability that the drug works" are nonsense - the drug either works or it does not; the truth is a fixed, unknown parameter, not a random variable. Probabilities only apply to repeatable random events, not to fixed states of the world.
The Bayesian view: probability is a quantified degree of belief, calibrated against evidence. Saying "there is a 70% probability that the drug works" is a statement about what you know, not about what is. The drug's true effect is fixed but unknown; the 70% is your uncertainty about it given the evidence you have. New data updates that belief via Bayes theorem.
Both views are internally consistent. They disagree about which kinds of statements are meaningful, which methods are admissible, and how you should phrase a conclusion. Once you commit to one, the rest of the machinery follows.
## What the Two Frameworks Actually Compute ## Confidence Intervals vs Credible IntervalsThis is the textbook example of two methods producing similar-looking numbers that mean different things.
A 95% frequentist confidence interval is a statement about the procedure, not the parameter. It means: if you ran this experiment many times and constructed an interval the same way each time, 95% of those intervals would contain the true parameter. It does not mean there is a 95% probability that the true parameter is in this particular interval - the parameter is fixed, so any specific interval either contains it or does not, with no probability attached.
A 95% Bayesian credible interval is a statement about the parameter. It means: given the data and the prior, there is a 95% probability that the true parameter lies in this interval. This is the statement most people incorrectly attribute to confidence intervals.
For large samples with weak priors, the two intervals are often nearly identical numerically. They are answering different questions and the interpretive labels matter - confusing them is one of the most common errors in applied statistics.
## When Frequentist Methods WinFrequentism is the right tool when the problem genuinely fits its assumptions:
Bayesian methods are the right tool whenever the data are limited, the priors are informative, or the question is not naturally repeatable:
To make the difference concrete, take a small, contrived example. A coin is flipped 10 times and lands heads 7 times. Is it biased?
Frequentist analysis. The null hypothesis is that the coin is fair (p = 0.5). Under the null, the probability of seeing 7 or more heads in 10 flips is given by the binomial tail: about 0.172. That is the p-value. At a 5% significance level (p < 0.05), the null is not rejected. The frequentist conclusion: the data are not inconsistent with a fair coin. No probability is attached to the hypothesis itself.
Bayesian analysis. Start with a uniform prior on the coin's bias (Beta(1, 1) - any bias is equally likely a priori). The posterior after observing 7 heads in 10 flips is Beta(8, 4). From that posterior we can compute the probability that the coin is biased toward heads: P(p > 0.5 | data) ≈ 0.89. The 95% credible interval for p is roughly (0.42, 0.89). The Bayesian conclusion: the data raise the probability the coin is biased toward heads to about 89%, but the credible interval is wide because 10 flips is not many.
Same data, two answers. The frequentist answer is a statement about how surprised we should be under the null; the Bayesian answer is a statement about what we believe about the coin. Neither is wrong. They answer different questions.
## The Common Misinterpretations - **Treating a p-value as the probability the null is true** - P(data | null) is not P(null | data). The two are linked by Bayes theorem but are almost never equal. A p-value of 0.04 does not mean there is a 4% chance the null is true - that statement requires a prior, which the p-value framework deliberately excludes. - **Calling a 95% confidence interval a 95% probability range** - The 95% refers to the procedure, not the parameter. A specific interval either contains the parameter or does not. The credible-interval interpretation people instinctively reach for is the Bayesian one - accidental Bayesianism wearing a frequentist label. - **Picking the prior to get the answer you want** - Bayesian methods are sometimes accused of being "subjective" because the prior is an input. The defence is to use weakly informative priors by default and to perform sensitivity analyses showing how the posterior changes if the prior is varied. A robust posterior survives across a range of reasonable priors; a fragile one does not, and the fragility itself is informative. - **Believing frequentism is prior-free** - Frequentist methods make many implicit modelling choices - likelihood functions, link functions, error distributions, stopping rules - that play the same role as priors. The choice is just less visible because it is buried in the model rather than stated explicitly. Neither framework is assumption-free. - **Stopping a frequentist experiment early because the result "looks good"** - Optional stopping in frequentist NHST inflates the false-positive rate dramatically. Bayesian methods are immune to this because the posterior reflects whatever data you have. Mixing the two cultures - running a frequentist test but peeking like a Bayesian - produces the worst of both worlds. ## How to Choose for a Specific Problem 1. **Ask what kind of question you are answering** - A statement about a procedure ("if I ran this 1,000 times, what fraction would reject the null?") is frequentist by construction. A statement about the world ("what is the probability this effect is positive?") is Bayesian by construction. The grammar of the question often dictates the framework. 2. **Ask whether prior information exists and is worth using** - If there is a meaningful base rate, historical data, or expert judgment, encoding it as a prior produces sharper inferences. If there is genuinely nothing prior to encode, the two frameworks tend to agree numerically and frequentism's syntactic simplicity is a fair tiebreaker. 3. **Ask whether the audience can interpret the output** - Posterior probabilities and credible intervals are directly interpretable; p-values and confidence intervals require careful framing. If the reader will reach for the wrong interpretation either way, lean toward whichever framework makes the wrong interpretation accidentally correct. 4. **Ask whether the design allows peeking** - If the experiment is sequential or might be stopped early, Bayesian methods preserve their interpretation; frequentist methods do not. Either commit to a pre-registered sample size and stick to it, or use a Bayesian framework that supports continuous monitoring. 5. **Ask what the regulatory or publishing environment expects** - Pharmaceutical trials, peer-reviewed medical research, and most economics journals still expect frequentist analyses. Tech-product A/B testing, modern machine learning, and bespoke decision support are increasingly Bayesian. Pick the framework that matches the audience, then use the other internally as a sanity check. ## Why the Argument Is Less Heated Than It Used to BeFor most of the twentieth century, the Bayesian–frequentist split was an ideological war. Frequentists accused Bayesians of subjectivism; Bayesians accused frequentists of incoherence. Two things have largely defused the conflict.
First, computational advances. Bayesian inference used to be analytically tractable only for a handful of conjugate-prior cases. Markov chain Monte Carlo, variational inference, and modern probabilistic programming languages (Stan, PyMC, NumPyro, Pyro) made it possible to fit Bayesian models for realistic problems on a laptop. Without that, much of applied statistics had no Bayesian option even if a practitioner wanted one.
Second, methodological humility. The replication crisis showed how easy it is to misuse p-values - most spectacularly, by stopping data collection when a test happened to be significant. The Bayesian world is not immune to bad practice, but the explicit prior and the continuous posterior make some of the worst frequentist failure modes harder to commit by accident. The professional consensus is now closer to "use whichever framework fits the problem and the audience" than to either camp's nineteenth-century certainties.
## Frequently Asked Questions **Q: What is the main difference between Bayesian and frequentist statistics?** Frequentists define probability as the long-run frequency of an event in repeated trials and treat parameters as fixed unknowns. Bayesians define probability as a degree of belief that can be updated with evidence and treat parameters as having their own distributions. Almost every methodological difference between the two flows from that single split. **Q: Is Bayesian or frequentist better?** Neither is universally better. Frequentist methods are well-suited to repeatable experiments with clean designs and regulatory contexts that require methodological objectivity. Bayesian methods are well-suited to one-off decisions, problems with strong prior information, and situations involving sequential evidence. **Q: What is a p-value, and what does it actually mean?** A p-value is the probability of observing data at least as extreme as the actual data, assuming the null hypothesis is true. It is not the probability the null hypothesis is true, and it is not the probability of getting the same result if the experiment were repeated. Most popular interpretations of p-values are technically wrong. **Q: What is the difference between a confidence interval and a credible interval?** A confidence interval is a frequentist construct: if you repeated the experiment many times, 95% of intervals constructed the same way would contain the true parameter. A credible interval is a Bayesian construct: given the data and prior, there is a 95% probability the true parameter lies in this specific interval. They look similar numerically but answer different questions. **Q: Why do regulators usually require frequentist analyses?** Because the prior in a Bayesian analysis is an input the analyst chooses, regulators worry it can be tuned to favour a desired result. Frequentist methods sidestep this by not having a prior at all - at the cost of being unable to formally incorporate genuine prior information. The trade-off is methodological objectivity in exchange for sometimes-suboptimal inference. **Q: Can Bayesian methods peek at data without inflating error rates?** Yes. Bayesian posteriors reflect whatever data has been observed so far; there is no notion of "alpha spending" or "sample size set in advance." This makes Bayesian methods natural for sequential trials, adaptive A/B testing, and any situation where stopping rules are flexible. The frequentist equivalent requires careful pre-registration and stopping-rule design. **Q: Do Bayesian and frequentist methods ever give the same answer?** Often, yes - particularly with large samples and weakly informative priors. The numerical answer converges; the interpretation does not. A 95% confidence interval and a 95% credible interval at large n might be the same two numbers and still mean different things. ## Further Reading on This SiteBayes theorem is the mathematical rule for updating a belief when new evidence arrives. It takes a prior probability, a likelihood, and a base rate, and returns a posterior probability - the answer to the question "given what I just saw, how likely is the thing I care about?"
It is short, exact, and almost universally counter-intuitive. The same equation that powers spam filters and medical diagnostics also explains why most positive cancer screens are false alarms, why courtroom statistics get reversed on appeal, and why so many forecasts feel obviously wrong in hindsight. The maths is high-school algebra. The hard part is the framing.
**TL;DR**Bayes theorem: P(A | B) = P(B | A) × P(A) / P(B). Translated: the probability of A given B equals the probability of B given A, scaled by how common A is to begin with, divided by how common B is overall. The base-rate term P(A) is where almost every wrong intuition lives.
## What is the actual Bayes theorem formula?Written formally:
P(A | B) = P(B | A) × P(A) / P(B)
Every symbol has a name:
An equivalent expanded form is often more useful because it forces you to write out the denominator:
P(A | B) = [P(B | A) × P(A)] / [P(B | A)P(A) + P(B | ¬A)P(¬A)]
## How is Bayes theorem derived in two lines?Bayes theorem falls out of the definition of conditional probability with one substitution. Start with the standard definition:
P(A | B) = P(A and B) / P(B)
By symmetry, the joint probability P(A and B) also equals P(B and A) = P(B | A) × P(A). Substitute that into the numerator:
P(A | B) = [P(B | A) × P(A)] / P(B)
That is the entire derivation. The formula is not a new principle - it is a re-arrangement of the conditional-probability definition that puts the term we usually want (the posterior) on the left, and the terms we usually have (likelihood, prior, marginal) on the right. For a refresher on the underlying definition, see conditional probability explained.
## Example 1: how does Bayes theorem solve the medical-test trap?This is the canonical Bayes example because the answer is so wrong-feeling. A disease affects 1 in 1,000 people. A diagnostic test is 99% sensitive (correctly flags 99% of true positives) and 99% specific (correctly clears 99% of true negatives). You test positive. What is the probability you actually have the disease?
Intuition says "about 99%." The actual answer is roughly 9%. Here is the calculation:
Plug into the expanded form:
P(A | B) = (0.99 × 0.001) / [(0.99 × 0.001) + (0.01 × 0.999)]
= 0.00099 / (0.00099 + 0.00999) = 0.00099 / 0.01098 ≈ 0.0902
So a positive result on a 99%-accurate test for a rare disease means about a 9% chance you actually have it. The other 91% of positives are false alarms from the much-larger pool of healthy people. This is the false positive paradox, and Bayes theorem is the cleanest way to see it.
**Why the intuition fails**People mentally collapse P(disease | positive) and P(positive | disease) into the same number. They are not the same. When the base rate is tiny, the denominator is dominated by false positives from the healthy majority, and the posterior is far lower than the test's accuracy suggests. This confusion is so common it has a name: the prosecutor's fallacy.
## Example 2: how does Bayes theorem power spam filtering?Naive Bayes classifiers - the original spam filters - apply this formula word by word. Suppose across a corpus of emails, 30% are spam (P(spam) = 0.30). The word "lottery" appears in 8% of spam emails (P("lottery" | spam) = 0.08) and in 0.5% of ham emails (P("lottery" | ham) = 0.005). An incoming email contains the word "lottery." How spammy is it on that signal alone?
P(spam | "lottery") = (0.08 × 0.30) / [(0.08 × 0.30) + (0.005 × 0.70)]
= 0.024 / (0.024 + 0.0035) = 0.024 / 0.0275 ≈ 0.873
So one word lifts the posterior from a 30% prior to an 87% posterior. The real classifier multiplies these likelihoods across many words (treating them as conditionally independent - the "naive" part), which is mathematically rough but empirically excellent for text classification.
## Example 3: how does sequential Bayesian updating work?Bayes is iterative. The posterior from one observation becomes the prior for the next. Suppose a company assesses a job candidate as having a 40% prior probability of being a strong hire. The technical interview is a noisy signal: a real strong hire passes 80% of the time; a weak hire passes 30% of the time. The candidate passes.
First update:
P(strong | pass) = (0.80 × 0.40) / [(0.80 × 0.40) + (0.30 × 0.60)] = 0.32 / 0.50 = 0.64
The candidate then passes a behavioural round with similar signal strength (80% pass rate among strong, 30% among weak). Plug 0.64 in as the new prior:
P(strong | both pass) = (0.80 × 0.64) / [(0.80 × 0.64) + (0.30 × 0.36)] = 0.512 / 0.620 = 0.826
Two noisy positive signals pushed the belief from 40% to 83%. This is what "updating in light of evidence" looks like in numbers, and it is the practical engine behind Bayesian thinking in everyday decisions.
## Example 4: how do you reverse a forecast with Bayes?A weather model predicts rain tomorrow with 70% confidence. Historically the model is well-calibrated: when it says 70%, rain happens 70% of the time. But the regional climatology for this date is dry - only 15% of historical days in this week have had rain. What posterior should you actually hold?
This case has no observation to update on yet; instead, treat the model's stated probability as a likelihood and reconcile it with the base rate. If the model issues a 70%-confidence rain forecast for 70% of actual rainy days and 30% of actual dry days (a perfectly calibrated model would have a specific likelihood profile - these numbers approximate it):
P(rain | model says 70%) = (0.70 × 0.15) / [(0.70 × 0.15) + (0.30 × 0.85)] ≈ 0.105 / 0.360 ≈ 0.29
A 70%-confidence forecast on a 15%-base-rate day translates to about a 29% real probability. Calibration only holds in aggregate; for any individual case the base rate matters enormously. This is the same machinery as base-rate neglect, just running in reverse.
## Example 5: how does Bayes solve the two-coins puzzle?A bag contains two coins. One is fair (50% heads). One is double-headed (100% heads). You pick a coin at random and flip it once: heads. What is the probability you picked the double-headed coin?
P(double | heads) = (1.0 × 0.5) / [(1.0 × 0.5) + (0.5 × 0.5)] = 0.5 / 0.75 = 0.667
One heads observation lifts the posterior from 50% to 66.7%. Two consecutive heads (re-applying Bayes with 0.667 as the new prior):
P(double | HH) = (1.0 × 0.667) / [(1.0 × 0.667) + (0.5 × 0.333)] = 0.667 / 0.833 = 0.800
Three heads in a row pushes it to about 0.889, and so on. Each observation is consistent with both hypotheses, but it is twice as consistent with one of them - and that ratio compounds.
## What are the most common Bayes-theorem mistakes? - **Ignoring the base rate** - Treating the likelihood as if it were the posterior. The 99% test does not give a 99% diagnosis when the disease affects 0.1% of the population. The prior dominates the calculation whenever the base rate is far from 50%. - **Confusing P(A | B) with P(B | A)** - These are almost never equal. The probability a defendant is guilty given a DNA match is not the same as the probability of a DNA match given guilt. Reversing the direction inflates apparent certainty enormously. - **Forgetting the denominator** - The marginal P(B) is the total probability of the evidence under every hypothesis. Skipping it produces numbers that look right but are not normalised - they will not sum to one across the hypothesis space. - **Using point estimates for likelihoods that have real uncertainty** - If the test sensitivity is 'about 99%' with a wide confidence interval, the posterior also has a wide interval. Treating the likelihood as a single number propagates false precision into the answer. - **Stopping at one update** - Bayes is iterative. New evidence updates the posterior, which becomes the prior for the next round. A single shocking observation rarely settles a question; a sequence of moderate observations usually does. ## How do you do Bayesian updating step by step? 1. **Write down the hypothesis space** - What are the possible answers to the question? For binary cases, A and not-A. For multiple hypotheses, list them explicitly. Bayes only works when the hypotheses are mutually exclusive and collectively exhaustive. 2. **Assign a prior to each hypothesis** - Use the base rate where one exists (population prevalence, historical frequency). When no data is available, an uninformative uniform prior is a defensible default - but flag it as such, because the posterior inherits the prior's uncertainty. 3. **Specify the likelihood of the observation under each hypothesis** - P(evidence | A) and P(evidence | not A). These are usually what the data gives you - sensitivity and specificity for a test, hit rates for a forecast, frequency tables for a classifier. 4. **Compute the marginal probability of the evidence** - Sum the likelihood times the prior across every hypothesis. This is the denominator that normalises the posterior. 5. **Divide to get the posterior** - Numerator: likelihood times prior for the hypothesis you care about. Denominator: the marginal you just computed. Out comes a probability between 0 and 1. 6. **Repeat for the next observation** - Substitute the posterior in as the new prior and re-run with the next piece of evidence. Two or three iterations usually resolve most questions worth asking. ## Where does Bayes theorem actually show up?The formula is mathematically universal, but a handful of domains rely on it constantly:
The formula is the formal version of a habit: change your mind in proportion to the strength of the evidence, weighted by how plausible the claim was to begin with. Most everyday situations do not need the numbers - they need the framing.
Three working rules approximate the calculation:
This is the everyday application covered in Bayesian thinking for everyday decisions - the mental discipline you can use without ever writing the formula down.
## Frequently Asked Questions **Q: What is Bayes theorem in one sentence?** Bayes theorem is the formula for updating the probability of a hypothesis when new evidence arrives: P(hypothesis | evidence) = P(evidence | hypothesis) × P(hypothesis) / P(evidence). **Q: Why is Bayes theorem so important?** Because it is the only mathematically correct way to combine a prior belief with new evidence. Every probabilistic forecast, diagnostic test interpretation, and Bayesian machine-learning model relies on the same equation. **Q: What is the difference between Bayesian and frequentist probability?** Frequentists define probability as the long-run frequency of an event in repeated trials. Bayesians define probability as a degree of belief that can be updated with evidence. Bayes theorem only makes sense under the second interpretation. **Q: What is a prior in Bayes theorem?** The prior is the probability assigned to a hypothesis before observing any evidence. It can come from a base rate, historical data, or a subjective judgment. The posterior depends on it, so a poorly chosen prior produces a poorly calibrated answer. **Q: Why does Bayes theorem give counter-intuitive results for rare diseases?** Because the false-positive pool from the much-larger healthy population dwarfs the true-positive pool from the small infected population. Even a highly accurate test will produce mostly false positives when the disease is rare - the maths is correct; the intuition is wrong. **Q: Can Bayes theorem be applied iteratively?** Yes - and that is one of its strengths. The posterior from one observation becomes the prior for the next. Sequential updating across many pieces of evidence is how Bayesian inference engines actually work in practice. **Q: What is the prosecutor's fallacy?** Confusing the probability of evidence given guilt with the probability of guilt given evidence. They are the two sides of Bayes theorem and are almost never equal. The fallacy systematically overstates the strength of forensic evidence in court. ## Further Reading on This SiteThis page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
**Answer**Daniel Kahneman's Thinking, Fast and Slow argues that the human mind runs two systems: a fast, intuitive, error-prone System 1 and a slower, deliberate, more accurate System 2. Most of our predictable mistakes happen when System 1 answers a question that needed System 2. The book is worth reading for the cognitive-bias catalogue alone, which underpins most of the modern behavioural-economics field.
Where to buy: Thinking, Fast and Slow on Amazon.co.uk →
If you only ever read one book on cognitive bias, this is the one. Daniel Kahneman's 2011 work is the founding text of the entire mental-models genre, and almost every other decision-making book on the market is a footnote to it. After more than a decade in print it remains the densest, most rigorous introduction to how human judgement actually works - and where it predictably fails.
This review covers what the book teaches, where it has aged less well than its admirers admit, and whether to buy it. (Yes - but read it slowly.) ## What is Thinking, Fast and Slow really about?Kahneman's central claim is that the mind runs two systems in parallel. System 1 is fast, automatic, intuitive, and almost entirely unconscious - it's what answers "two plus two" before you finish reading the question. System 2 is slow, effortful, sequential, and lazy - it's what you use to do long division or to evaluate a complex legal argument.
The trouble is that System 2 is lazy in a specific, predictable way: it defers to System 1 unless something forces it to engage. Most of the time you feel like you're thinking, but you're really just rationalising the answer System 1 already supplied. Kahneman spends the next 400 pages cataloguing exactly how System 1 fails - and the catalogue is the most useful map of human irrationality ever assembled in one volume.
**What kind of book is it?** A pop-science synthesis of forty years of academic research by the man who created most of it. Kahneman won the 2002 Nobel Prize in Economics for prospect theory, the work he did with the late Amos Tversky. The book is part autobiography, part literature review, part field manual. ## What the Book Teaches BestProspect theory is the centrepiece, and it's the chapter that earned Kahneman his Nobel. The classical economic view assumes people maximise expected utility - they prefer a 50% chance of £100 to a guaranteed £40. Kahneman and Tversky showed empirically that people overwhelmingly don't. We weight losses roughly twice as heavily as equivalent gains (this is loss aversion in its formal form), and we overweight small probabilities while underweighting large ones. The classical view isn't slightly wrong; it's systematically and predictably wrong, and prospect theory describes the actual shape of the deviation. This single chapter rewrites how you should read any pricing decision, insurance offer, or salary negotiation.
Anchoring is the second big idea, and it's the one most readers underestimate. Once a number is in your head - any number, even one obviously irrelevant - your subsequent judgements drag toward it. Kahneman documents experiments where rolling a weighted die produced answers that systematically tracked the rolled number, even though participants were told the roll was random. The implication is uncomfortable: any conversation that involves a price, a quantity, or a probability has already been half-decided by whichever party named the first number. We cover this further in anchoring bias and how to counter it.
The availability heuristic rounds out the trio. People estimate the probability of events by how easily examples come to mind - which is why fear of plane crashes is rampant and fear of car crashes (vastly more probable per mile) is rare. Vivid stories dominate base rates. This is the single most important habit a working analyst can build: when someone says "X is likely", the immediate next question should be "compared to what reference class?". The availability heuristic post has the working examples.
## Other Ideas Worth Internalising - **The planning fallacy** - Almost all plans are wildly optimistic - we systematically underestimate time, cost, and risk, even when we have data on how badly previous similar plans went. Kahneman's prescription is the outside view: consult the base rate of comparable projects before trusting the inside view of your own. - **Regression to the mean** - Extreme outcomes tend to be followed by less extreme ones, regardless of what intervention you applied. Sports coaches who praise good performances and criticise bad ones therefore wrongly conclude that criticism works and praise doesn't - it's just regression doing the work. - **The narrative fallacy** - Humans construct coherent stories from incomplete data and then trust the stories. Most business case studies, most market post-mortems, and a meaningful slice of clinical research are narratives papered over noise. Once you see it, you can't un-see it. - **WYSIATI (What You See Is All There Is)** - We form confident judgements based on the information immediately available to us, ignoring the absence of evidence we'd need to make a sound call. This is the engine behind most overconfidence, and it's almost impossible to defeat without explicit checklists. - **The peak-end rule** - When remembering an experience, we average its peak intensity with its end - not its total length or sum. A two-week holiday that ended badly is remembered worse than a one-week holiday that ended well, regardless of how much pleasant time the longer one contained. ## What Hasn't Aged WellWhen Thinking, Fast and Slow was published in 2011, social psychology was about to enter a brutal decade of self-examination. Many landmark findings from the 1990s and 2000s failed to replicate when other labs ran the same experiments at proper sample sizes. Kahneman's book leans heavily on some of these findings, particularly in the chapter on priming.
In 2017, Kahneman himself acknowledged this in an open letter on his Facebook page. He wrote that he had "placed too much faith" in social-priming findings and that the relevant chapter in the book "should not be cited as evidence". This is a remarkably honest correction from an author at his level - but it does mean Chapter 4 ("The Associative Machine") should be read sceptically. Specifically, the famous Florida-effect priming study (where participants exposed to elderly-related words walked more slowly out of the lab) has not replicated.
The rest of the book is on much firmer ground. Prospect theory, anchoring, the planning fallacy, and the availability heuristic have all been replicated extensively. The replication crisis hit a specific sub-literature within social psychology, not the heuristics-and-biases programme Kahneman built with Tversky. So treat Chapter 4 as historical interest rather than current science, and trust the rest. The deeper point - that we have hard-wired patterns of misjudgement we can name and predict - is unchanged.
## Is It Readable?This is the most common complaint and it's fair. Thinking, Fast and Slow is 500 pages long, dense with experimental detail, and written by an academic. Kahneman is a good prose stylist but he is not Malcolm Gladwell. Many readers buy it, get through the first 150 pages, and then leave it on the bedside table for two years.
The way to read it is in chunks of one chapter at a time, with a notebook open. Each chapter is essentially self-contained and runs about 10–15 pages. Skip nothing in Parts I and IV (System 1/System 2, and prospect theory) - those are the load-bearing sections. Parts II, III, and V can be sampled rather than read straight through.
If the slow pace defeats you, two options. First, the original 1974 Tversky-Kahneman Science paper "Judgment Under Uncertainty: Heuristics and Biases" is six pages and covers anchoring, availability, and representativeness in one sitting. Second, Annie Duke's Thinking in Bets builds on the same foundations and is half the length and twice the pace - it's a faster door into the same room.
## Where the Book Helps Most The book pays its highest dividend in three domains. Personal finance is one - once you've internalised loss aversion and the disposition effect, you'll catch yourself wanting to sell winners and hold losers, and you'll override the instinct. Hiring is another - the chapter on the illusion of validity is the strongest case ever written for structured interviews over gut-feel hiring. And probability work generally - every bias the book describes shows up in spades the moment you start estimating a number you can't verify. It is less useful for purely interpersonal decisions, creative work, or anything that depends on tacit knowledge rather than explicit judgement. Kahneman is the wrong guide for choosing a partner or deciding whether to start a band. He is the right guide for choosing an investment, evaluating a forecaster, or designing a clinical trial. The boundary is roughly: anywhere a number could be attached, the book applies - and a numerate, structured approach like a decision tree will compound the value further. ## VerdictThinking, Fast and Slow is the canonical text on cognitive bias and the founding work of the modern decision-science literature. It is the best single book in print for understanding why intelligent people make predictable mistakes - and for building the language you need to spot them in yourself.
Three caveats. The priming chapter has aged badly and should be read with the 2017 author's note in mind. The prose is slow and academic; many readers struggle to finish it. And a few chapters in the middle drag - Parts II and V can be skimmed without missing the core argument.
Buy it, read it over six months in chapter-sized doses with a notebook open, then re-read the prospect-theory section every couple of years. Pair it with Thinking in Bets for the working-practitioner perspective and with Superforecasting for the calibration drills. Together, those three books are the standard reading list for anyone serious about decision quality.
**The cheaper-than-the-book habit** After each chapter, write three sentences in a decision journal about a real decision you made that week and which bias from the chapter most applied. By the time you finish the book you'll have 38 cases - that's the practical residue worth the £14. ## Frequently Asked Questions **Q: Is Thinking, Fast and Slow worth reading in 2026?** Yes. The replication crisis affected one chapter (priming, Chapter 4), not the book's main contributions. Prospect theory, anchoring, the planning fallacy, and the availability heuristic have all replicated extensively. The book remains the canonical reference for cognitive bias. **Q: What should I read first - Kahneman, Duke, or Tetlock?** If you want depth and you'll commit six months, start with Kahneman. If you want practical application in three weeks, start with Annie Duke's Thinking in Bets. If you want to actually get better at numerical forecasting, start with Philip Tetlock's Superforecasting. All three reinforce each other; the order matters less than reading them all eventually. **Q: Which chapters can I skip?** Chapter 4 ("The Associative Machine") on priming has aged worst - read it for historical context only. Chapter 35 ("Two Selves") and Chapter 38 ("Thinking About Life") are good but optional unless you're particularly interested in well-being research. The load-bearing chapters are 1–3 (System 1/2), 10–18 (heuristics), and 25–32 (prospect theory and framing). **Q: Is the audiobook any good?** It's narrated competently but not memorably. The book is reference-heavy - graphs, tables, experimental descriptions - and the audio version makes those harder to absorb. Most readers do better with the paperback and a notebook. The audiobook works as a re-listen after you've read it once, but not as a first encounter. **Q: How does this book compare to Predictably Irrational by Dan Ariely?** Ariely's book is shorter, punchier, and easier to finish, but lighter on theory. It draws on similar research but presents it as standalone behavioural-economics experiments rather than a unified two-systems framework. Read Ariely first if you want to be hooked; read Kahneman next for the deeper structure. **Q: Is the two-systems framework still considered accurate?**It's a useful metaphor more than a literal neuroscientific model. Most cognitive scientists today treat System 1 and System 2 as a teaching shorthand for two distinct types of processing rather than two physically separate systems. The categorisation still predicts behaviour reliably, which is what matters for practical decision-making, even if the underlying neural picture is more distributed.
**Q: What's the best single takeaway from the book?** Slow down on decisions that look obvious. The biases Kahneman catalogues mostly attack situations where System 1 produces a confident answer and System 2 never wakes up to check. Building the habit of pausing on apparently-obvious calls - especially under time pressure or emotional load - captures most of the practical value of the entire book. --- # Decision Trees: A Visual Framework for Complex Choices Source: https://expectedvalue.co.uk/blog/decision-trees-visual-framework/ # Decision Trees: A Visual Framework for Complex Choices Decision trees turn tangled choices into a diagram you can actually solve. Build one in five steps, work through three examples, and avoid common traps. ## What a Decision Tree Actually Is **Answer**A decision tree maps out a complex choice as a sequence of branching outcomes, each with its own probability and payoff. Draw your decision at the root, branch out to each option, then branch again at each chance node, and finally score the expected value of each path. Best for decisions with three or more steps, multiple uncertain inputs, or where teams need to align on assumptions.
A decision tree is a diagram of a choice that branches forward through time. You sketch every decision you control, every uncertain event you don't, and the outcomes at the end. Then you collapse the tree back into a single number - an expected value for each option - by rolling the maths backwards from right to left.
The technique is older than most readers assume. It was formalised by Howard Raiffa at Harvard in the 1960s and has been standard practice in medical decision analysis, oil-and-gas exploration, and operations research ever since. The reason it never went away is simple: it forces you to write down the things you'd otherwise leave vague - the probability of each scenario, the value of each outcome, the order in which information arrives. Most bad decisions are bad because those numbers stayed in someone's head.
There are exactly three node types. Master these and you can model almost any decision:
A pros-and-cons list treats every consideration as equal in weight and certain to occur. That's why it so often produces the wrong answer: a single rare-but-catastrophic con can dominate ten mild pros, but the list has no way to show it.
Decision trees fix three specific failures of unstructured deliberation:
They force you to multiply, not just add. A 5% chance of a £200,000 loss is not the same kind of thing as a 70% chance of a £10,000 gain - but on a pros-and-cons list they're indistinguishable bullet points. The tree makes the multiplication visible.
They force you to be explicit about probabilities. Once you have to put a number on each branch of a chance node, you can no longer hide behind vague language like "there's a real risk that...". You have to commit to a probability. If you can't - if your honest answer is "somewhere between 10% and 60%" - that's useful information too, and you can run the tree at both endpoints.
They make sequential structure visible. Most real decisions aren't one-shot; they're a chain. You choose, an outcome happens, you choose again. A tree captures that ordering, and the value of information shows up automatically - branches that let you wait for a signal before committing usually beat branches that lock you in early.
A related framework that complements decision trees is the pre-mortem - imagining the decision has already failed and reasoning backwards from the failure mode. Pre-mortems generate the chance-node branches; the tree quantifies them.
## Building a Decision Tree in Five Steps Every decision tree follows the same construction recipe. The mechanics are simple; the discipline of doing each step honestly is where the work lives. 1. **Frame the decision** - Write the question in one sentence: "Should we expand into the German market in 2026?" Be specific about the time horizon and the alternative - "should we do X" is incomplete without "...instead of what?". 2. **List the options** - Draw a decision node and a branch for each option you'll seriously consider. Three to five branches is the sweet spot - fewer and you're probably anchoring on a default; more and you're not really deciding, you're brainstorming. 3. **Identify the key uncertainties** - For each option, what's the one or two uncertain things that determine whether it turns out well? Don't list every possible factor - pick the ones that change the answer if they flip. Add a chance node for each, with two to four branches. 4. **Assign probabilities and payoffs** - Put a probability on each branch leaving a chance node (they must sum to 1.0) and a monetary value at every terminal node. Start with base rates if you have them - see our guide to base rate neglect for why this matters - then adjust for your specific circumstances. 5. **Roll back to find the expected value** - Start at the terminal nodes on the right. At each chance node, calculate the EV - sum of (probability × value) across its branches - and write that number on the chance node itself. At each decision node, pick the branch with the highest EV and write that number on the decision node. Keep working leftward until you reach the original decision. The branch with the highest rolled-back EV is the EV-maximising choice. ## Worked Example: The CounterofferImagine Maya, a senior product manager on £85,000. She's been offered a role at a competitor on £105,000. Her current employer has hinted that they'll counter if she resigns. She has three options:
Let's build the tree.
- **Branch 1: Take the new offer**: Certain: £105k × 3 years = £315,000 - **Branch 2a: Resign → strong counter (£100k)**: P = 0.40 → £100k × 3 = £300,000 - **Branch 2b: Resign → weak counter (£90k)**: P = 0.35 → £90k × 3 = £270,000 - **Branch 2c: Resign → no counter, take new offer anyway**: P = 0.25 → £315,000 (but reputational cost) - **Branch 2 rolled-back EV**: 0.40 × 300k + 0.35 × 270k + 0.25 × 315k = £293,250 - **Branch 3: Stay put**: Certain: £85k × 3 + standard 4% rises = £265,302Rolling back: Branch 1 has EV £315,000. Branch 2 has EV £293,250. Branch 3 has EV £265,302. The EV-maximising choice is Branch 1 - take the offer and leave.
Notice what the tree reveals that a list wouldn't. Branch 2 looks superficially attractive - "I'll get a counter" - but once you multiply, the chance of a weak counter drags the expected value below Branch 1. The counter-game only beats taking the offer outright if you believe the probability of a strong counter is above ~70%, which most people overestimate.
It also exposes a hidden cost the framing usually hides: in Branch 2c, you've burned political capital by resigning, even though you end up at the new employer anyway. The £315,000 figure there is the same as Branch 1's, but the path was worse. Some practitioners assign a soft penalty here - say -£10,000 in reputational damage - which pushes Branch 2's EV down further.
This kind of structured reasoning is the same engine behind Bayesian thinking in everyday decisions - making the prior probabilities explicit instead of leaving them as intuitions you can't argue with.
## Worked Example: Surgery vs Watchful Waiting Medical decision analysis is where the technique earned its keep. The framework lets a doctor and patient quantify the trade-off between an aggressive intervention and a conservative one, given imperfect diagnostic information. Consider a hypothetical: a 60-year-old patient has a borderline scan result for a slow-growing tumour. The clinician estimates a 30% probability it's malignant. The options are immediate surgery, or watchful waiting with a follow-up scan in six months. We'll measure outcomes in quality-adjusted life years (QALYs) - one year of full health is one QALY, one year at 50% quality is 0.5 QALYs. Numbers below are illustrative rather than clinical: - **Surgery: P(success) = 0.92**: 20 QALYs gained - **Surgery: P(major complication) = 0.05**: 12 QALYs gained - **Surgery: P(mortality) = 0.03**: 0 QALYs (with surgical decision regret weighting) - **Surgery rolled-back EV**: 0.92 × 20 + 0.05 × 12 + 0.03 × 0 = 19.0 QALYs - **Watch: P(benign as suspected) = 0.70**: 20 QALYs (no intervention needed) - **Watch: P(malignant, caught next scan) = 0.20**: 0.92 × 18 + 0.05 × 10 + 0.03 × 0 = 17.06 QALYs (delayed surgery) - **Watch: P(malignant, late discovery) = 0.10**: 8 QALYs (worse prognosis after delay) - **Watch rolled-back EV**: 0.70 × 20 + 0.20 × 17.06 + 0.10 × 8 = 18.21 QALYsSurgery wins on raw EV - 19.0 vs 18.21 QALYs. But the gap is small (less than 1 QALY) and depends entirely on the 30% malignancy estimate. If the true probability is 20%, watchful waiting wins. If it's 40%, surgery wins decisively. The tree doesn't tell you the answer - it tells you which input you should care about most.
This is the single most useful output of decision-tree analysis in real life. The point is rarely to compute a single number and act on it. The point is to identify the sensitivity - which input would have to change, and by how much, before your conclusion flips. In medical decision analysis, this is called a one-way sensitivity analysis. In business, it's called a tornado chart. Either way, you're asking: what would have to be true for me to change my mind?
**Run the tree twice** Run every tree at your best-guess probabilities, then re-run it at the boundaries of what you think is plausible (say, ±15 percentage points on the key probability). If the EV ranking doesn't change, you can act with confidence. If it does, you've identified what to investigate before deciding. ## Worked Example: Entering a New MarketBusiness expansion decisions illustrate the most powerful feature of decision trees - modelling decisions in sequence, where information arrives over time and lets you avoid commitment.
A UK fintech is considering launching in Germany. Two options: launch immediately at £4M cost, or run a six-month pilot in Berlin at £600k cost first. The full launch's payoff depends on the local regulatory environment, which is mid-consultation.
The pilot effectively buys information. The tree captures this by adding a second decision node after the pilot's chance node - once the regulatory signal arrives, you can choose again.
- **Path A: Full launch now, favourable regs (P=0.5)**: Net £18M over 5 years - **Path A: Full launch now, unfavourable regs (P=0.5)**: Net -£2M over 5 years - **Path A rolled-back EV**: 0.5 × 18M + 0.5 × (-2M) = £8.0M - **Path B: Pilot → favourable signal → launch (P=0.45)**: £18M - £600k pilot = £17.4M - **Path B: Pilot → mixed signal → launch anyway (P=0.20)**: EV £6M - £600k = £5.4M - **Path B: Pilot → unfavourable signal → walk away (P=0.35)**: -£600k pilot cost only - **Path B rolled-back EV**: 0.45 × 17.4M + 0.20 × 5.4M + 0.35 × (-0.6M) = £8.7M The pilot path wins by £700k in expected value, even though it spends £600k on the pilot. The reason: it lets you walk away from a £2M loss 35% of the time. That optionality is worth more than the pilot costs. This is the textbook "value of information" calculation. Whenever you can spend a known amount to learn something material before committing a larger amount, the tree will usually favour learning first - provided the information actually reduces uncertainty (a pilot that doesn't tell you anything useful is just a tax). This logic underpins much of second-order thinking and is the discipline that separates good operators from bold ones. ## When Decision Trees MisleadThe technique is powerful, which means the mistakes are correspondingly costly. Five common failure modes:
The tree gets too big to think about. Beyond roughly fifteen terminal nodes, the diagram stops aiding intuition. Either collapse low-impact branches into a single "other" bucket or split the problem into a sequence of smaller trees. If you can't draw it on a single sheet of A4, you've overspecified the problem.
The probabilities are guesses dressed as data. A decision tree run with made-up probabilities is just a list of opinions wearing a maths costume. Always start with base rates from real reference classes - historical success rates of comparable launches, published clinical outcomes, observed conversion rates from prior experiments. Adjust from there, but commit to the base rate as your prior.
Base rates get ignored entirely. The classic decision-tree failure is letting an inside view dominate. Maya in the counteroffer example might feel like her counter would be strong because her relationship with the boss is good, but the base rate for "counters that match a 23% external uplift" is well under 40%. Anchor on the outside view first; adjust modestly for the specifics.
Real options are missed. Many decisions look like "do it / don't do it" but actually have a third branch: "do a smaller version, then decide". Pilots, prototypes, MVPs, two-week trials - all are real options that branches on the tree can capture. Skip them and you'll systematically over-commit. This is one of the highest-leverage corrections you can make to your own thinking; thinking probabilistically is essentially the habit of always asking what the next-cheapest piece of information would be.
Terminal values are computed too narrowly. The terminal node at the end of every path needs to include everything - direct payoff, opportunity cost, reputational impact, time spent, tax, and second-order effects. A spreadsheet that only counts the headline number will systematically over-value high-risk, high-headline branches. For long-horizon decisions, also account for the timing of cash flows: a £1 in five years is not the same as £1 today, and the conditional probability of each branch can shift if the underlying conditions change between now and then.
## Tools for Building Decision Trees - **Pen and paper** - For everything under about ten terminal nodes, this is the right tool. The discipline of sketching the structure by hand is half the value of the exercise. - **draw.io / Lucidchart / Whimsical** - Free or freemium diagramming tools that handle trees up to roughly thirty nodes cleanly. Best for shareable diagrams you want a team to review. Whimsical is the fastest of the three for getting a clean diagram in under five minutes. - **Spreadsheets (Excel, Google Sheets)** - Once you need sensitivity analysis or scenario tables, a spreadsheet beats a diagram. Put each path on a row, each node's probability and value in named cells, and the rolled-back EV in a totals column. - **TreeAge Pro** - The professional choice for medical decision analysts and operations researchers. Handles trees with thousands of nodes, runs Monte Carlo sensitivity automatically, and exports clinical-quality reports. Pricey; only worth it for full-time decision analysts. - **Python / R libraries** - For analysts comfortable with code: scikit-learn for machine-learning trees, dtreeviz for visualisation, and PyMC or Stan for trees with Bayesian-updated probabilities. Overkill for personal decisions; essential for production decision systems. ## Further ReadingTwo books are worth owning. Decisive by Chip and Dan Heath is the most readable introduction to structured decision-making - it doesn't dwell on the maths but it captures the spirit of the discipline better than any pure decision-analysis text. Thinking in Bets by Annie Duke is the practitioner's companion, written by a former professional poker player who teaches Fortune 500 executives. We've covered the latter in detail in our Thinking in Bets summary - start there if you want the key ideas before buying the book.
For calibrating the probabilities you'll feed into your trees, Superforecasting by Philip Tetlock is the canonical text on training yourself to produce numerical estimates that actually correspond to reality. If you've never measured your own calibration, you almost certainly believe you're better than you are - see our guide to probability calibration training for a way to fix that with twenty minutes a week.
And if you're building decision trees for financial choices specifically, you'll get more from them once you've internalised the Kelly criterion for bet sizing - the maths that tells you not just whether a bet is positive-EV but how much of your bankroll to commit.
## Frequently Asked Questions **Q: What's the difference between a decision tree and a flowchart?** A flowchart maps a process - what step happens after each previous step. A decision tree maps a choice - what outcome arises from each combination of decisions and uncertain events, with probabilities and values attached. Flowcharts are for describing systems that already exist; decision trees are for deciding what to do before the system runs. **Q: How many branches should a chance node have?** Two to four, almost always. Five branches is usually a sign you're modelling at too fine a grain - collapse "high / medium-high / medium / medium-low / low" into "good / neutral / bad". The exception is when you have real frequency data and the granular branches change the EV materially. If the answer is the same with three branches as with seven, use three. **Q: Should I include tiny-probability branches?** Only if the magnitude is large enough to matter. A 0.1% probability of a £50,000 loss has an EV impact of £50 - usually negligible. A 0.1% probability of bankruptcy has the same EV impact in raw maths but matters enormously because of ruin avoidance. Apply the same rule as with insurance: tiny probabilities of catastrophic outcomes deserve their own branch even when their EV impact is small. **Q: What if I don't know the probabilities?** Use the outside view as a starting point - historical base rates from comparable situations. Then run the tree at the lower and upper ends of your uncertainty range. If the EV ranking doesn't flip across that range, you can act without further analysis. If it does flip, you've identified the single most valuable thing to investigate next. **Q: Do decision trees work for one-off decisions?** Yes, with two adjustments. First, weight the variance of each branch, not just the expected value - for a one-shot decision you only get one outcome, so a high-variance branch can blow up even if its EV looks attractive. Second, consider ruin and irreversibility - if a bad outcome would prevent you from playing again, treat it as a disqualifying constraint rather than le …[truncated for length] --- # Bayesian Thinking for Everyday Decisions Source: https://expectedvalue.co.uk/blog/bayesian-thinking-everyday-decisions/ # Bayesian Thinking for Everyday Decisions Bayesian thinking - the art of changing your mind rationally. How to update beliefs with evidence, with examples from interviews, medicine, and investing. ## Do you already think like a Bayesian? **Answer**Bayesian thinking is the discipline of updating your beliefs proportionally to the strength of new evidence. Start with a prior probability, encounter evidence, multiply through Bayes' theorem to get a posterior. The practical lesson: strong evidence should shift you a lot, weak evidence should shift you a little, and pre-existing strong priors should resist weak counter-evidence.
Every time you change your mind, you're doing Bayesian reasoning - formally, you're computing a conditional probability of your hypothesis given the new evidence. You believed something, new evidence arrived, and your belief shifted. The problem isn't that you don't update - it's that you update badly.
Bayesian thinking, named after the 18th-century Presbyterian minister Thomas Bayes, is simply a formal framework for doing what your brain already attempts: combining what you previously believed (your prior) with new evidence to arrive at an updated belief (your posterior). The formula itself - Bayes' theorem - is elegant but intimidating:
P(H|E) = P(E|H) × P(H) / P(E)
Don't worry about the notation. The intuition is what matters, and it's beautifully simple: the strength of your updated belief depends on what you believed before AND how surprising the new evidence is. If you already thought something was likely and the evidence confirms it, you should believe it more strongly. If you thought something was unlikely but the evidence strongly suggests otherwise, you should update - but perhaps not as much as the evidence alone would suggest.
This single idea - that prior beliefs and new evidence must be weighed together - is the most powerful thinking tool most people never learn to use deliberately. It's also the foundation that makes expected value calculations actually work: the better your probability estimates, the better your EV calculations, and the better your decisions.
## Where do your priors actually come from?Your prior probability - what you believe before seeing new evidence - is the starting point for every Bayesian update. Priors come from several sources:
Base rates are the most important and most neglected. If 2% of job applicants get an offer at a particular company, your prior probability of getting an offer should start at roughly 2% before you factor in anything specific about your interview. If 1 in 10,000 people your age have a particular disease, that's your prior before any test results come in.
Your own track record shapes priors too. If you've started three businesses and two failed, your prior for the next one succeeding is informed by that track record. This is legitimate - as long as your sample isn't too small or too biased.
Expert consensus provides useful priors when you lack personal data. What do doctors, scientists, or seasoned professionals in the relevant field believe? Their aggregate view often encodes decades of evidence you haven't personally reviewed.
The critical mistake is having strong priors based on nothing. Gut feelings, cultural assumptions, things you heard once at a dinner party - these masquerade as informed priors but are often just noise. A Bayesian thinker asks: why do I believe this, and how confident should I actually be?
The strength of your prior matters enormously. A weak prior (say, 50/50 - you genuinely don't know) will shift dramatically with even modest evidence. A strong prior (say, 95% confident) requires substantial evidence to move significantly. This is actually correct behaviour - if you have very good reasons to believe something, a single piece of contradictory evidence shouldn't overturn it. But it should nudge it.
## How do you update on evidence, with an example?Let's walk through Bayesian updating with a concrete example that uses natural frequencies - a method that makes the maths intuitive without requiring any formulas.
Scenario: You've applied for a job. The company hires roughly 5% of applicants who reach the final interview stage. You think your interview went well.
Step by step:
Your prior: 5% chance of getting an offer (the base rate for final-stage candidates).
Evidence 1: The interviewer spent 45 minutes with you instead of the scheduled 30, asked detailed questions about your start date, and introduced you to team members. Let's say this happens for about 60% of candidates who eventually get offers, but only about 15% of those who don't.
Using natural frequencies: imagine 1,000 candidates at final stage. 50 get offers, 950 don't.
The positive signals moved your estimate from 5% to 17% - a meaningful update, but nowhere near certainty. This is a key Bayesian insight: even strong-seeming evidence doesn't override a low base rate as much as your intuition suggests.
Evidence 2: Two days later, the recruiter emails asking for your references. This happens for 80% of candidates who get offers and 10% of those who don't.
Updating from our new base of 17.3%:
Now we're talking. Two pieces of evidence, each moderately diagnostic, have taken you from 5% to 63%. But notice you're still not at 90% - because the base rate was genuinely low, and each piece of evidence, while positive, wasn't conclusive.
Evidence 3: A week of silence. No call, no email. Among candidates who get offers, only 20% experience a full week of silence before hearing. Among those who don't, 70% do.
Updating from 63.2%:
The silence was negative evidence that pulled your estimate back down. A Bayesian thinker doesn't ignore evidence that contradicts their hopes.
## Why do people get medical diagnosis probability so wrong?You wake up with a headache. Not just any headache - a persistent, throbbing headache that's lasted three days. You Google your symptoms. Brain tumour appears in the results. Your anxiety spikes.
Let's apply Bayesian thinking.
Prior: Brain tumours affect roughly 10 in 100,000 people per year in the UK, or about 0.01%. This is your base rate.
Evidence: You have a persistent headache. Persistent headaches occur in roughly 60% of brain tumour cases, but they also occur in about 4% of the healthy population in any given month (from tension, dehydration, stress, poor sleep, and dozens of other mundane causes).
The natural frequency calculation:
Your probability went from 0.01% to 0.15% - a fifteen-fold increase, which sounds dramatic. But 0.15% is still vanishingly small. There's a 99.85% chance your headache is caused by something mundane.
This is why doctors don't order MRI scans for every headache. The base rate is so low that even a symptom strongly associated with the disease barely moves the needle. You'd need multiple converging symptoms - headache plus vision changes plus unexplained nausea plus progressive worsening - before the posterior probability climbed to a level warranting expensive investigation.
The lesson: When you're worried about a rare outcome, always start with the base rate. Vivid, frightening possibilities feel more probable than they are - a cognitive trap we explored in depth in our piece on why your brain is bad at risk.
## How do you update an investment thesis the Bayesian way?You hold shares in a tech company. Your thesis: the company will grow revenue at 20%+ annually for the next three years. You'd put your confidence at around 70% when you bought the shares.
Then Q1 earnings come in: revenue growth was 12%, missing analyst expectations of 18%.
The non-Bayesian investor does one of two things:
Both responses are wrong. The Bayesian investor asks: how much should one disappointing quarter update my belief?
Consider:
A single miss is only mildly diagnostic - it's roughly twice as likely if the thesis is wrong versus right. Running the numbers from a 70% prior, one quarterly miss takes you to about 54%. Still more likely than not, but your confidence has justifiably weakened.
Now if Q2 also misses, and the company revises guidance downward? Each additional piece of negative evidence compounds. Two misses plus a guidance cut might take you from 54% to 25%. At some point, the evidence overwhelms your prior, and the rational move is to exit.
Key insight: Bayesian updating in investing prevents both premature panic and stubborn holding. It gives you a principled framework for deciding when enough evidence has accumulated to change course. This connects directly to optimal bet sizing - as your confidence in a thesis drops, the Kelly Criterion says your position size should shrink proportionally.
## How does weak evidence stack up over time?Bayesian thinking is particularly useful for reading ambiguous social situations. Consider this scenario: you're wondering whether your partner is unhappy in the relationship.
Your prior: Based on general relationship satisfaction rates and your history together, maybe you'd start at 15% - possible but unlikely.
Now you notice several small things over a few weeks:
Individually, each of these signals is almost meaningless. Any one of them has a mundane explanation that's far more likely than "partner is unhappy." If you updated on any single signal in isolation, you'd barely move from 15%.
But Bayesian updating is cumulative. Four independent weak signals, each mildly more consistent with unhappiness than happiness, compound. If each signal is 1.5 times more likely to occur if your partner is unhappy (a very modest diagnostic value), four such signals multiply: 1.5^4 = 5.06. Your 15% prior updates to roughly 47%.
That's a meaningful shift - from "probably fine" to "genuinely uncertain" - and it happened through the accumulation of individually trivial evidence. This is one of Bayesian thinking's most powerful features: it tells you that many weak signals can be as informative as one strong signal.
Compare with one strong signal: Your partner directly says "I think we need to talk about our relationship." That single statement might be 10 times more likely if they're unhappy - taking your 15% prior to about 64% in one jump. Dramatic, but not qualitatively different from what four weak signals achieved.
The practical takeaway: Pay attention to patterns, not just dramatic events. Bayesian reasoning validates the intuition that "something feels off" - even when you can't point to any single piece of evidence that's conclusive.
## What are the most common Bayesian thinking mistakes?Understanding Bayes' theorem doesn't automatically make you good at applying it. Here are the most common failure modes:
1. Anchoring too hard on priors
Some people, having learned that base rates matter, become so anchored on their prior beliefs that virtually no evidence can shift them. This is the mirror image of base rate neglect - and it's equally wrong. If your prior is 5% but you see five independent pieces of strong confirmatory evidence, stubbornly staying near 5% isn't rational caution; it's dogmatism. Strong evidence should move you substantially, even from a low prior.
2. Overcorrecting on vivid evidence
A single dramatic anecdote - a friend who smoked until 95 and never got cancer - can overwhelm careful probabilistic reasoning if you're not vigilant. Vivid, emotionally charged evidence feels more diagnostic than it actually is. One data point is almost never enough to significantly update a well-established base rate. Our tendency to overweight vivid examples is one of the cognitive biases that systematically distort our risk perception.
3. Treating dependent evidence as independent
Bayesian updating works cleanly when each piece of evidence is independent. But in practice, evidence often clusters. Three news articles reporting a company is in trouble might all be sourcing the same rumour. Two symptoms might share a common cause (stress) rather than independently pointing to a serious diagnosis. When evidence is correlated, each additional piece provides less new information than you'd think.
4. Ignoring the base rate entirely
The classic error. Someone tells you a business idea is "genius" - but 90% of startups fail. A test comes back positive - but the disease affects 1 in 100,000 (the false positive paradox in action). The specific evidence feels compelling because it's right in front of you; the base rate is abstract and easy to overlook. Always ask: how common is this outcome in general?
5. Failing to update at all
Perhaps the most insidious mistake. You form a belief and then simply... stop updating. New evidence arrives - a failed prediction, a contradicted assumption, a changed circumstance - and you either don't notice or find ways to explain it away. The Bayesian framework only works if you actually apply it when inconvenient evidence appears.
## How do you actually do Bayesian thinking day to day?You don't need to run calculations every time you make a decision. The goal is to internalise the logic of Bayesian updating so it becomes a thinking habit.
Start with "What's the base rate?"
Before evaluating any specific evidence, ask yourself: how common is this outcome in general? This single question eliminates the most frequent error in probabilistic reasoning. Feeling certain you'll get that promotion? What percentage of people at your level actually get promoted each year? That's your starting point.
Ask "How diagnostic is this evidence?"
When something happens that seems relevant, ask: would this outcome be equally likely whether my hypothesis is true or false? If yes, it's not diagnostic - it shouldn't change your belief at all. An interviewer being friendly might happen 80% of the time regardless of whether you're getting the job. That's nearly useless as evidence. An interviewer discussing specific start dates? That's far more diagnostic.
Quantify your uncertainty
Force yourself to put numbers on beliefs, even rough ones. "I'm about 60% sure this project will finish on time" is far more useful than "I think it'll probably be fine." Numbers let you track how your beliefs shift over time and check whether you're calibrated. This habit of quantifying uncertainty is what makes expected value thinking practical rather than theoretical.
Keep a decision journal
Write down important beliefs with probability estimates. When new evidence arrives, record your update and your reasoning. Over time, you'll develop an intuition for how much different types of evidence should shift your beliefs - and you'll catch yourself when bias is creeping in.
Embrace uncertainty as information
A Bayesian thinker is comfortable saying "I don't know" or "I'm at 50/50 on this." Uncertainty isn't a failure of analysis; it's an honest assessment. The goal isn't to be certain - it's to be calibrated: when you say 70%, things should happen about 70% of the time.
Update in both directions
The true test of Bayesian thinking isn't whether you update when evidence supports your view - everyone does that. It's whether you update when evidence contradicts it. Make it a habit to ask: what evidence would change my mind? If you can't answer that question, you're not reasoning; you're rationalising.
Bayesian updating tells you what to believe given evidence. Pair it with second-order thinking - which forces you to forecast how the world responds to your actions - and you have a complete loop: form a belief, act on it, update on the system's reaction, act again.
- **Prior**: What you believed before seeing new evidence (based on base rates, experience, expert consensus) - **Likelihood**: How probable the evidence is if your hypothesis is true vs. false - **Posterior**: Your updated belief after combining prior and evidence - **Strong prior + weak evidence**: Belief barely shifts - and that's correct - **Weak prior + strong evidence**: Belief shifts substantially toward the evidence - **Multiple wea …[truncated for length] --- # The Gambler's Fallacy: Why You're Wrong About 'Due' Outcomes Source: https://expectedvalue.co.uk/blog/gamblers-fallacy/ # The Gambler's Fallacy: Why You're Wrong About 'Due' Outcomes The Gambler's Fallacy: why we wrongly believe random outcomes are 'due'. The cognitive trap, where it costs you, and how to think clearly under randomness. **Answer** The gambler's fallacy is believing that past random outcomes affect the probability of future ones in an independent system. After ten red spins on roulette, black is not 'due' - the next spin is still ~48% red. The error confuses independent and dependent events; truly independent trials have no memory.On 18 August 1913, in the Casino de Monte-Carlo, a roulette wheel landed on black. Then black again. And again. By the time the ball had landed on black for the fifteenth consecutive time, a crowd had gathered around the table.
They were betting on red. Frantically. Surely, they reasoned, red was due.
The ball landed on black. Then black. Then black again. By the twentieth black in a row, players were emptying their wallets onto the red. The streak finally ended at twenty-six consecutive blacks - a sequence so improbable (roughly 1 in 67 million) that it has its own Wikipedia entry. The casino made an absolute fortune that night, primarily from people convinced that the wheel owed them red.
This is the gambler's fallacy: the deeply intuitive but completely wrong belief that past random outcomes change the probability of future ones. It's one of the most expensive mistakes in cognitive psychology - and once you learn to spot it, you'll see it everywhere from casino floors to stock markets to your own decision-making.
## What Is the Gambler's Fallacy?The gambler's fallacy is the belief that if a particular random event has occurred more or less frequently than expected in the past, it is more or less likely to occur in the future - when in fact the events are independent and the probability hasn't changed at all.
The key word is independent. An event is independent when its outcome doesn't affect the next outcome. A roulette wheel has no memory. A fair coin has no memory. A fair die has no memory. They don't keep count of past spins or flips and adjust their behaviour to balance things out.
A few canonical examples:
If the math is so simple, why does the gambler's fallacy feel so obviously true? Several deep features of human cognition are conspiring against us.
The Law of Large Numbers gets misapplied. The mathematical Law of Large Numbers states that as the number of trials grows very large, the observed frequency converges toward the true probability. Flip a coin a million times, you'll get very close to 50% heads. This is true. What people implicitly assume - incorrectly - is that this convergence happens through short-term correction. They believe the universe is actively rebalancing the books, that a streak of heads must be 'corrected' by tails. It isn't. Convergence happens through dilution: each new flip is so small a contribution to the cumulative percentage that streaks get drowned out by sheer volume, not cancelled out.
Pattern recognition is over-tuned. Our brains evolved to detect patterns in noisy environments - that rustling in the grass might be a predator, that arrangement of clouds might mean rain. False positives (seeing a pattern where none exists) were cheap. False negatives (missing a real pattern) were expensive. We are descendants of the over-pattern-matchers. This serves us well in many contexts but is catastrophic in genuinely random environments, where there are no patterns to find but our brain insists on finding them anyway.
Representativeness heuristic. Daniel Kahneman and Amos Tversky famously showed that people judge probability by how 'representative' a sequence looks. People asked to write down a 'random' sequence of coin flips will produce sequences with far too many alternations (HTHTH...) because that looks random. A real random sequence frequently produces clusters of three, four, five identical outcomes - but those don't feel random, so we expect them to correct themselves. They don't.
Local representativeness. A particularly insidious version: we expect even small samples to look like the underlying probability. If a coin is 50/50, we expect any run of 6 flips to contain about 3 heads. So when we see HHHHHH, our brain screams 'this is wrong, it must correct.' But small samples genuinely vary - and the gambler's fallacy is what happens when we treat that variation as a debt that must be repaid.
## How Casinos Exploit the Gambler's FallacyCasinos understand this bias intimately. Walk past any roulette table in Las Vegas, Macau, or Monte Carlo and you'll see an electronic display board listing the last 10-20 results. Why? Officially: 'so players can see the action.' Actually: because players who see five blacks in a row are statistically far more likely to bet heavily on red - and that bet is no better than any other bet. The display is a profit-maximising tool dressed as transparency.
The same logic explains why casinos display roulette history but not slot machine history (slot machines are not technically independent in the same way, but their outcomes are still essentially random per spin). Anywhere you see a 'last results' board, ask yourself: who benefits from showing me this?
A real number to anchor on. A European roulette wheel gives the casino a 2.7% house edge on every spin. Whether you bet on red, black, a single number, a column, or any combination, the expected value is the same: you lose 2.7% of your stake on average. The previous outcomes change nothing. A streak of five reds doesn't make the next spin a better bet. It's still -2.7% expected value, every spin, forever.
This is also why 'progressive betting systems' like the Martingale (double your bet after every loss) eventually destroy bankrolls. The Martingale assumes that a long losing streak must be corrected - that a win is mathematically due. It isn't. Streaks of any length are possible, and even a modest losing streak can hit table limits or empty your account before the 'correction' arrives. Many gamblers have learned this the hard way.
## The Gambler's Fallacy in Sports BettingSports betting is genuinely different from roulette in one important way: the outcomes are not perfectly independent. Skill, fitness, momentum, weather, and matchups all contribute to results. A team really can be on a roll. A striker really can be in form.
But this is also where it gets dangerous, because the gambler's fallacy mutates into something subtler. Bettors switch between two contradictory beliefs - sometimes within a single conversation:
Which one is right? Sometimes neither. Often, the streak is being driven by factors the bettor isn't tracking - favourable fixtures, an injury return, an opposition slump - and the streak ends when those factors change, not because of any cosmic balancing act.
The practical lesson: in sports markets, do not bet on or against streaks based on the streak alone. The bookmakers have already priced in any genuine momentum effect. If you think a team is overrated because they're 'due' a loss, your only edge is if your model of why they'll lose is more accurate than the market's. The streak itself isn't the edge.
## The Hot-Hand Fallacy: A Twist in the TaleThe hot-hand fallacy is the gambler's fallacy's mirror image: the belief that a string of successes makes future success more likely. A basketball player who's hit four shots in a row is 'hot' - pass them the ball.
For decades, the academic consensus was that the hot hand was a pure illusion - pattern-matching on randomness. A famous 1985 study by Gilovich, Vallone, and Tversky analysed Philadelphia 76ers shooting data and found no evidence of streakiness beyond random chance.
But here's the twist: in 2015, economists Joshua Miller and Adam Sanjurjo found a subtle statistical bias in how those original studies counted streaks. When you correct for it, the hot hand actually exists in basketball - players who've made several shots in a row do have a slightly elevated probability of making the next one. The effect is small (a few percentage points), but it's real.
The correction is technical (it's called Miller-Sanjurjo bias and applies to any analysis of finite sequences), but the broader lesson is more interesting: real-world processes that involve human skill can have momentum, while purely random processes (roulette, lottery, dice) cannot. The mistake is applying the wrong model to the wrong situation.
Practical rule: before reasoning about streaks, ask whether the underlying process is genuinely independent (coins, dice, slot machines) or whether it involves skill, fatigue, or context (sports, business performance, market timing). The cognitive errors look identical, but the corrections are different.
## The Fallacy Outside the CasinoThe gambler's fallacy doesn't stay in casinos. It contaminates decisions far away from any table.
Investing. Investors regularly believe a stock that's fallen for several days is 'due' a bounce, or that a fund that's outperformed for years is 'due' a drop. Sometimes mean reversion is real (valuation does pull prices back over long horizons), but often the streak is driven by underlying changes in the company or market - and 'reverting to the mean' through pure timing is a coin flip with extra steps.
Hiring. Interview panels who've seen three weak candidates in a row become more lenient with the fourth. Logically, the fourth candidate's quality is independent of the others. Practically, fatigue and the desire to find someone - anyone - bias the assessment. This is the gambler's fallacy plus motivated reasoning, and it's everywhere in recruitment.
Medical decisions. A doctor who's diagnosed five flu cases in a row may be slightly more likely to dismiss the sixth patient's atypical symptoms as 'probably flu.' The base rate of flu hasn't changed, but their recent run has shifted their priors in a non-evidence-based way.
Childbirth and family planning. Couples with three children of the same sex sometimes report believing the fourth is 'bound to be different.' It isn't. Each conception is roughly independent, ~50/50.
Lottery players. People who track 'overdue' numbers and play them disproportionately. Lottery balls have no memory. The expected value of every ticket is identical (and negative).
## How to Avoid the Gambler's FallacyKnowing about the gambler's fallacy is not enough - most people who can recite the textbook definition still fall for it in practice. A few practical techniques actually help:
Ask whether the events are genuinely independent. This is the single most useful habit. Before reasoning about a streak, explicitly ask: does the outcome of event A change the probability of event B? If the answer is no (coins, dice, lotteries, well-shuffled cards), the streak is meaningless. If the answer is yes (sports, market dynamics, skill-based games), reason about the underlying mechanisms, not the streak.
Quantify, don't intuit. When you find yourself thinking 'X is due,' write down the actual probability. After ten consecutive heads, the probability of another head is 50%. Writing it down forces the system 2 part of your brain to engage.
Recognise the words. 'Due,' 'overdue,' 'on a streak,' 'cold,' 'on a run,' 'about to break,' 'has to balance out' - these phrases are red flags. When you hear them (or say them), pause.
Distrust your sense of randomness. Genuinely random sequences produce more clustering than people expect. If a streak feels suspiciously long, it probably isn't - six heads in a row will happen roughly 1 in 64 times, which means you'll see it occasionally if you flip enough coins.
Keep a decision log. Write down predictions before outcomes happen. Keeping yourself honest with a track record exposes how often your 'due' instincts are wrong. (We cover the discipline of calibration training in more depth elsewhere - it's the single best protection against this and most other biases.)
Know that knowing isn't enough. The most surprising research on cognitive biases is that awareness of a bias doesn't reliably eliminate it - including in people who study them professionally. The defence is structural: rules of thumb, decision frameworks, and pre-committed processes that don't rely on you spotting the bias in the moment.
**A useful test** If you're tempted to bet - or invest, or decide - based partly on what's happened recently, ask: 'Would I make this same decision if I had no idea what just happened?' If the answer is no, you're probably reasoning from streaks rather than from the underlying probabilities. ## Connecting It to Better Decision-MakingThe gambler's fallacy belongs to a family of biases that share a common root: our discomfort with genuine randomness. Closely related are:
If there's a single skill that protects against all of these, it's thinking in probabilities rather than certainties. People who internalise probabilistic thinking don't ask 'will this happen?' They ask 'how likely is this to happen, what would change my estimate, and how much should I act on a given probability?'
That shift - from binary, narrative thinking to graded, probabilistic thinking - is the foundation of every reliable framework for making decisions under uncertainty, from expected value calculations to Kelly criterion bet sizing to Bayesian updating. It also helps to be fluent in the language random outcomes are usually quoted in - see our guide to probability vs odds for a full walkthrough of decimal, fractional, and American formats and what they really mean. The gambler's fallacy is what happens when narrative thinking ('the streak must end') replaces probabilistic thinking ('the next event has its own probability, independent of the streak').
**Q: Is the gambler's fallacy ever correct?** Only when events are not actually independent. In sampling without replacement (drawing cards from a deck without reshuffling), past outcomes do affect future probabilities. If you've drawn three aces from a 52-card deck, the probability of the next card being an ace really has dropped. But for independent events - coin flips, roulette spins, lottery draws - the gambler's fallacy is always wrong. **Q: How is the gambler's fallacy different from regression to the mean?**Regression to the mean is a real statistical phenomenon: extreme outcomes tend to be followed by less extreme ones, on average. The gambler's fallacy is the false belief that this happens through compensating streaks. The difference is that regression operates over many trials and reflects the influence of luck, while the gambler's fallacy claims individual streaks must correct in the short term. Regression: real and slow. Gambler's fallacy: imaginary and fast.
**Q: Why do casinos display recent results if the games are random?** Specifically because the games are random. Knowing that the wheel doesn't have memory has no effect on the next spin's probability - but knowing the recent history triggers the gambler's fallacy in players, leading them to bet more, often on outcomes they perceive as 'due.' The display is a profit-driving tool, not transparency. The same logic applies to lottery 'overdue numbers' websites. **Q: Does the gambler's fallacy apply to the stock market?** Partially. Stock markets are not independent in the strictest sense - fundamentals, sentiment, and economic conditions create real correlations. But for short-term price movements, betting that a stock is 'due' to bounce after a few down days is mostly the gambler's fallacy in financial costume. Mean reversion is real over long horizons (years) but unreliable over short ones (days/weeks). **Q: Can I train myself out of the gambler's fallacy?** Partially, but not completely. Awareness helps modestly, and structured techniques (writing down probabilities, using decision frameworks, keeping prediction journals) help more. But the bias is built deep into how human cognition handles randomness, and even statisticians sometimes catch themselves doing it. The best defence is structural: don't let yourself bet, invest, or decide based on streaks unless you can articulate a non-streak-based reason. --- # Second-Order Thinking: How to See Around Corners Source: https://expectedvalue.co.uk/blog/second-order-thinking/ # Second-Order Thinking: How to See Around Corners Most decisions are made on first-order effects, but second-order consequences are where the surprises live. A framework for seeing past the obvious. ## Why is first-order thinking not enough? **Answer** Second-order thinking, popularised by Howard Marks in The Most Important Thing, is reasoning about the consequences of consequences rather than just the immediate effect. First-order: 'X is a good company so its stock will rise'. Second-order: 'X is widely known to be a good company so its stock already prices that in; the next move depends on whether reality beats or misses the consensus expectation'. Practical heuristic: ask 'and then what?' two or three times before acting; most expensive decisions look obvious from the first order and reverse at the second or third.Most decisions are made on the first thing the decision feels like it will do. The obvious effect. The one anyone can see by glancing at the situation.
First-order thinking sounds like: 'Raising the minimum wage will raise wages for low-paid workers.' 'Cutting interest rates will boost growth.' 'A bigger bonus will motivate the team.' 'Banning a thing I disapprove of will reduce that thing.' Each statement is plausibly true on its own. None of them are complete.
Second-order thinking is the discipline of pausing before you act and asking: and then what? What does this set in motion? Who responds? What incentives does this create? What was previously impossible that is now possible - and vice versa? The decisions that look obvious in first-order terms are often catastrophic in second-order terms, which is why almost every interesting failure in policy, investing, or business looks blindingly stupid in retrospect.
## What Second-Order Thinking Actually IsSecond-order thinking is the practice of forecasting beyond the immediate consequence of an action to the chain of effects it produces - particularly the responses of other agents (people, markets, institutions) to the new state of the world.
It's a small idea with outsized leverage. Three rules of thumb capture most of it:
If you remember nothing else, remember this question: and then what happens? Ask it twice.
## Worked Example 1: Rent ControlFirst-order thinking: Rents are too high. Cap them. Now rents are not too high.
Second-order: With a cap, landlords convert rental units to condos for sale, withdraw units from the market, or invest less in maintenance. Supply contracts. The set of available rentals shrinks.
Third-order: Existing tenants under the cap stay put for years longer than they otherwise would (their below-market rent is too valuable to give up), reducing turnover and worsening the housing-shortage feedback loop for new arrivals.
The first-order effect is real - rents are lower for the units that exist. But the second- and third-order effects mean fewer units exist, less mobility, slower maintenance, and ultimately worse outcomes for the people the policy was designed to help. Most economists across the political spectrum agree on this not because of ideology but because the second-order effects are predictable from basic supply-and-demand reasoning. The cost is hidden in the absence of housing that was never built - which is exactly the kind of cost first-order thinking can't see.
## Worked Example 2: Cutting Interest RatesFirst-order: Lower rates make borrowing cheaper. Companies invest more. Households spend more. Growth picks up.
Second-order: Cheap money inflates the price of yield-bearing assets - bonds, dividend stocks, property. Pension funds chase higher returns into riskier assets to meet their obligations. Zombie companies that should have failed survive on cheap debt, locking up labour and capital that more productive firms could otherwise use.
Third-order: When rates eventually rise, the asset-price adjustment is sharper because the build-up was larger. The zombies fail in a wave rather than as a steady trickle. Pension funds with embedded leverage face margin calls (see UK Liability-Driven Investment, autumn 2022). The system is more fragile because of the period of stability.
This isn't an argument against rate cuts. It's an argument that the cost of any decision is the full chain of consequences, not just the immediate effect anyone can see. People who pre-paid the second-order cost in 2022 had been thinking about it since 2015.
## Worked Example 3: A Bigger BonusFirst-order: Pay people more for hitting targets. They hit targets more.
Second-order: People optimise for the measurable targets and let the unmeasured but important things slip. Sales bonuses lead to over-promising; engineering velocity bonuses lead to gaming velocity points; surgical reward systems based on outcome rates lead surgeons to refuse difficult cases.
Third-order: Once a metric is gamed, it stops working as a measure of the underlying thing it was meant to track. The organisation can no longer tell whether the underlying thing is improving. Trust in metrics generally erodes. New metrics are introduced; they get gamed too.
Goodhart's law - when a measure becomes a target, it ceases to be a good measure - is the second-order effect of incentive design, and it is the reason every honest manager has a deep ambivalence about KPIs. The fix isn't 'pick better metrics' (people will game those too). It's accepting that any metric will degrade once people optimise for it, and rotating measurement, holding people to qualitative judgement alongside numbers, and tolerating the messiness of human evaluation.
## Chesterton's Fence: A Defensive ApplicationG.K. Chesterton's parable: you come across a fence in a field. You want to remove it because it serves no obvious purpose. Chesterton's rule is that you don't get to remove the fence until you've explained why it was put there in the first place.
The parable is a defensive form of second-order thinking. Existing institutions, rules, and traditions usually exist because someone, at some point, was solving a real problem with them. The first-order view says 'this is inefficient, remove it.' The second-order view says: and then what was it doing that we'll now stop doing?
This applies just as strongly to internal company processes ('why is this approval step here?'), to legal frameworks ('why is this regulation on the books?'), and to your own habits ('why do I always check X before Y?'). Sometimes the answer is genuinely 'no good reason, remove it.' Often the answer is 'because the last time we didn't, an expensive thing happened.' You don't get to know which until you ask.
## Where Second-Order Thinking Pays Off MostInvesting. First-order: 'Earnings up → stock up.' Second-order: 'Earnings up but expectations were higher → stock down.' Markets price expectations, not absolute outcomes. Howard Marks (Oaktree Capital) writes about this constantly - most institutional investors are first-order thinkers, which is why second-order thinkers can outperform them.
Public policy. Almost every policy with unintended consequences is a first-order win that ignored the second-order response of the people affected. Drug prohibition shifted markets to more dangerous synthetic alternatives. Three-strikes laws changed plea-bargaining dynamics. The general lesson: people respond to incentives, even when the response is illegal or socially costly.
Technology. First-order: 'A new tool makes existing tasks faster.' Second-order: 'It changes which tasks are worth doing, and which jobs are worth having.' This is the entire history of mechanisation, and is what's currently playing out for knowledge work as AI tooling matures.
Personal decisions. First-order: 'Take the higher-paying job.' Second-order: 'You'll see less of your kids during the years you'd remember most.' First-order: 'Skip the gym today.' Second-order: 'You're slowly building a habit of skipping the gym.' Most regret is second-order regret.
Negotiation. First-order: 'Get the best deal.' Second-order: 'You'll be working with this person again - what does winning hard now cost you next time?'
Engineering. First-order: 'Add this feature, the customer asked.' Second-order: 'You'll have to maintain it forever, and it constrains every future design decision.' The cost of code is in maintenance, not creation.
## How to Actually Do It - **Write down the first-order effect explicitly** - Before reasoning about consequences, write the obvious case in one sentence. 'If we do X, the immediate effect is Y.' Most second-order failures start with people skipping this and assuming the first-order effect is obvious enough not to state. - **Ask 'and then what?' three times** - Once. Twice. Thrice. Most decisions only need two iterations to surface the important second-order effect; three iterations forces you to think about feedback loops and equilibrium states. - **List who responds and how** - Second-order effects almost always come from other agents adapting. Make them explicit: customers, competitors, regulators, your own team, the markets. What does each one do differently in the new world? - **Identify the time lag** - First-order effects are immediate. Second-order effects show up over months or years. If your decision rule depends on the second-order effect not having time to develop, you're betting on a short window - which is rarely a good bet. - **Pre-mortem the second-order failure mode** - Imagine it's a year from now and the decision has clearly backfired through second-order effects. Write the post-mortem. What chain of events caused the failure? This is the same idea as Daniel Kahneman's pre-mortem technique applied specifically to second-order risks. - **Steelman the status quo** - Before changing something, force yourself to articulate why the current arrangement might be good - even if you disagree with it. This is Chesterton's fence in checklist form. Skip the change if you can't find a coherent reason for the existing setup. ## When Second-Order Thinking FailsSecond-order thinking has its own failure modes - three in particular.
Analysis paralysis. Every decision has effects that ripple infinitely. At some point you have to act on incomplete information. The discipline isn't to think more but to think one level deeper than the people you're competing with. If everyone else is first-order thinking, second-order is enough. If everyone else is second-order thinking, you'll need third-order - and even then, returns to depth diminish quickly.
Overconfidence in causal chains. It's easy to invent a plausible-sounding second-order story that turns out to be wrong. Real-world systems have many feedback loops, and your model might miss the dominant one. Mark Twain put it cleanly: it ain't what you don't know that gets you into trouble. It's what you know for sure that just ain't so. Second-order chains feel insightful precisely because they're elaborate; that's also how they fool you.
Mistaking complexity for depth. A long chain of second-order reasoning is not the same as a deep one. A short, well-calibrated forecast usually beats a sprawling, poorly-grounded one. If you find yourself five steps deep with low confidence at each step, your overall confidence should be very low - multiplied probabilities collapse fast. See risk vs uncertainty for the distinction between situations where probabilities can be estimated and situations where they can't.
## Frequently Asked Questions **Q: Where did the term 'second-order thinking' come from?**The framing has roots in systems theory and economics. Howard Marks popularised it for investors in The Most Important Thing (2011), and Ray Dalio uses similar language in his decision-making frameworks. The underlying idea - that the response of agents to a change is part of the cost of the change - is older, going back at least to Frédéric Bastiat's 1850 essay 'That Which Is Seen, and That Which Is Not Seen'.
**Q: How is second-order thinking different from systems thinking?** Systems thinking is broader - it's the discipline of mapping all the parts of a system, their connections, and the feedback loops between them. Second-order thinking is one practical heuristic from systems thinking: when you act on the system, what does the system do back? You can do useful second-order thinking without building a full systems-thinking map. **Q: Is second-order thinking the same as 'galaxy-brained' reasoning?**Almost the opposite. Galaxy-brained reasoning is the failure mode where chains of plausible-sounding logic conclude something obviously wrong (often that you should defect from a sensible commitment). Good second-order thinking has a check against this: are the second-order effects you're forecasting actually likely, or do you only believe them because they support the conclusion you wanted?
**Q: How does this connect to expected value?**Expected value is about quantifying the value of an outcome given probabilities. Second-order thinking is about identifying which outcomes belong in the EV calculation in the first place. If you only EV the first-order outcome, your calculation is right but the inputs are incomplete. See expected value explained for the maths and base rate neglect for how unaided intuition gets the inputs wrong.
**Q: Should I always think second-order?** No. Most decisions don't matter enough to justify the effort. Reserve second-order thinking for high-stakes, low-reversibility decisions: career moves, large investments, hires, product strategy, policy choices. For everyday decisions ('which restaurant?', 'reply to this email how?'), first-order is fine and second-order is overthinking. **Q: What's the easiest way to start practising it?** Pick one important decision per week and write the second-order effects in a short note before deciding. The act of writing forces precision. After a few months, compare what you wrote against what actually happened. Calibration improves with feedback. You'll find that some types of second-order forecast (incentives, market reactions) you do well, and others (technology adoption, social trends) you struggle with - knowing your own pattern is half the value. ## Related ReadingSecond-order thinking pairs naturally with several adjacent decision-making concepts:
A doctor tells you that you've tested positive for a rare disease. The test is 99% accurate. How worried should you be?
Most people say very worried. The test is 99% accurate, after all. But if the disease affects 1 in 10,000 people, the real probability that you actually have the disease is roughly 1%. The other 99% of positive results are false alarms - what statisticians call the false positive paradox.
This is base rate neglect - our systematic failure to account for how common or rare something is before updating our beliefs with new evidence. It is one of the most pervasive and consequential cognitive biases, and it distorts decisions in medicine, investing, hiring, criminal justice, and everyday life.
If you've read about thinking in probabilities, you already know that our brains are poorly calibrated for risk. Base rate neglect is a specific mechanism behind that miscalibration, and understanding it will sharpen every probabilistic judgment you make.
## What is the famous medical test problem?Let's work through the classic example with real numbers. Suppose:
Now imagine we test 1,000,000 people.
People who actually have the disease: 100 (that's 1 in 10,000 of one million)
People who don't have the disease: 999,900
Total positive results: 99 + 9,999 = 10,098
Probability you actually have the disease given a positive test: 99 / 10,098 = 0.98%
Read that again. A 99% accurate test, and the chance you're actually sick is less than 1%. The base rate of the disease (1 in 10,000) is doing almost all the work. The test's accuracy is almost irrelevant when the condition is rare enough.
This is precisely why mass screening programmes for rare conditions are so problematic. The UK's National Screening Committee explicitly considers base rates when evaluating whether to roll out population-wide tests - because a flood of false positives causes anxiety, unnecessary follow-up procedures, and wasted resources.
This is Bayesian thinking in action: the prior probability (base rate) matters enormously, and no amount of evidence quality can overcome a sufficiently low prior.
**The Base Rate Test**Whenever someone gives you a conditional probability ('this test is 99% accurate'), immediately ask: what is the base rate? How common is the thing being tested for? Without the base rate, the accuracy figure is almost meaningless.
## How do base rates change investing decisions?Base rate neglect is rampant in investing, and it costs people real money.
Consider the base rates that most investors conveniently ignore:
Yet people routinely pick individual stocks, choose active fund managers, and try day trading. Why? Because they focus on the vivid stories - the friend who bought Tesla early, the fund manager profiled in the Financial Times - while ignoring the base rate of failure.
The base rate for stock-picking success is abysmal. If you're picking individual stocks, you're not competing with average investors - you're competing with quantitative hedge funds, PhD researchers, and algorithms that process information in microseconds. The base rate of outperformance for a retail stock-picker is somewhere in the neighbourhood of 10-20% over a decade, and much of that is luck.
This is directly connected to the concept of expected value. Even if you can identify a stock that will 10x, the expected value of your stock-picking strategy depends on the base rate of your picks being correct multiplied by the payoff. For most people, the expected value of passive indexing is simply higher.
The Kelly Criterion makes this even more explicit: optimal bet sizing depends on your actual edge, and if your edge is zero or negative (as the base rates suggest for most stock-pickers), the Kelly-optimal allocation to individual stock bets is zero.
## What do base rates tell you about startups and hiring?The base rates for startups are well-documented and sobering:
Yet every founder believes they're in the 10% that will succeed. This is textbook base rate neglect combined with overconfidence bias. When Y Combinator surveys its applicants, the average founder estimates their chance of success at around 60% - six times the actual base rate.
Does this mean nobody should start a company? Not at all. But it means you should calibrate your confidence to the base rate and then adjust based on specific evidence. If you have domain expertise, a proven co-founder, early revenue, and a defensible market position, your odds are genuinely better than 10%. But if you're a first-time founder with an untested idea, your starting estimate should be much closer to the base rate than your gut tells you.
Companies routinely fall victim to base rate neglect in hiring. An interviewer spends 45 minutes with a candidate who is articulate, confident, and charming. They come away convinced this person will be a top performer.
But what's the base rate? Research suggests that unstructured interviews predict job performance with a correlation of about 0.20 - barely better than chance. Even structured interviews achieve only about 0.40. The base rate of any single hiring signal being predictive is low.
The implication: rather than trusting your gut feeling from an interview, you should weight the base rate of candidates from similar backgrounds performing well in similar roles, then modestly update based on what you observe. A structured process with multiple independent signals (work samples, references, structured questions) gives you more data points to update away from the base rate - but the base rate should always be your starting point.
Base rate neglect shows up in mundane decisions too:
Kahneman and Tversky identified base rate neglect in the 1970s, and decades of research have clarified why it happens:
1. The representativeness heuristic. We judge probability by how well something matches a mental prototype. If someone 'looks like' a successful entrepreneur (confident, articulate, Stanford MBA), we assign high probability to their success - ignoring that the base rate of startup success is low regardless of how someone presents.
2. Vivid evidence crowds out statistics. A single compelling story (a friend who beat cancer after a positive test, a startup that defied the odds) is psychologically more powerful than dry statistics. Our brains weight narrative evidence far more heavily than base rate data.
3. Base rates are abstract. The number '1 in 10,000' doesn't feel like anything. A positive test result feels like something. We process feelings faster than statistics, and in the absence of deliberate analysis, feelings win.
4. We anchor on the most available number. When told a test is '99% accurate,' that 99% figure becomes an anchor. We adjust insufficiently from it, even when the base rate should dominate our calculation.
5. Ego protection. Accepting base rates means accepting that we're probably average - that our stock picks probably won't beat the market, that our startup probably won't succeed, that our judgement probably isn't as good as we think. Base rate neglect is partly motivated reasoning: we want to believe we're the exception.
## How do you train yourself to think in base rates?The good news is that base rate neglect is one of the more correctable cognitive biases. Here's a practical framework:
Before evaluating any specific case, find the base rate. This is your prior probability - the starting point before any specific evidence.
Make it a habit to never form a judgement without first establishing the base rate. Write it down. Put a number on it.
Once you have the base rate, update it with specific evidence - but do so properly. The framework from Bayesian thinking for everyday decisions applies directly:
In the medical test example: start with 1/10,000, then update with the positive test result. The update is significant (the test is 99% accurate), but it only brings you to about 1% - nowhere near the 99% that naive intuition suggests.
Research by Gerd Gigerenzer shows that people reason about base rates much better when information is presented as natural frequencies rather than percentages.
Instead of: 'The test is 99% accurate and the disease prevalence is 0.01%'
Think: 'Out of 10,000 people, 1 has the disease. If we test all 10,000, we'll get 1 true positive and about 100 false positives. So 1 out of 101 positive results is real.'
Natural frequencies make the base rate visible. The numbers become concrete and countable rather than abstract.
Build a mental (or physical) library of base rates for domains you care about:
The more base rates you have memorised, the harder it becomes to neglect them.
Calibration training - where you make probabilistic predictions and track your accuracy - is one of the most effective ways to overcome base rate neglect. Over time, you develop better intuitions about prior probabilities because you've been forced to confront the gap between your estimates and reality.
Philip Tetlock's research on superforecasters found that the best predictors were distinguished not by intelligence but by their willingness to start with base rates and update incrementally. They resisted the temptation to anchor on vivid narratives.
**The Superforecaster's Secret**In Superforecasting, Philip Tetlock found that the single most important habit of the best predictors was starting with the 'outside view' - the base rate - before considering the specifics of any individual case. They asked 'how often does this kind of thing happen?' before asking 'what's special about this particular case?'
## When are base rates a competitive advantage?Here's the optimistic conclusion: because most people neglect base rates, those who don't have a genuine edge.
In investing, simply accepting the base rate that most stock-picking fails - and indexing accordingly - puts you ahead of the majority of investors over a 20-year period. You don't need to be smarter. You just need to accept the base rate.
In business, founders who honestly assess the base rate of failure and plan accordingly (maintaining runway, testing assumptions cheaply, having a plan B) survive at higher rates than those running on pure optimism.
In medicine, doctors trained in Bayesian reasoning order fewer unnecessary follow-up tests, reducing patient anxiety and healthcare costs.
Base rate thinking is not pessimism. It's calibration. It means starting with reality and adjusting from there, rather than starting with hope and ignoring the evidence. The goal isn't to never take risks - it's to take risks with your eyes open, understanding the true probabilities rather than the ones your intuition fabricates.
As we explore in Thinking in Probabilities, the world rewards those who see it clearly. And seeing clearly starts with base rates.
For a deeper dive into the mathematical foundations of updating beliefs, see our guide to Bayesian Thinking for Everyday Decisions. And for more on how to apply probabilistic reasoning to practical decisions about money and risk, explore Expected Value Explained and The Kelly Criterion.
For further reading on the psychology of judgement and decision-making, many of the best books on probabilistic thinking cover base rate neglect in depth - particularly Kahneman's Thinking, Fast and Slow and Gigerenzer's Reckoning with Risk.
**Q: What is base rate neglect?** Base rate neglect (also called the base rate fallacy) is a cognitive bias where people ignore or underweight the general prevalence of something (the base rate) when evaluating a specific case. For example, ignoring how rare a disease is when interpreting a positive test result, leading to a dramatic overestimate of the probability of actually being ill. **Q: What is the difference between base rate neglect and the base rate fallacy?** They refer to the same phenomenon. 'Base rate fallacy' emphasises the logical error, while 'base rate neglect' emphasises the psychological mechanism - our tendency to ignore or underweight prior probabilities. Both terms are used interchangeably in the literature. **Q: How does base rate neglect relate to Bayes' theorem?** Bayes' theorem is the mathematical framework for correctly combining base rates (prior probabilities) with new evidence (likelihoods) to arrive at updated probabilities (posteriors). Base rate neglect occurs when people skip the prior probability step and focus only on the new evidence, leading to miscalculated posteriors. **Q: Can you give a simple example of the base rate fallacy?** Suppose 1 in 100 people in a car park are car thieves. A security guard correctly identifies suspicious behaviour 90% of the time, but also falsely flags 10% of innocent people. If the guard flags someone, the chance that person is actually a thief is only about 8% - not 90% - because the base rate of thieves (1%) is so low. Most flagged people are innocent. **Q: How can I avoid base rate neglect in my own decisions?** Develop the habit of asking 'what is the base rate?' before evaluating any specific case. Use natural frequencies (e.g., '1 out of 100 people') instead of percentages, which makes base rates more intuitive. Study Bayesian reasoning and practise calibration exercises to improve your probabilistic intuitions over time. --- # The Kelly Criterion: How to Size Your Bets Optimally Source: https://expectedvalue.co.uk/blog/kelly-criterion-optimal-bet-sizing/ # The Kelly Criterion: How to Size Your Bets Optimally The Kelly Criterion tells you exactly how much to stake when you have an edge. Formula derivation, fractional Kelly, history, and worked examples for 2026. **Answer**The Kelly Criterion tells you what fraction of your bankroll to stake on a positive-expected-value bet to maximise long-run growth. The formula: f* = (bp - q) / b, where b is net odds, p is win probability, and q = 1 - p. The formula was published by J.L. Kelly Jr. at Bell Labs in 1956 and popularised by Edward Thorp in blackjack and quant finance. Most practitioners use fractional Kelly (half or quarter) to reduce variance at small cost to growth.
## Risk-of-ruin calculator Even with an edge, a small bankroll can be wiped out by a bad run. This shows the probability of eventually going broke staking one unit at a time on even-money bets. Interactive calculator. Computes: Risk of eventually going broke. ## Why is bet sizing such a hard problem? Imagine you've found a coin that lands heads 60% of the time. You have £10,000 to bet on repeated flips, and each bet pays even money (win = double your stake, lose = lose your stake). You clearly have an edge. But how much should you bet on each flip? Bet too little, and you're leaving money on the table. Bet too much, and a run of bad luck will wipe you out before your edge has time to compound. The extremes are obvious - betting 1% of your bankroll is overly conservative, while betting 100% guarantees eventual ruin (you only need one loss). This is the bet-sizing problem - separate from the question of whether you have an edge at all (which is what expected value tells you) - and it has an elegant mathematical solution: the Kelly Criterion, developed by John L. Kelly Jr. at Bell Labs in 1956. Originally designed for information theory problems, it quickly found applications in gambling, investing, and anywhere else that decisions must be sized under uncertainty. ## Who invented the Kelly Criterion?John Larry Kelly Jr. (1923-1965) was a Texas-born physicist who joined Bell Labs in 1953 after a PhD from the University of Texas at Austin. He worked in the same group as Claude Shannon, the father of information theory, and the 1956 paper that gave us the Kelly Criterion sits squarely inside Shannon's intellectual world.
The paper - A New Interpretation of Information Rate, published in the Bell System Technical Journal - wasn't framed as a gambling formula. Kelly was answering a Shannon-style question: if you receive a noisy signal that gives you a probabilistic edge over the market price of a bet, what is the maximum rate at which information (and therefore money) can flow to you? The answer turned out to be a logarithmic utility function, and the optimal staking fraction that maximises long-run growth dropped out as a corollary.
Kelly himself never gambled on his own formula. He died of a brain haemorrhage in 1965 at age 41, having spent his career on speech synthesis and information theory rather than betting. The formula's translation to gambling was the work of Edward Thorp, an MIT mathematician who used it to size his blackjack bets at Reno and Lake Tahoe casinos in the early 1960s. Thorp later extended the same framework to convertible bond arbitrage at Princeton-Newport Partners, one of the first quantitative hedge funds.
The path from a 1956 information-theory paper to modern quantitative finance runs directly through Thorp's 1962 book Beat the Dealer, which made counting cards public and gave the Kelly Criterion its first popular audience. By the 1980s the formula was core curriculum at every quant trading desk.
## What is the Kelly formula?For a simple bet with two outcomes (win or lose), the Kelly Criterion says to stake the following fraction of your bankroll:
f* = (bp - q) / b
Where:
For our biased coin at even money:
Kelly says to bet exactly 20% of your current bankroll on each flip. This fraction maximises the expected logarithm of wealth - or equivalently, maximises the long-term compound growth rate of your bankroll.
## How is the Kelly formula derived? The derivation is short enough to do in one sitting and worth understanding because it explains why Kelly is uniquely correct rather than one option among many. Setup. You bet a fraction f of your bankroll W on a wager that pays net odds b with win probability p. After one bet, your bankroll is either W(1 + bf) on a win or W(1 - f) on a loss. Step 1: Expected log-wealth. Because gains compound multiplicatively rather than additively, we want to maximise the expected logarithm of wealth rather than expected wealth itself. After N bets, log-wealth grows linearly with the per-bet expected log-return: E[log(WN)] = log(W0) + N × E[log(1 + return)] Step 2: One-bet expected log-return. G(f) = p × log(1 + bf) + q × log(1 - f) This is the expected geometric growth rate per bet as a function of the stake fraction f. Step 3: Differentiate and set to zero. dG/df = pb / (1 + bf) - q / (1 - f) = 0 Multiplying through and rearranging: pb(1 - f) = q(1 + bf) pb - pbf = q + qbf pb - q = pbf + qbf pb - q = bf(p + q) f* = (bp - q) / b (since p + q = 1) Step 4: Second derivative check. The second derivative is negative everywhere on (0, 1/b), confirming f* is a maximum and not a minimum. Kelly is therefore the unique fraction that maximises expected log-wealth - any other fraction grows your bankroll more slowly in the long run. What this proves. The derivation is constructive: nothing was assumed about your risk preferences or your time horizon beyond the geometric-growth framing. If you bet repeatedly and care about long-run multiplicative wealth, Kelly is optimal. If you care about something else - drawdown psychology, finite horizons, uncertain edge - that's where fractional Kelly comes in. ## Why does Kelly maximise growth?The Kelly Criterion doesn't maximise expected value of your bankroll (that would mean betting everything). Instead, it maximises the expected geometric growth rate. This distinction is crucial.
With a 60% coin at even money, betting 20% each time:
The expected growth factor per bet is:
That's about 2% growth per bet. After 100 bets, your expected bankroll is £10,000 × 1.020^100 ≈ £72,400.
Now consider what happens if you over-bet at 50% of bankroll:
Despite having a clear edge, over-betting produces a growth factor below 1.0 - your bankroll will shrink over time! This is the tragedy of over-betting: you can have a genuine advantage and still go broke.
The deeper reason Kelly works - and why log-wealth optimisation matters - is the concept of ergodicity: for repeated personal bets, the time-averaged growth rate diverges from the ensemble expected value, and Kelly aligns you with the path you actually live.
- **5% of bankroll**: Growth factor: 1.005 per bet (conservative) - **10% of bankroll**: Growth factor: 1.012 per bet - **20% of bankroll (Kelly)**: Growth factor: 1.020 per bet (optimal) - **30% of bankroll**: Growth factor: 1.017 per bet (over-betting) - **40% of bankroll**: Growth factor: 1.005 per bet (severely over-betting) - **50% of bankroll**: Growth factor: 0.953 per bet (bankroll shrinks!) - **100% of bankroll**: Growth factor: 0.000 (guaranteed ruin) ## Full Kelly or fractional Kelly: which should you use? In theory, full Kelly is optimal. In practice, almost everyone uses a fraction of Kelly - typically between one-quarter and one-half. Here's why: 1. Probability estimates are uncertain Kelly assumes you have a true probability p in hand. In practice that's almost never the case - you have a guess at p, and the gap between measurable risk and unmeasurable uncertainty matters enormously here. The distinction between risk and uncertainty explains why fractional Kelly is the safe default rather than a fudge factor. Kelly assumes you know the true probability. But in real life, your estimate of p might be wrong - and as we explore in Thinking in Probabilities, our brains are systematically biased in how we assess likelihood. If you think you have a 60% edge but it's actually 52%, full Kelly based on the wrong probability can lead to severe over-betting. Calibration training is the practical fix: it shrinks the gap between your stated probabilities and reality, so the input to Kelly is honest. 2. Drawdowns are psychologically brutal Full Kelly can produce drawdowns of 50-80% from peak. Most humans cannot stick to a strategy through such volatility. A smaller stake produces smoother growth, which means you're more likely to actually follow through. 3. Kelly assumes infinite time horizon If you need the money within a specific timeframe, the variance of full Kelly might be unacceptable. Half-Kelly gives you 75% of the growth rate with significantly reduced variance. 4. Real-world frictions Transaction costs, taxes, liquidity constraints, and the inability to bet exactly the right fraction all erode the theoretical optimality of full Kelly. The common recommendation among quantitative traders and professional gamblers is half-Kelly - it captures most of the growth rate at a fraction of the variance and survives realistic levels of probability-estimate error. - **Full Kelly (20%)**: Max growth: 1.020/bet | Max drawdown: ~60% | P(halving): ~15% - **3/4 Kelly (15%)**: Growth: 1.018/bet | Max drawdown: ~45% | P(halving): ~5% - **Half Kelly (10%)**: Growth: 1.012/bet | Max drawdown: ~30% | P(halving): ~1% - **Quarter Kelly (5%)**: Growth: 1.005/bet | Max drawdown: ~15% | P(halving): <0.1% ## How does Kelly handle multiple simultaneous bets? The single-bet Kelly formula assumes you make one wager, observe the result, then make the next. Real bettors and portfolio managers rarely have that luxury - you have positions open on five sports matches concurrently, or hold a dozen stocks at once. Independent simultaneous bets. When the outcomes are statistically independent, you can apply Kelly to each bet in isolation - but only if the combined stake doesn't exceed your bankroll. A practical adjustment is to compute the unconstrained Kelly fraction for each bet, sum them, and if the total exceeds 1.0, scale every position proportionally. This preserves the ratio of allocations while keeping you solvent. Correlated bets. Correlation is where naive Kelly goes badly wrong. Suppose you've identified value on three Premier League matches all played on the same wet weekend - rain helps underdogs, so a wet weekend means your three bets are positively correlated. The single-bet Kelly formula treats them as independent and over-allocates because it ignores the compounding downside of all three losing together. The correct fix is the multivariate Kelly formulation, which in its continuous form looks identical to Markowitz portfolio optimisation: f* = Σ⁻¹ × μ Where Σ is the covariance matrix of returns and μ is the vector of expected excess returns. The matrix inverse punishes correlation: two highly correlated bets get a smaller combined allocation than the sum of their independent Kelly fractions. Practical heuristic. If you don't want to estimate a full covariance matrix, a useful rule of thumb is: estimate the correlation between each pair of bets, multiply each pair's combined Kelly stake by (1 - ρ), and shrink the total until the worst-case drawdown is acceptable. Half-Kelly with this correlation haircut survives almost every realistic correlated-bet scenario. The practical lesson is the same in finance and sports: always ask what correlates these positions before sizing them. A perfectly Kelly-sized portfolio of correlated bets is no longer Kelly-sized at all. ## How do you use Kelly for investing?The Kelly framework extends beyond simple win/lose bets to continuous outcomes like stock returns. For an investment with expected return μ and variance σ², the Kelly fraction is approximately:
f* = μ / σ²
This is sometimes called the Merton share in finance. It says you should invest more aggressively when:
For example, if a stock index has expected excess return of 6% per year and volatility of 16%:
Full Kelly suggests 234% allocation - i.e., heavy leverage! This illustrates why fractional Kelly is essential in practice. Half-Kelly gives ~117%, and quarter-Kelly gives ~59%, which aligns much better with conventional investment wisdom for equity allocation.
Warren Buffett, though he likely doesn't use the formula explicitly, has described his approach in Kelly-like terms: "When you have a big edge, bet big. When you don't, don't bet at all." His concentrated portfolio reflects Kelly thinking - large positions where he has high conviction.
## How did Edward Thorp apply Kelly in the real world?If Kelly is the formula, Edward Thorp is the practitioner who proved it worked. Thorp was an MIT mathematics PhD who, in 1960, discovered that blackjack could be beaten by tracking the count of high cards remaining in the deck. The card-counting edge was small - typically 0.5 to 1.5% on individual hands - but it was real, and Kelly was the natural tool for sizing bets given that small edge.
Thorp's 1962 book Beat the Dealer documented the strategy, ran the math on a £10,000 bankroll played at Reno and Lake Tahoe, and made the Kelly Criterion famous outside Bell Labs. The book triggered a wave of casino countermeasures - more decks, faster shuffles, surveillance cameras - that still shape blackjack today.
Thorp then translated the same framework to financial markets. In 1969 he co-founded Princeton-Newport Partners, widely considered one of the first quantitative hedge funds. The fund traded convertible-bond arbitrage and warrant pricing using statistical edges and Kelly-sized positions. Between 1969 and 1988 it returned roughly 19% annually with very low correlation to the S&P 500 - a profile that became the template for the modern quant industry.
Thorp's practical lessons from those years are worth repeating:
Thorp's autobiography A Man for All Markets remains the best primary-source account of using Kelly in practice across two completely different markets.
## Worked example: betting Premier League footballSuppose you've built a model that identifies value in football match outcomes. Your model rates Team A to win at 55% probability, but the bookmaker is offering odds of 2.10 (implying only ~48% probability).
Calculating the Kelly stake:
Full Kelly says bet 14.1% of your bankroll. Using half-Kelly (a more sensible practical approach), you'd bet 7.05%.
With a £5,000 bankroll, that's:
If your model is well-calibrated and you can find several such bets per week, even quarter-Kelly produces substantial long-term growth. The discipline is in never exceeding your Kelly fraction, even when you feel especially confident about a particular bet.
## Why can't most people actually run Kelly? The math says full Kelly is optimal for long-run growth. The behaviour of real bettors says it's almost never the right choice. The gap between the two is psychology. Drawdown tolerance is the binding constraint. Full Kelly with a 60% coin produces an expected peak-to-trough drawdown of around 60% over a 1000-bet horizon. That's not an unlucky tail outcome - it's the baseline. Few humans can watch their bankroll halve and then halve again and still bet the same fraction on the next flip. The behavioural response is almost always to cut stake size after a drawdown, which converts a Kelly bettor into an inadvertent anti-martingale bettor and destroys the formula's optimality. Loss aversion biases the recovery. Behavioural finance has documented for fifty years that humans feel losses roughly twice as keenly as equivalent gains (prospect theory, Kahneman and Tversky, 1979). A bettor running full Kelly during a 50% drawdown isn't comparing the current state to the long-run growth path - they're comparing it to the peak, and the emotional pain is dominating the formula. The practical consequence: full Kelly looks fine in simulations and is unsustainable in real accounts. The half-Kelly compromise. Half-Kelly captures three-quarters of the long-run growth rate at roughly half the drawdown depth. That trade is almost always worth it because the strategy only delivers its long-run growth if you stick to it - and the probability that you stick to a strategy is a strong inverse function of how big the drawdowns get. The discipline problem applies to professional traders too.Bayes' theorem matters more than test accuracy. ## False-positive calculator: what a positive test really means Adjust the test accuracy and how common the condition is. A test that is 99% accurate can still make a positive result more likely wrong than right when the condition is rare - that is the paradox. Interactive calculator. Computes: Chance you actually have it, given a positive result; Of every 100 positive results, this many are false alarms.Here's a question almost everyone - including most doctors - gets wrong:
A test for a disease is 99% accurate. The disease affects 1 in 1,000 people. Your test comes back positive. What's the probability you actually have the disease?
Most people answer somewhere between 90% and 99%. The intuitive logic feels airtight: the test is 99% accurate, so a positive result must mean you almost certainly have it.
The correct answer is about 9%.
If you find that hard to believe, you're in good company. In a famous 1978 study at Harvard Medical School, only 18% of doctors got the answer right - and the median answer they gave was 95%, off by a factor of more than ten. The mistake has a name: the false positive paradox, and it's the single most consequential probability error in modern life.
Let's unpack why your intuition is so badly miscalibrated, and why understanding this matters for medical decisions, drug testing, airport security, and the criminal justice system.
## What Is the False Positive Paradox?The false positive paradox is the counterintuitive fact that when you test for a rare condition, even a highly accurate test produces mostly wrong positive results.
The paradox has two ingredients:
When these two things combine, the small false-positive rate gets multiplied by the very large population of healthy people, producing a flood of false alarms that drowns out the real positives. This is sometimes called the base rate fallacy - failing to account for the underlying prevalence of the condition.
This is just Bayes' theorem doing its job. Bayes tells us that the probability of a hypothesis (you have the disease) given evidence (positive test) depends on three things: the test's true positive rate, its false positive rate, and the prior probability of the hypothesis. Most people anchor entirely on test accuracy and ignore the prior. That's the mistake.
## The Math, Step by StepLet's redo the opening problem with concrete numbers, because numbers cut through intuition.
Imagine 100,000 people are tested for a disease that affects 1 in 1,000 (so 100 of them actually have it).
The test is 99% accurate, meaning:
Now let's count.
True positives: 99% of 100 = 99 people (sick, correctly identified)
False positives: 1% of 99,900 = 999 people (healthy, wrongly flagged)
Total positive results: 99 + 999 = 1,098
If you got a positive test, you're in this group of 1,098. Your probability of actually being sick is:
99 ÷ 1,098 ≈ 9%
Not 99%. Not 90%. About 9%. The other 91% of positive tests are false alarms - sitting in waiting rooms wondering if their lives are about to change, while the math quietly says they're almost certainly fine.
**Why this is genuinely surprising** When the false-positive rate (1%) is larger than the disease prevalence (0.1%), false positives mathematically must outnumber true positives - even with a 99% accurate test. The smaller the prevalence, the worse this gets. For a 1-in-10,000 disease tested with a 99% accurate test, only 1% of positives are real. ## Real-World Examples Where This Matters ## 1. Mammography ScreeningFor women in their 40s, breast cancer prevalence in any given year is roughly 0.4%. Mammography typically has around an 87% true positive rate and a 9% false positive rate (numbers from US Preventive Services Task Force literature).
Apply the maths: out of 10,000 women screened, about 40 will have cancer (35 detected), and about 9,960 will not (around 896 false alarms). A positive mammogram in this group means roughly a 4% chance the cancer is real.
This is why screening guidelines have become more cautious over the past decade. The downstream costs of false positives - anxiety, biopsies, occasional unnecessary mastectomies - are real, and at low base rates the false positives can outweigh the benefit of catching the rare true case.
## 2. Workplace and Sports Drug TestingStandard urine drug screens have false positive rates of around 5-10% depending on the substance. If only 5% of employees are using a particular drug, and the test has a 5% false positive rate, then about half of all positive tests will be false alarms - even with a sensitive test.
This is why responsible drug testing programmes always use a confirmatory second test (typically gas chromatography-mass spectrometry) before any consequences are imposed. The first test is a screen, not a verdict. People have been fired, kicked out of competitions, and had their careers destroyed because employers and athletic bodies treated initial positive results as conclusive.
## 3. Airport Security and WatchlistsSuppose airport screening has a 99% true positive rate and a 0.1% false positive rate for catching a terrorist. Sounds good, until you do the maths.
If there's roughly 1 terrorist per 1 million flyers, screening 1 million people produces about 1 true positive (correctly flagged) and about 1,000 false positives (innocent travellers detained). So for every real threat caught, about 1,000 innocent people are wrongly flagged.
This is why no-fly lists and predictive policing systems generate so many complaints - the math guarantees that at very low base rates, the vast majority of flagged people are innocent. It doesn't mean these systems are useless, but it does mean every flagging needs to be treated as a starting point for investigation, not as a conclusion.
## 4. Spam FiltersEmail is one place the false positive paradox works in your favour rather than against. Spam is the common category - about 50-80% of all email traffic is spam - so a positive identification (this is spam) is usually correct even when the filter isn't perfect.
But filters are deliberately tuned with very low false-positive rates, because the cost of marking a legitimate email as spam (missing a job offer, an invoice, a family message) is enormously higher than the cost of letting one spam through. This is a Bayesian decision built into the design: when the cost of false positives is high, you adjust your threshold even if it lets some true positives slip past.
## 5. The Prosecutor's FallacyIn court, expert witnesses sometimes testify that 'the chance of a random person matching this DNA evidence is 1 in 10 million'. Juries often interpret this as 'there's only a 1-in-10-million chance the defendant is innocent'. Those are not the same statement, and confusing them is called the prosecutor's fallacy.
In a country of 60 million people, a 1-in-10-million match rate produces about 6 random matches. If the police searched a database of millions of profiles to find a match, the prior probability of any given match being the real culprit is much less than 100%. This isn't a hypothetical - the famous Sally Clark case in the UK, in which a mother was wrongly convicted of murdering her two infant sons partly on the basis of misapplied probability, hinged on exactly this kind of error.
Whenever you hear a number like '99.9% accurate' presented in a trial, the question to ask is: what was the prior? Without that, the headline accuracy is meaningless.
## Why Your Intuition Fails HereThere are two related cognitive biases that produce this error:
Base rate neglect. When humans are given specific evidence (a positive test, a vivid description, a stereotype), we tend to ignore the base rate - the underlying frequency of the thing in the general population. We focus on the immediate evidence and forget how common the condition actually is. This is closely tied to the broader phenomenon of base rate neglect, which causes mistakes far beyond medical testing.
Confusion of conditional probabilities. P(positive test | sick) and P(sick | positive test) are different quantities. The test's accuracy tells you the first; you usually want the second. People treat these as interchangeable when they are emphatically not.
There's also a deeper reason this is hard: human cognition evolved in environments where rare events were genuinely rare and most signals were trustworthy. We didn't evolve to think clearly about scaling effects, where a small percentage applied to a huge population produces a number that swamps the rare true positives.
## How to Think Clearly About Test ResultsWhen you encounter any test result - medical, security, statistical, legal - work through these three questions before drawing conclusions:
1. What's the prior probability?
How common is the condition in the relevant population? If you don't know, the test result alone is uninterpretable. Ask, look it up, or hold the conclusion loosely.
2. What's the false positive rate?
Not just the headline accuracy - the specific rate at which healthy/innocent/normal cases get flagged. This is often buried in technical literature and not communicated to patients or the public.
3. Use natural frequencies, not percentages
Research by Gerd Gigerenzer (a leading risk-communication scientist) shows that people reason more accurately when probabilities are framed as counts: 'out of 1,000 people, 1 has the disease, 99 healthy people will test positive' is much clearer than '0.1% prevalence and 9.9% false positive rate'. Reframe the problem in terms of populations and counts whenever you can.
**The single best heuristic** When you get a positive result for any rare thing, your first response should not be panic - it should be 'let's confirm this with an independent second test.' False positives are independent across well-designed tests, so a second positive multiplies the evidence dramatically and gets you closer to a real answer. ## Common Mistakes to AvoidMistake 1: Treating accuracy as the only number that matters.
A 99% accurate test sounds excellent. For a common condition, it is. For a 1-in-100,000 condition, it's an alarm system that produces 999 false alerts for every real one. Accuracy is meaningless without prevalence context.
Mistake 2: Assuming independent tests stay independent.
If you take the same flawed test twice, the false positive isn't independent - the same systematic error that flagged you the first time will flag you again. To get the multiplicative benefit of a second test, the second test has to use a different mechanism (different lab, different chemistry, different methodology).
Mistake 3: Updating only when you get a positive.
Bayes works in both directions. A negative result on a test for a rare condition is strong evidence you don't have it (because false negatives become rare when prevalence is low). People over-update on positives and under-update on negatives.
Mistake 4: Forgetting that screening changes the prior.
If you only test people with symptoms, the prior is much higher than the population base rate, and positive predictive value rises accordingly. This is why doctors test based on clinical suspicion rather than screening everyone - it dramatically improves the value of the test.
## Frequently Asked Questions **Q: Is the false positive paradox the same as the base rate fallacy?** They're closely related. The base rate fallacy is the general cognitive error of ignoring how common something is when interpreting evidence. The false positive paradox is one specific consequence - that for rare conditions, even highly accurate tests produce mostly false positives. Read more in our guide to base rate neglect. **Q: Does the false positive paradox mean medical screening is useless?** No - but it means screening is most valuable when the population being screened has elevated risk (raising the prior probability) and when positive results are followed up with confirmatory testing. Universal screening of low-risk populations often produces more harm than benefit, which is why guidelines have become more selective over time. **Q: How do confirmatory tests fix the problem?** If two tests are independent (different mechanism, different lab) and each has a 1% false positive rate, the chance of both giving a false positive is roughly 1% × 1% = 0.01%. Combining tests dramatically reduces the false positive rate, which is why responsible programmes always use a screening test followed by confirmatory testing. **Q: Why don't doctors learn this in medical school?** They sometimes do, but the famous Harvard study (and many follow-ups) shows the lesson rarely sticks. Practical clinical reasoning about probabilities is hard, and doctors are typically trained on pattern recognition rather than Bayesian updating. This is changing slowly, but the gap is real and consequential. **Q: Where can I learn the underlying maths?** Start with our guide to Bayesian thinking and our overview of probabilistic reasoning. Gerd Gigerenzer's book Reckoning with Risk is excellent for the medical-testing context, and Daniel Kahneman's Thinking, Fast and Slow covers the cognitive biases that make these errors so common. ## The TakeawayThe false positive paradox is one of those rare ideas that, once you understand it, changes how you read the news. Medical screening recommendations, drug testing programmes, security systems, criminal trials, COVID test results, AI fraud detection - they all run into this paradox, and most public discussion of them ignores it.
The core insight is simple: a test result is not the same as a conclusion. To go from result to conclusion, you have to combine the test's accuracy with the prior probability of the condition. Skip that step and you end up scared of vanishingly small risks, missing real ones, or punishing innocent people on the basis of mathematics nobody bothered to do.
If you take one thing from this post, take this: whenever you see a percentage that seems alarming, ask what fraction of the relevant population it's a percentage of. That single habit will protect you from more bad reasoning than any other rule of thumb in probabilistic thinking.
For the deeper machinery behind this, our guides on Bayesian thinking for everyday decisions and base rate neglect build out the toolkit. And for a sister concept that catches people out in different ways, see correlation vs causation.
--- # Sunk Cost Fallacy: When to Quit and When to Persist Source: https://expectedvalue.co.uk/blog/sunk-cost-fallacy/ # Sunk Cost Fallacy: When to Quit and When to Persist The sunk cost fallacy: why we keep investing in losing decisions. The bias, the psychology, and a clear test for when to quit and when to persist. **Answer**The sunk cost fallacy is letting past unrecoverable investments influence forward-looking decisions. Money already spent shouldn't enter your decision about whether to spend more. The rational frame: only marginal future cost and marginal future benefit matter from this point forward.
You bought concert tickets three months ago for £80. The night arrives, and it's pouring with rain. You're exhausted, you've come down with a cold, and honestly you'd rather stay in with a box set. But you go anyway - because you already paid for the tickets.
Sound familiar? You've just been caught by one of the most pervasive cognitive biases in human decision-making: the sunk cost fallacy.
The reasoning feels completely natural. You spent money, so you should "get your money's worth." But here's the thing - that £80 is gone whether you go to the concert or not. The only question that matters is: will you enjoy the next three hours more at the concert or on your sofa? The past expenditure is irrelevant to that decision.
This bias doesn't just affect our Friday night plans. It shapes business strategy, government policy, investment decisions, and some of the most consequential choices in history.
## What Is the Sunk Cost Fallacy?A sunk cost is any expenditure - money, time, effort, emotion - that has already been incurred and cannot be recovered. The sunk cost fallacy is our tendency to continue an endeavour because of what we've already invested, rather than evaluating it purely on its future prospects.
The term entered mainstream economics through Richard Thaler's 1980 paper Toward a Positive Theory of Consumer Choice, and was rigorously demonstrated in a now-classic 1985 study by Hal Arkes and Catherine Blumer (The Psychology of Sunk Cost, Organizational Behavior and Human Decision Processes). Arkes and Blumer showed that people who had paid more for a theatre subscription were significantly more likely to attend shows in bad weather than people who had paid less - even though the ticket price was already gone in both cases.
Rational decision-making demands that we ignore sunk costs entirely. Every decision should be based on expected future outcomes - what economists call the marginal analysis. If you're familiar with expected value thinking, you'll recognise this immediately: the correct question is always "what's the expected value of continuing versus stopping from this point forward?"
But our brains don't work that way. We carry the weight of past investment into every forward-looking decision, and it warps our judgement in predictable ways.
## Why does the sunk-cost trap feel so logical?Three psychological mechanisms drive the sunk cost fallacy:
Daniel Kahneman and Amos Tversky demonstrated that losses hurt roughly twice as much as equivalent gains feel good. When we abandon a project, we crystallise the loss - we have to admit the money is gone. Continuing lets us maintain the comforting fiction that the investment might still pay off. This connects directly to how we misjudge probabilities - we overweight the pain of certain loss against an uncertain future gain.
Once we've publicly committed to a course of action, abandoning it feels like admitting we were wrong. The more we've invested, the harder it becomes to walk away, because the implied admission of error grows proportionally. Psychologists call this escalation of commitment - each additional investment makes the next one more likely, not less. Barry Staw's 1976 paper Knee-Deep in the Big Muddy documented this pattern in managerial decision-making, and it has been replicated countless times since.
We have a deep-seated aversion to "waste." If you quit a book halfway through, those hours feel wasted. If you abandon a renovation, the money spent so far feels thrown away. But this confuses two different things: the resources are gone regardless. The only waste that matters is the future time, money, or effort you'd pour into something that isn't working.
There's also a perception layer to this trap. The more vividly we remember the original investment - the meetings, the late nights, the cheque written - the heavier it weighs on our judgement. This is closely linked to the availability heuristic: vivid, easily recalled details disproportionately shape our decisions, even when they're logically irrelevant. The lavish launch party for a failing product should have no bearing on whether to fund it for another year. Our brains disagree.
## What are the most expensive sunk-cost mistakes in history?Perhaps the most famous example - so famous that economists sometimes call the sunk cost fallacy the "Concorde fallacy." By the mid-1970s, both the British and French governments knew that Concorde would never be commercially viable. The development costs had ballooned far beyond projections, and the market for supersonic passenger travel was a fraction of what had been forecast.
But both governments kept funding the project. Why? Because they'd already spent billions. Walking away would have meant admitting that the entire investment was a loss. So they poured in more money, eventually spending over £1.3 billion (roughly £11 billion in today's money) on an aircraft that never turned a profit.
The rational calculation was straightforward: the money already spent was gone. The only question was whether future spending would yield a positive return. The answer was clearly no - but the sunk costs made it politically impossible to stop.
Robert McNamara, US Secretary of Defence during the Vietnam War, later acknowledged that escalation of commitment played a central role in prolonging the conflict. As casualties mounted, each loss became a reason to continue - "we can't let their sacrifice be in vain" - rather than a signal to reassess the strategy.
By 1967, internal analysis showed the war was likely unwinnable under the current approach. But withdrawal would have meant that the 20,000 American lives lost by that point had been "for nothing." So the commitment deepened, eventually costing over 58,000 American lives and millions of Vietnamese casualties.
Every large company has at least one product that should have been killed years ago. Microsoft's Windows Phone consumed billions over nearly a decade. Google poured resources into Google+ long after it was clear the social network had failed to gain traction.
The pattern is remarkably consistent: a major investment is made, early results disappoint, but instead of cutting losses, leadership authorises more funding. Each round of additional investment raises the emotional stakes of abandonment.
## Where does the sunk-cost fallacy show up in everyday life?The sunk cost fallacy isn't confined to governments and corporations. It shapes deeply personal decisions - and several everyday examples are so universal they appear in almost every textbook on the topic:
Casinos have understood sunk cost psychology for over a century, which is why they replace your money with chips at the door. Once cash becomes coloured plastic, it stops feeling like real money - and any chips you've already lost feel like sunk investment in "the night." Players routinely keep betting after losses they would never have authorised at the start of the evening, chasing what's gone rather than evaluating whether the next bet has positive expected value.
This is why experienced gamblers fix their stake before sitting down and treat anything beyond it as off-limits. They are pre-committing past the point where the sunk cost fallacy can hijack them - a discipline that mirrors the Kelly criterion's emphasis on bet sizing based on current edge, not historical balance.
You're forty minutes into a film. It's clearly a dud. The plot is incoherent, the dialogue is wooden, and you're checking your watch. Yet most people stay to the end - because they paid for the ticket.
The ticket money is gone whether you stay or leave. The only question is whether the next ninety minutes are better spent in the cinema or doing literally anything else. The same logic applies to bad books, half-finished podcasts, and overlong meetings.
At an all-you-can-eat buffet, people routinely eat past the point of comfort because they want to "get their money's worth." The price is already paid. The only question is whether each additional plate makes the evening better or worse - and after the third helping, the answer is reliably worse.
Careers: "I've spent five years building expertise in this field - I can't switch now." Those five years are gone. The question is whether the next five years are better spent here or elsewhere.
Relationships: "We've been together for eight years - I can't just throw that away." The time together has already been spent. The only question is whether the relationship makes your future better or worse.
Education: "I'm two years into this degree and I hate it - but I can't quit now." You absolutely can, and sometimes should. The tuition already paid is irrelevant to whether finishing the degree has positive expected value.
Possessions: That gym membership you never use, the expensive kitchen gadget gathering dust, the half-finished online course. We keep paying, keep storing, keep feeling guilty - all because of what we've already spent.
In each case, the sunk cost creates a powerful emotional anchor that distorts our assessment of future options.
## How does the sunk-cost fallacy hurt investors?The sunk cost fallacy is particularly destructive in investing. You buy a stock at £50. It drops to £30. Rather than evaluating whether the stock is worth buying at £30 right now, you hold on, waiting to "get back to even."
This is irrational - and it is a textbook case of anchoring bias: your purchase price is mathematically irrelevant to the stock's future prospects. The only question is: given what you know today, is this stock a better use of your capital than the alternatives?
It also stacks dangerously with confirmation bias in investing - once you're holding a losing position, you start unconsciously filtering news and analysis, weighting bullish takes more heavily than bearish ones to justify staying in. The combination of sunk-cost reasoning and confirmation bias is one of the main reasons retail investors hold losers far too long.
Think about it using expected value. If the stock has a 40% chance of recovering to £50 and a 60% chance of falling to £20, the expected value of holding is:
EV = (0.4 × £50) + (0.6 × £20) = £20 + £12 = £32
If selling at £30 and reinvesting elsewhere has better expected value, you should sell - regardless of what you originally paid.
Professional traders understand this intuitively. They evaluate positions based on current information and forward-looking probability, not on their entry price. The Kelly criterion, for instance, sizes bets based entirely on the current edge and odds - historical cost doesn't enter the formula.
The amateur investor, by contrast, sorts their portfolio into "winners" (above purchase price) and "losers" (below), selling winners and holding losers. This behaviour, known as the disposition effect (documented by Hersh Shefrin and Meir Statman in 1985), is the sunk cost fallacy in action.
## When is persistence actually rational, not a sunk-cost trap?Understanding the sunk cost fallacy doesn't mean you should abandon everything the moment it gets difficult. Sometimes persistence is exactly the right call. The key is distinguishing between rational persistence and sunk-cost-driven stubbornness.
Persistence is rational when:
New information still supports the original thesis. If you started a business based on a market opportunity, and the opportunity is still real - even if progress is slower than hoped - continuing may be justified. The sunk cost fallacy only applies when you're continuing because of past investment, not despite it.
Switching costs are genuinely high. Sometimes the cost of abandoning and starting fresh exceeds the cost of pushing through. This isn't a sunk cost argument - it's a legitimate forward-looking calculation about transition expenses, lost momentum, and rebuilding time.
You're in a temporary dip, not a structural decline. Seth Godin's The Dip captures this well: most worthwhile pursuits have a difficult middle phase. The question is whether you're experiencing a predictable dip that rewards persistence, or a dead end that never gets better.
The option value of continuing is high. Some endeavours create valuable options even if the primary goal fails. A startup that's struggling might pivot successfully. A degree you're unsure about might open doors you can't foresee.
The critical test: are you evaluating the future based on forward-looking evidence, or are you clinging to the past? A Bayesian thinker updates their beliefs as new evidence arrives. If the evidence keeps pointing the same way - this isn't working - the rational response is to update your beliefs and act accordingly.
## How do you decide when to quit? 1. **Ignore what you've already spent** - Mentally write off all past investment - money, time, emotional energy, reputation. This is the hardest step, but it's non-negotiable. Those resources are gone. You cannot un-spend them by continuing. 2. **Evaluate purely on expected future value** - Based on what you know right now, what's the probability-weighted outcome of continuing? Be honest about the likelihood of success and the magnitude of the potential payoff. Don't let optimism bias inflate your estimates. 3. **Compare against the next best alternative** - Every hour and pound spent continuing is an hour and pound not spent on something else. What's the opportunity cost? The best alternative might be another project, a different investment, or simply freeing up mental bandwidth. 4. **Ask the fresh-start question** - If you were starting from scratch today - with no history, no prior investment, no emotional attachment - would you choose this path? If the answer is no, you're likely being held hostage by sunk costs. This framework works for everything from stock positions to career changes to home renovation projects. The fresh-start question (step 4) is especially powerful because it strips away the emotional weight of prior commitment and forces a clean evaluation. ## What are kill criteria and why should you set them?The best defence against the sunk cost fallacy is to decide your quitting conditions before you begin, while you're still thinking clearly.
Pre-commitment strategies include:
Setting explicit kill criteria. Before launching a project, define the conditions under which you'll stop. "If we haven't reached 1,000 users by month six, we shut it down." "If the stock drops below £25, I sell regardless." These pre-defined triggers bypass the emotional escalation that happens in the moment.
Time-boxing investments. Rather than open-ended commitments, allocate fixed time or budget windows. "I'll give this side project three months and £2,000. At the end, I'll evaluate from scratch."
Keeping a decision journal. Writing down why you committed to something - the evidence, the assumptions, the expected outcome - makes it possible to reassess later without revisionist memory. When the data turns sour, you can compare reality against the original thesis instead of an inflated retroactive version of it. Decision journals are arguably the single most effective sunk-cost antidote, because they force the comparison to be evidence-based rather than emotional.
Scheduling regular reviews. Build in periodic reassessment points where you explicitly ask: "Knowing what I know now, would I start this today?" Calendar these in advance - don't wait until you feel like reassessing, because that feeling comes too late.
Appointing a devil's advocate. Ask a trusted friend, colleague, or adviser to challenge your reasoning. Tell them: "I need you to argue for quitting, and I'll argue for continuing. Let's see which case is stronger." This externalises the conflict and reduces the ego investment.
These approaches work because they separate the decision from the emotion. You're effectively making a pact with your future, rational self to overrule your future, biased self.
## When should you update your bet on a project?The sunk cost fallacy is, at its core, a failure to update beliefs in response to evidence. A Bayesian approach to everyday decisions provides a natural antidote.
When you start a project, you have a prior belief about its chances of success. As time passes, you gather evidence: user numbers, revenue figures, market signals, personal satisfaction. Each piece of evidence should update your belief.
If you started with 70% confidence that a venture would succeed, but six months of data have been consistently disappointing, your updated probability should be significantly lower. The sunk cost fallacy is what happens when you refuse to let that update happen - when you cling to the original 70% because acknowledging the drop means acknowledging the loss.
A useful discipline: after each major milestone or review point, explicitly state your updated probability of success. Write it down. If the number keeps falling, at some point it crosses a threshold where the expected value of continuing becomes negative. That's your signal.
This connects back to a fundamental principle of probabilistic thinking: holding your beliefs with appropriate uncertainty and being willing to change them …[truncated for length] --- # A Probabilistic Framework for Career Decisions Source: https://expectedvalue.co.uk/blog/probabilistic-framework-career-decisions/ # A Probabilistic Framework for Career Decisions Job offers, pivots and salary negotiations are bets under uncertainty. Treat them like bets and you make consistently better calls. **Answer**
Frame career decisions as a tree of probability-weighted outcomes rather than a single bet. List the realistic paths, assign rough probabilities to each based on your evidence, score the expected value of each path on the things you actually care about (money, time, optionality), and pick the one with the highest expected outcome - not the one that feels safest. Revisit annually.
## Should you negotiate? An expected-value calculator Weigh the upside of asking for more against the (usually small) chance the offer is pulled. If the expected value is positive, the maths says ask. Interactive calculator. Computes: Expected value of asking for more; Success chance you would need just to break even.Almost every important career decision is a bet under uncertainty. You take a job not knowing whether the company will thrive, accept a promotion not knowing whether you will enjoy the work, or invest a year learning a new skill not knowing whether the market will reward it. Yet most career advice - "follow your passion", "trust your gut", "go where the opportunity is" - pretends the uncertainty does not exist.
The probabilistic framework on this page borrows directly from the way professional gamblers, traders and forecasters approach decisions. It will not tell you which job to take. What it will do is force you to be honest about the odds, the payoffs and the size of your bet - which is most of what separates good career decisions from bad ones.
The framework has four steps: frame the decision as a bet, estimate the payoffs and probabilities, calculate expected value, and then apply Kelly-style sizing to control downside. Each step exists to defuse a specific failure mode that career decisions usually fall into.
## How do you frame a career decision as a bet?The classic failure is asking "is this the right job for me?" That question has no answer because it presupposes a single objectively correct outcome. The probabilistic version is different: what does the distribution of outcomes look like, and how does that distribution compare to my next-best alternative?
To frame a career decision as a bet, write down four things:
Writing these down explicitly does most of the work. A surprising fraction of career decisions feel hard because the alternative has not been articulated, or because the time horizon has not been bounded, or because the decision-maker has not noticed how easily reversible the bet actually is. Thinking in probabilities covers why this kind of explicit framing beats gut feeling.
## How do you estimate payoffs and probabilities?For each option (the action and the alternative), sketch three scenarios - pessimistic, base case and optimistic - and assign each a rough probability. The numbers do not need to be precise. What matters is that the scenarios are specific and the probabilities add up to 100 percent.
Current job: stable role at a large tech company, £90k base, £15k bonus, predictable hours.
Offer: senior engineer at a Series B startup, £85k base, £25k of equity vesting over 4 years, longer hours.
| Scenario | Probability | 3-year financial outcome | 3-year career outcome |
|---|---|---|---|
| Startup fails / lays off engineering team | 30% | ~£255k earned, equity worthless | Startup experience on CV; need to job-hunt again |
| Base case - company survives, modest growth | 50% | ~£270k earned, equity worth £15–30k | Promoted to staff/principal level; broader skill set |
| Acquisition or strong Series C/D | 20% | ~£270k earned, equity worth £80–200k | Senior role at a larger company post-acquisition; meaningful equity event |
The base case looks roughly comparable to staying. The downside is bounded (the engineer can find another job). The upside is non-trivial. The decision is now visible - and it is a much better decision than "is this startup going to make it?"
If you find yourself unable to articulate three scenarios, that is information: it usually means you have not done enough due diligence on the role yet, not that the decision is impossible.
**Calibration over precision** Beginners worry about whether the probability is 25% or 30%. Almost always it does not matter. What matters is whether your probabilities are roughly calibrated - i.e. when you say 70%, do things actually happen seven out of ten times in your life? If you have never tracked your predictions, your probabilities are probably miscalibrated by a factor of two. This is fixable with practice. ## How do you calculate the expected value of a career move?Expected value is the probability-weighted average of the payoffs. For a career decision, it answers a sharper question than "will this work out?" - namely, what is the average outcome across all the futures I think are plausible?
Continuing the engineer example, looking only at financial outcomes:
On pure financial EV, staying wins by about £20k - but that is before accounting for the career-trajectory upside in the optimistic case (which is where most of the long-tail value sits) and before accounting for the option value of having startup experience on the CV.
This is the point at which most people stop and pick the higher-EV option. That is a mistake. Two adjustments are essential:
The Kelly Criterion captures the first adjustment formally; our Kelly criterion guide covers the maths.
## How do you size the bet on a career decision?The fourth and most overlooked step: even if a career bet is positive-EV, you can still lose badly if you bet too much. Bet size in career terms is roughly: how much of your runway, your reputation, your relationships and your time are you committing irreversibly?
The Kelly Criterion says you should bet a fraction of your bankroll equal to your edge divided by the odds. In career terms, the rough analogue is: commit a fraction of your time, capital and reputation in proportion to how confident you are and how reversible the bet is. A high-conviction, easily-reversible decision (e.g. taking a 3-month contract) deserves a bigger commitment than a low-conviction, irreversible one (e.g. relocating across the world for a job).
## What does this look like in real-life career decisions?Negotiating a salary feels risky because the loss feels vivid ("they might withdraw the offer") and the gain feels abstract ("a few percent more"). The probabilistic framing fixes this. Probability of an offer being withdrawn during a polite, well-researched negotiation: typically under 5%. Probability of a 5–15% improvement: typically 50–70%. EV is overwhelmingly positive. The main reason people fail to negotiate is loss aversion, not bad maths.
Apply the framework: action (transition into PM), alternative (stay technical), horizon (3–5 years), exit options (return to engineering - usually feasible within 18 months of a switch). Estimate scenarios: PM-success base case, PM-bad-fit pessimistic case, PM-thrives optimistic case. Most people who do this exercise discover that the downside is much smaller than they thought, because returning to engineering is genuinely available - but they were treating the decision as one-way.
Bet size: 200–400 hours over a year. Payoffs: a useful but not job-changing skill (base case), a clear job-market premium (optimistic), wasted time (pessimistic). The killer here is usually the sunk cost fallacy at the 100-hour mark - you commit further because you have already invested, rather than because the EV still looks good.
Don't compare offers feature-by-feature. Compute the EV of each one over a 3-year horizon, then ask which one has the better distribution of outcomes. Two offers can have identical expected compensation but very different variance - a fact that disappears completely if you only look at the headline package.
## Which biases distort career thinking?Three biases to watch for, in approximate order of damage caused:
None of these biases disappear when you learn about them. What changes is the cost of acting on them: an unaddressed bias quietly steers years of your life; a named bias is a failure mode you can specifically check for.
## Is there a template you can copy?For your next significant career decision, fill in the following template before deciding:
This template will not tell you the answer. It will, however, surface the parts of your thinking that are usually hidden - and that is the part that turns out to matter most.
## Frequently asked questions **Q: Doesn't expected value over-rationalise an emotional decision?** EV is a tool for surfacing the structure of a decision, not replacing your judgement. The framework explicitly accounts for non-financial outcomes - relationships, autonomy, energy - by asking you to score them as part of each scenario. The point is to make the trade-offs visible rather than to reduce a life decision to a spreadsheet. **Q: How precise do my probability estimates need to be?** Less precise than you think. The exercise of forcing yourself to write down three scenarios that sum to 100% catches most of the value. The probabilities themselves are usually robust to changes of ±10–15% without flipping the decision. **Q: How is this different from a pros-and-cons list?** A pros-and-cons list typically has no probabilities, no time horizon, and no comparison to the alternative. It treats each consideration as if it carries equal weight, which buries the few factors that actually drive the decision. The probabilistic framework forces those factors to the surface. **Q: What if the decision is irreversible?** Almost no career decision is truly irreversible - but some are much more reversible than others. The right adjustment is on bet size, not on whether you do the analysis. For genuinely one-way decisions (relocation with family, doctorate enrolment), the framework simply demands more time spent on Step 4 - narrowing the downside before committing. **Q: Should I track my predictions afterwards?** Yes - this is how you calibrate. Write down the probabilities you assigned at the time and revisit them 12–18 months later. Most people discover their pessimistic scenarios were too pessimistic and their optimistic scenarios assumed too much continuity. Bayesian thinking for everyday decisions covers how to update your priors as new information arrives.If this framework was useful, the natural next reads are expected value explained for the foundation, the Kelly criterion for bet sizing, and the sunk cost fallacy for the bias that hurts career decisions most.
And before you commit, run a pre-mortem on the decision to surface failure paths your optimism is hiding, then put it through second-order thinking - career bets in particular have heavy second-order effects (your reputation, your network, the optionality you keep or burn) that the EV calculation alone won't surface.
--- # Expected Value in Poker: Pot Odds, EV & Decisions Source: https://expectedvalue.co.uk/blog/expected-value-poker/ # Expected Value in Poker: Pot Odds, EV & Decisions Expected value in poker - pot odds, implied odds, EV of bluffs and folds. Worked hand examples showing how the maths translates into better decisions. **Answer** Expected value (EV) in poker is the average outcome of a decision repeated infinitely. A profitable poker decision has positive EV; the goal is to act on +EV opportunities even when individual hands lose. Example: facing a £20 bet to win a £80 pot with a 30% chance of winning. EV = 0.30 × £80 - 0.70 × £20 = £24 - £14 = +£10. Make this call every time. Skilled poker is the discipline of finding +EV spots, sizing bets to extract maximum value when ahead, and folding -EV spots that feel emotionally compelling. Expected value in poker is the single most useful idea a player can carry to the table. It is what separates the winning regular from the losing recreational player who has been studying for ten years and still cannot tell you why a particular call was profitable. The maths underneath is the same expected value framework we cover in our main expected value explained guide; poker is just a domain where the calculations are constant, the consequences are immediate, and the ego cost of being wrong is unusually high. This post walks through the core EV applications in poker: pot odds, implied odds, the EV of calling versus folding, the EV of bluffs, and the EV of value bets. Every section uses concrete numbers from realistic hands, because EV in poker only sticks once you have done the arithmetic yourself a few times. **Prerequisites** This guide assumes basic familiarity with no-limit hold'em - pre-flop, flop, turn, river, and the standard betting actions (check, bet, call, raise, fold). If terms like 'pot' and 'big blind' are new, work through a beginner's no-limit primer first; the concepts here will land much harder once the structure is familiar. ## What does expected value mean at the poker table?Expected value at the table is the average outcome of a decision if you could repeat the same situation an infinite number of times. The formula is identical to the general one:
EV = (Probability of winning × Amount won) − (Probability of losing × Amount lost)
In poker, the "amounts" are chips and the "probabilities" come from the cards: the cards you see, the cards your opponent could have, and the cards still to come. Every meaningful decision in a poker hand - call, fold, bet, raise - has an EV. The decision with the highest EV is, by definition, the correct one.
The word "correct" here is unusually load-bearing. A correct (positive-EV) decision can lose money on the night. An incorrect (negative-EV) decision can win the pot. Variance - luck - masks the EV of any individual hand. The skill in poker is not seeing the future; it is making the highest-EV decision available, hand after hand, and trusting the maths to play out across enough hands.
The rest of this guide is the toolbox you use to compute EV in the situations that come up most often.
## What are pot odds and why are they the cheapest EV calc?Pot odds are the ratio of the cost of a call to the size of the pot you are calling into. They are the simplest possible EV calculation and the one most beginners learn first.
The setup: the pot is £100. Your opponent bets £50, making the pot £150. You have to put in £50 to win £150 (£100 already in the pot, plus their £50 bet). Your pot odds are 50:150, or 1:3 - for every £1 you risk, you stand to win £3.
To break even on a call, you need to win at least 25% of the time. The maths:
If your hand wins more than 25% of the time, the call is profitable. If it wins less than 25% of the time, the call is unprofitable.
The quick way to think about pot odds at the table:
| Pot odds | Bet to call | Pot before call | Break-even win rate |
|---|---|---|---|
| 1:1 | £50 | £50 | 50% |
| 1:2 | £50 | £100 | 33% |
| 1:3 | £50 | £150 | 25% |
| 1:4 | £50 | £200 | 20% |
| 1:5 | £50 | £250 | 17% |
The bigger the pot relative to the bet, the smaller the win rate you need to make calling profitable.
This is where common drawing situations get their famous heuristics. A flush draw on the flop has roughly nine outs (cards that complete it). The rule of "4 and 2" estimates your odds: nine outs × 4 ≈ 36% by the river, nine outs × 2 ≈ 18% by the turn. So with a flush draw on the flop, facing a half-pot bet (which gives you 1:3 pot odds, requiring 25% equity), the call is profitable by a comfortable margin if there is no further bet on the turn - exactly when you should consider implied odds.
## What are implied odds in poker?Pot odds only count the money currently in the pot. Implied odds add the money you can reasonably expect to win on later streets if you hit your hand.
The setup: you have an open-ended straight draw on the flop (8 outs, roughly 32% by the river). The pot is £20, your opponent bets £20. You need to call £20 to win £40, which gives you 1:2 pot odds and a 33% break-even point. Your raw equity is just below this threshold, so on a pure pot-odds basis the call is marginal-to-bad.
But your opponent has £200 left in their stack. If you hit your straight on the turn or river, they are likely to pay off another £40-60 with a good top pair. That £40-60 is implied money - money you will win, on average, when you hit, that the pot odds calculation does not currently see.
Factor it in:
With implied odds, the marginal call becomes clearly profitable.
Implied odds are estimates, not guarantees. They depend on:
Good players adjust implied odds upward against loose, station-y opponents and downward against tight, fit-or-fold ones. Badly applied, implied odds become a self-justifying excuse to call too wide; well applied, they capture real expected money the pot odds do not.
## What are reverse implied odds?Reverse implied odds are the mirror of implied odds: the money you will lose on later streets when you make a hand that is good enough to call with but not good enough to win the pot.
The classic case is a weak top pair on a coordinated board. You call the flop with top pair, weak kicker; the turn brings a possible flush; you call again because "the pot is too big to fold"; the river brings a possible straight; you call again. By showdown you have invested far more than the original pot odds suggested, and you are losing to your opponent's better top pair, made flush, or straight.
Reverse implied odds are why drawing to dominated hands (low pairs, weak suited connectors against a raise) is such a slow leak. The hand looks like it has equity. It does. But every time the equity "realises" into a made hand, you are about to lose more money - sometimes much more - to a hand that has you crushed.
The practical rules:
Let's walk through a complete EV calculation for a meaningful river decision.
The situation. You are heads-up on the river. You hold A♠ Q♠ on a board of K♠ 7♠ 4♥ 2♣ J♦. Your hand is ace-high - likely no good against any pair, but you have ace-high blocker value. The pot is £200. Your opponent bets £100 (a half-pot bet). You have to decide whether to call £100 to win £300.
Pot odds: 100:300, or 1:3. You need to win 25% of the time to break even.
Hand reading. From their pre-flop call, flop check-call, turn check-call, river bet, you assign them a range. Reasonable assumptions:
Your estimated equity against their range:
Total equity ≈ 25%.
EV of calling:
Break even - the call is not profitable, but not a leak either.
EV of folding: £0 by definition. Folding closes the action with no further loss.
In this case, calling and folding are equivalent in expected value. Most experienced players would lean toward the fold, on the principle that the assumed bluff frequency (25%) is generous and the cost of being wrong about it is greater than the cost of folding marginally too tight.
The point of the exercise is not the specific numbers - they will always be estimates. The point is that the decision becomes structured: estimate the opponent's range, estimate your equity against it, compare the EV of calling against zero. Once you have done this calculation a hundred times, you start doing it instinctively at the table in seconds.
## What is the EV of a bluff?Bluffing flips the EV calculation. You have a hand that cannot win at showdown. Your only path to winning the pot is to make your opponent fold.
The framework:
Bluff EV = (Fold % × Pot won) − (Call % × Bluff cost)
The setup. The pot is £100. You bet £75 as a bluff. The required fold frequency to break even:
If your opponent folds more than 43% of the time, the bluff is profitable. If they fold less than 43%, it is unprofitable.
The size of the bluff matters a lot:
| Bluff size | Pot before bluff | Required fold % to break even |
|---|---|---|
| 1/3 pot | £100 | 25% |
| 1/2 pot | £100 | 33% |
| 2/3 pot | £100 | 40% |
| Pot | £100 | 50% |
| 1.5x pot | £100 | 60% |
| 2x pot | £100 | 67% |
Smaller bluffs need fewer folds to be profitable. Bigger bluffs need more folds but win more when they succeed. Sizing your bluffs based on the actual fold frequency you expect is one of the highest-leverage applications of EV thinking in poker.
Where bluff EV most often goes wrong:
Value betting is the everyday positive-EV action in a winning poker game. You have a hand likely to be best at showdown, your opponent has a worse hand they will pay off with, you bet, they call, you win more than they wanted to give you.
The framework:
Value bet EV = (Call % × Bet) − (Fold % × 0)
Unlike bluffs, the EV of a value bet is unambiguously positive whenever your hand has more equity than your opponent calls with. The questions are sizing and frequency, not whether to bet.
Sizing the value bet:
The 'thin' value bet - betting hands like top pair weak kicker on a draw-heavy river - is where many decent players leak EV. The bet is correct against opponents who pay off thinly; it is wrong against opponents who only call with bigger pairs and better. Adapting bet size and frequency to opponent type is the practical work.
## Does EV work the same in tournaments and cash games?Almost everything above applies cleanly to cash games, where every chip has the same monetary value as every other chip. Tournament EV is more complex because the relationship between chips and money is non-linear.
In a tournament, doubling your stack does not double your equity in the prize pool. Late-game pay jumps mean a chip won is worth less than a chip lost - losing a flip near the bubble can knock you out for nothing, while winning the same flip only nudges you up the leaderboard. This is captured by the Independent Chip Model (ICM), which converts chip stacks into expected dollar/pound equity given the prize structure.
The practical consequence: marginal cash-game spots that are clearly +cEV (chip EV) become −$EV (dollar EV) under ICM pressure. The classic example is a flip for stacks on the bubble - calling all-in for 50% equity is +cEV (you risk x chips to win x chips, a wash in chip terms) but −$EV (you risk a likely cash for a small chance of doubling up).
For most players, the rule is simple: cash-game EV thinking applies in early-stage tournaments and deep-stack situations. ICM-aware EV thinking takes over in the late stages, on the bubble, and at final tables. If you are serious about tournament play, learn the basics of ICM after you are comfortable with cash-game EV.
## Which biases distort EV thinking at the poker table?EV in poker is mathematically clean. The reason humans struggle to apply it is that several well-documented cognitive biases conspire to make us deviate from EV-optimal decisions.
The gambler's fallacy - the belief that random outcomes are "due" to balance out - leads players to chase losses ("I have lost five flips, the next one has to be mine") or play scared after a hot streak. Each hand is independent; previous results do not change current EV.
The sunk cost fallacy - the urge to keep investing because you have already invested - explains why so many players call rivers that the pot odds clearly say to fold. The chips already in the pot are not yours; the only question is the EV of the call you are about to make.
Loss aversion - feeling losses about twice as strongly as equivalent gains - makes players fold profitable calls because the £100 they might lose hurts more than the £200 they might win helps. EV does not weight losses more than gains; humans do.
Hindsight bias - the feeling that an outcome was "obvious" once it happens - leads players to second-guess EV-correct decisions that lost money. The river card you should have folded to was unknowable before it landed; the call was either +EV or −EV based on the information available at the time.
A healthy poker EV mindset is partly arithmetical (do the calculation) and partly behavioural (notice when these biases are pulling you off the maths).
## How does EV sizing connect to bankroll management?An individually +EV decision can still bust your bankroll if you size it too large relative to your total stack. This is the same insight behind the Kelly criterion for bet sizing - knowing your edge tells you the right direction; it does not tell you how much to risk.
For poker specifically:
If you find yourself making EV-correct decisions but going broke anyway, the problem is bankroll sizing, not the decisions themselves. Conversely, if you have an enormous bankroll but lose money long-term, the problem is your EV calculations.
## What are the pros and cons of strict EV-based play? **Pros** - Decisions become structured - every hand reduces to a small set of estimable numbers (pot odds, equity, opponent range) - Tilt becomes harder - when your standard is 'did I make the +EV play?' rather than 'did I win?', a bad beat is just variance - Bluff and value-bet sizing become principled rather than feel-based - You stop chasing draws that do not have the pot odds and stop folding draws that do - You start identifying −EV plays others make at the table - the source of most of your profit comes from their leaks, not your wizardry - Long-run results align more closely with skill …[truncated for length] --- # Position Sizing: How Much to Bet When You Have an Edge Source: https://expectedvalue.co.uk/blog/position-sizing-kelly-criterion/ # Position Sizing: How Much to Bet When You Have an Edge Knowing you have an edge is half the battle - sizing positions correctly is the other half. How to size bets so you don't blow up or waste your advantage. **Answer**The Kelly criterion is a mathematical formula for optimal bet (or trade) sizing: stake_fraction = (edge / odds), where edge is your expected return per unit risked and odds are the payoff ratio. For a 60% chance of winning even money, full Kelly = 0.6 - 0.4 = 20% of bankroll. In practice almost no professional uses full Kelly: edge estimates carry error, and the variance is brutal. Fractional Kelly (typically half or quarter Kelly) sacrifices a small amount of expected log-growth for a much smoother equity curve and survives misestimation.
Most investing advice focuses on what to buy. Very little focuses on how much to buy - and that's a problem, because position sizing is often the difference between a profitable strategy and a catastrophic one.
You can have the best stock picks in the world and still go broke if you bet too aggressively on each one. Conversely, you can have a modest edge and compound it into serious wealth if you size your positions intelligently.
This guide connects the theory from our Kelly Criterion article to practical portfolio decisions. We'll work through real examples with stocks, crypto, and sports betting - and explain why most people systematically overbet.
## Why Position Sizing Matters More Than Stock Picking Imagine two investors. Both identify the same undervalued stock with a genuine 60% chance of returning 50% and a 40% chance of losing 30%. Investor A puts 80% of their portfolio into it. Investor B puts 15%. The expected value of the trade is identical for both: (0.6 × 50%) + (0.4 × -30%) = +18%. That's a great bet. But watch what happens over multiple similar bets. Investor A, betting 80% each time, will eventually hit a losing streak. Three losses in a row takes their portfolio from £10,000 to £10,000 × 0.76³ = £4,390. They've lost 56% of their capital and now need a 127% gain just to get back to even. That's the kind of hole most people never climb out of. Investor B, betting 15% each time, survives the same losing streak with £10,000 × (1 - 0.15 × 0.30)³ = £8,715. A 13% dip. Uncomfortable, but recoverable. The edge was the same. The only difference was position sizing. **The asymmetry of losses** A 50% loss requires a 100% gain to recover. A 75% loss requires a 300% gain. This asymmetry is the fundamental reason why position sizing matters - large losses are exponentially harder to recover from than small ones. ## Full Kelly vs Fractional KellyThe Kelly Criterion gives us a mathematically optimal bet size. For a simple binary outcome:
Kelly fraction = (bp - q) / b
Where:
Using our stock example: b = 50/30 = 1.667, p = 0.6, q = 0.4
Kelly = (1.667 × 0.6 - 0.4) / 1.667 = (1.0 - 0.4) / 1.667 = 36%
So the Kelly Criterion says to put 36% of your portfolio into this trade. That's the bet size that maximises long-term compound growth.
But here's the catch: full Kelly is extremely volatile. While it maximises the geometric growth rate, the ride is brutal. Drawdowns of 50-80% are normal under full Kelly - mathematically expected, in fact.
This is where fractional Kelly comes in. Most professional bettors and quantitative investors use somewhere between one-quarter and one-half Kelly:
In our example, half Kelly = 18% and quarter Kelly = 9%. Both are profitable. The question is how much volatility you can stomach.
## Worked Example: A Stock PortfolioLet's say you have a £50,000 portfolio and you've identified three opportunities:
Stock A - Undervalued retailer
Stock B - Biotech with pending trial results
Stock C - Dividend play with catalyst
Total allocated: £25,313 (50.6% of portfolio). The remaining 49.4% stays in cash or index funds - and that's fine. Kelly never says you have to be fully invested.
**The cash buffer is a feature, not a bug** Having 30-50% of your portfolio in cash when using Kelly sizing isn't being timid - it's being optimal. That cash is dry powder for the next opportunity. Full Kelly across a diversified portfolio naturally leaves a cash reserve. ## Worked Example: CryptoCrypto markets are particularly instructive because the asymmetry between gains and losses is extreme.
Say you believe Bitcoin has a 40% chance of doubling in the next year and a 60% chance of dropping 40%.
Kelly fraction: (2.5 × 0.4 - 0.6) / 2.5 = 16%
Half Kelly: 8% of your portfolio.
That might feel low if you're bullish on crypto. But consider: this allocation survives Bitcoin going to zero five times over before you're wiped out. Meanwhile, someone who put 50% of their portfolio into Bitcoin at the same odds is facing potential ruin.
The lesson: high-conviction, high-volatility assets demand smaller position sizes, not larger ones. The higher the potential loss, the more conservative your sizing needs to be.
## Risk of Ruin: Why Overbetting KillsRisk of ruin is the probability that you'll lose enough capital to effectively end your investing career (or at least set it back by years).
For a simple model, the risk of ruin when betting fraction f of your bankroll with probability p of winning is approximately:
Risk of ruin ≈ ((1-p)/p)^(bankroll/unit_bet)
The key insight: risk of ruin drops exponentially as you decrease bet size. Going from 40% of your portfolio per bet to 20% doesn't halve your risk - it might reduce it by a factor of ten or more.
This is why professional poker players, sports bettors, and quant funds are so disciplined about sizing. A 2% edge with 1% position sizes has essentially zero risk of ruin over any reasonable timeframe. The same 2% edge with 25% position sizes has a meaningful probability of blowing up.
Rules of thumb for avoiding ruin:
Overbetting is the default human tendency, and there are several psychological reasons:
Overconfidence in edge estimation. If you think your probability is 70% but it's actually 55%, full Kelly on your wrong estimate could be dramatically larger than optimal. Since we're almost always overconfident in our estimates, fractional Kelly provides a crucial safety margin.
Ignoring correlation. Kelly assumes each bet is independent. In reality, your five tech stock picks probably move together. If the market drops, they all drop. You need to account for correlation by reducing total exposure.
Outcome bias. After a winning streak, it's tempting to increase position sizes. But a winning streak doesn't mean your edge has grown - it might mean you've been lucky, and the variance is about to swing the other way.
Social proof. When you see someone on social media bragging about putting their entire portfolio into one stock that doubled, you don't see the thousands of people who tried the same thing and lost 80%. Survivorship bias makes reckless sizing look smart.
## A Practical FrameworkHere's a step-by-step approach to sizing any position:
Estimate your edge honestly. What's the probability distribution of outcomes? Be conservative - if you think there's a 70% chance of profit, model it as 60%.
Calculate Kelly. Use the formula above, or for continuous outcomes, estimate the expected return and variance.
Apply a fractional multiplier. Use half Kelly if you're confident in your edge estimate, quarter Kelly if you're not.
Cap at 5% of portfolio. Regardless of what Kelly says, no single position should exceed 5% of your total portfolio unless you have extreme conviction and limited downside.
Check correlation. If you already have three tech stocks, your fourth tech position should be sized as if the combined tech exposure is one position.
Re-evaluate after large moves. If a position doubles, it now represents a larger share of your portfolio. Consider trimming back to your target size.
This framework won't maximise your returns in any single year. But it will keep you in the game long enough for compounding to work - and that's the whole point.
## The Bottom LinePosition sizing is the unglamorous skill that separates professionals from amateurs. It's not exciting, it won't make for good dinner party conversation, and it'll often feel like you're leaving money on the table.
But over a 20 or 30-year investing career, the person who sizes their bets correctly will almost certainly end up wealthier than the person who picks better stocks but bets recklessly.
The maths is clear: the optimal bet is almost always smaller than you think. When in doubt, bet less.
For the theoretical foundation behind all of this, read our guide to the Kelly Criterion. The same fractional-Kelly logic applies anywhere you have an edge against a probabilistic counterparty - see our guide to how prediction markets work for a worked example on event-contract pricing. And if you're interested in the cognitive biases that make us systematically overbet, our piece on thinking in probabilities covers the psychological side.
--- # Loss Aversion: Why Losses Hurt Twice as Much Source: https://expectedvalue.co.uk/blog/loss-aversion/ # Loss Aversion: Why Losses Hurt Twice as Much The cognitive bias that makes us irrationally cautious - and how to counter it. **Answer** Loss aversion is the cognitive pattern, identified by Kahneman + Tversky in 1979, where losses feel roughly twice as painful as equivalent gains feel pleasurable. A loss of £100 hurts about as much as a gain of £200 feels good. The effect is robust across cultures and decision domains and is one of the foundations of prospect theory. Practical implication: investors hold losing positions too long (selling locks in the loss-pain) and exit winners too early; nudging defaults that frame outcomes as losses rather than foregone gains shifts behaviour reliably.Imagine someone offers you a coin flip. Heads, you win £150. Tails, you lose £100. The expected value is +£25 - a clearly positive bet that you should take every time it's offered. Most people refuse it.
This is loss aversion: the deep-seated psychological tendency to feel losses more intensely than equivalent gains. Daniel Kahneman and Amos Tversky's research suggests the asymmetry is roughly 2:1 - losing £100 hurts about twice as much as winning £100 feels good. The result is a steady drip of decisions that look conservative on the surface but quietly destroy expected value.
This article unpacks where loss aversion comes from, the specific ways it warps decisions in investing, negotiations, and consumer behaviour, and the practical strategies that genuinely help you push back against it.
## What Is Loss Aversion?Loss aversion is a key plank of prospect theory, the model Kahneman and Tversky introduced in 1979 to explain how people actually evaluate risky choices - as opposed to the rational, expected-utility-maximising agent assumed by classical economics.
The model has two big departures from textbook rationality:
Two related effects fall out of the same machinery. The endowment effect means we value things we already own more highly than identical things we don't - because giving them up registers as a loss. Status quo bias means we default to the current state of the world even when change would be a clear improvement, because any change creates the possibility of loss.
**The 2:1 ratio** Kahneman and Tversky's original estimate of the loss-aversion coefficient was around 2.25 - losses weigh roughly 2.25× more than equivalent gains. Later replications produce a range from about 1.5 to 2.5 depending on the population and the size of the stakes. The exact number matters less than the direction: losses are systematically over-weighted. ## Loss Aversion in Investing: The Disposition EffectThe cleanest financial fingerprint of loss aversion is the disposition effect: investors sell winners too early and hold losers too long. The mechanism is straightforward. Selling a winner locks in a gain, which feels good. Selling a loser realises a loss, which feels terrible - even though the realised and unrealised losses are mathematically identical at that moment. So we postpone the pain by refusing to sell.
Terrance Odean's well-known 1998 study of US brokerage accounts found exactly this pattern: winners were realised at a meaningfully higher rate than losers, despite winners going on to outperform on average after sale. Holding losers wasn't a clever value play - it was the loss-aversion reflex masquerading as patience.
The same bias shows up in subtler forms. Investors check their portfolios more often in flat or rising markets and less often in falling ones, because seeing a loss is painful. They also under-rebalance - moving money out of losing positions back to target weightings requires acknowledging the loss, so it gets postponed indefinitely.
For a more rational frame on when to size up or down, see our guide to position sizing and the Kelly criterion, which treats the question purely in terms of edge and bankroll rather than feelings about a single trade.
## Loss Aversion in Negotiations and Workplace DecisionsOutside markets, loss aversion shapes everyday choices that don't involve a stock price at all.
Marketers have understood loss aversion for decades, even before they had a name for it. Some of the most common applications:
None of this is sinister on its own - loss aversion is a real preference, and accommodating it isn't manipulation. But recognising the pattern lets you decide which ones you want to play along with.
## How to Counteract Loss Aversion 1. **Reframe in terms of total wealth, not changes** - Prospect theory predicts loss aversion only when outcomes are evaluated as gains and losses from a reference point. If you anchor on your total net worth instead, a £1,000 swing on a single position becomes a small percentage move - emotionally much easier to evaluate dispassionately. 2. **Apply expected value thinking explicitly** - Calculate the EV of the decision in front of you and let the number lead. The +EV coin flip from the intro is mathematically a clear take. See our guide to expected value to build the habit of computing it before acting. 3. **Make the decision once, not many times** - Loss aversion is much weaker for decisions made in aggregate. If you commit in advance to taking 100 +EV bets of similar size, you'll happily accept the portfolio even if you'd refuse any single bet in isolation. This is the principle behind systematic, rules-based investing - and a major reason index funds outperform hand-picked portfolios for most investors. 4. **Use pre-commitment devices** - Set rules that fire automatically: stop-losses, rebalancing schedules, automatic transfers, dollar-cost averaging. The rule does the painful action so the loss-averse part of you doesn't have to. The single most effective antidote to the disposition effect is a pre-set selling rule. 5. **Track opportunity costs alongside realised costs** - Loss aversion thrives when one side of the ledger is invisible. Holding a losing position is a real cost - the cash is tied up, accruing nothing - but it doesn't show up on a P&L. Forcing yourself to compare 'hold' against 'sell and redeploy' on equal terms removes the asymmetry. ## When Loss Aversion Is Actually UsefulIt's worth pushing back on a one-sided narrative. Loss aversion is not a defect that needs to be eliminated - it's an evolved heuristic that works well in environments where downside outcomes are catastrophic and upside outcomes are bounded. Refusing a 50/50 bet for half your net worth is loss-averse, but it's also correct: the utility curve for real wealth is concave enough that the bet is genuinely bad, even at fair odds.
The bias becomes a problem in modern, repeated, low-stakes financial decisions where outcomes are roughly symmetric and you'll face thousands of similar choices over a lifetime. A useful rule of thumb: if the loss would be survivable and the decision will repeat, override loss aversion with expected value. If the loss would be ruinous or the decision is one-shot, listening to loss aversion is probably wise.
## Frequently Asked Questions **Q: Is loss aversion the same as risk aversion?** No. Risk aversion is the preference for a certain outcome over an uncertain one with the same expected value - it can be perfectly rational. Loss aversion is the asymmetric over-weighting of losses versus gains around a reference point, and it produces decisions that violate basic expected-utility theory. **Q: Does loss aversion go away with experience?** Partially. Professional traders, market makers, and poker players show smaller loss-aversion coefficients than the general population - but they don't eliminate it entirely. Process and pre-commitment do more of the work than willpower. **Q: How does loss aversion relate to the sunk cost fallacy?** They're close cousins. Sunk cost fallacy is partly explained by loss aversion: walking away converts an unrealised loss into a realised one. Read more in our deep dive on the sunk cost fallacy. **Q: Is the 2:1 ratio universal?** It's a population average, not a constant. Replications find values from roughly 1.5x to 2.5x depending on the demographic, the framing, and the size of the stakes. Some people, in some domains, show very little loss aversion; a few even show loss-seeking behaviour. **Q: What's the single most useful debiasing technique?** For most people: pre-commitment. Decide your rule in advance, when you're calm, and let the rule make the painful decisions. Stop-losses, automatic rebalancing, and rules-based investing all exploit this principle. **Bottom line** Loss aversion is the bias that makes a £100 loss hurt about twice as much as a £100 gain feels good. It's behind the disposition effect, sunk-cost commitments, and most successful marketing tactics. The most reliable counter is process: expected-value calculations, pre-commitment rules, and aggregating decisions rather than evaluating them one at a time. --- # Thinking in Probabilities: Why Your Brain Is Bad at Risk Source: https://expectedvalue.co.uk/blog/thinking-in-probabilities/ # Thinking in Probabilities: Why Your Brain Is Bad at Risk Your brain systematically misjudges probability. Learn the cognitive biases that distort risk perception and how to calibrate better. ## Why is your brain so bad at probability? **Answer** Probabilistic thinking is the habit of replacing 'will X happen?' with 'what are the odds X happens?' It forces you to confront uncertainty quantitatively rather than dodging it. Used well, it improves decisions in poker, investing, medical diagnosis, and any domain where outcomes aren't deterministic.Humans evolved to survive on the African savannah, not to calculate conditional probabilities. Our ancestors needed fast, decisive responses to immediate threats - not careful Bayesian reasoning about base rates. The result is a brain that's remarkably capable in many domains but consistently terrible at judging likelihood.
This isn't a character flaw; it's a design feature. When a rustle in the grass might be a lion, the cost of a false positive (running away from nothing) is trivial compared to the cost of a false negative (being eaten). Evolution optimised us for survival-relevant decisions, not mathematically correct ones.
The problem is that modern life demands probabilistic thinking. Investing, medical decisions, business strategy, career planning - all require accurate assessment of uncertain outcomes. But not all uncertainty is alike - getting the distinction between risk and uncertainty right determines which tools actually apply. Getting probability right is the foundation of expected value thinking, which underpins rational decision-making. And our built-in mental shortcuts (heuristics) that served us well in ancestral environments now produce systematic errors (biases) in these domains.
Let's examine the most damaging of these biases and what you can do about them.
A closely related foundation is fluency with the language of probability - the difference between a 25% chance and 3-to-1 odds, and how decimal, fractional and American odds all encode the same number. Our guide to probability vs odds covers the conversions and the common mistakes.
## The Availability HeuristicThe availability heuristic makes us judge the probability of events by how easily examples come to mind. If something is vivid, recent, or emotionally charged, we overestimate its likelihood.
Example: After watching news coverage of a plane crash, people dramatically overestimate the probability of dying in a flight - even though the annual risk is roughly 1 in 11 million for commercial aviation. Meanwhile, the drive to the airport (approximately 1 in 5,000 annual risk of a fatal car accident) feels completely safe because car crashes rarely make national news.
The numbers: In the UK, you're about 2,200 times more likely to die in a car accident than in a plane crash per year of travel. Yet aviation anxiety is common while driving anxiety is rare.
How it distorts decisions:
Corrective: When you notice yourself thinking "that seems likely," ask: "Am I judging probability, or am I judging how easily I can picture this happening?" They're not the same thing. Look up the actual base rate.
## Base Rate NeglectBase rate neglect occurs when people focus on specific evidence about an individual case while ignoring how common the outcome is in general.
The classic demonstration is the taxi problem (Kahneman & Tversky, 1972):
A city has 85% green taxis and 15% blue taxis. A witness to a hit-and-run says the taxi was blue. The witness correctly identifies colours 80% of the time. What's the probability the taxi was actually blue?
Most people say about 80% - anchoring on the witness's accuracy. The correct answer, via Bayes' theorem, is 41%.
Here's why:
The base rate (85% green) is so dominant that even a fairly reliable witness is more likely to be wrong than right in this scenario. People find this deeply counterintuitive because we instinctively anchor on the diagnostic evidence (witness testimony) and neglect the prior probability.
Real-world impact: This is why medical screening tests produce so many false positives for rare diseases - a phenomenon known as the false positive paradox. If a disease affects 1 in 1,000 people and a test is 95% accurate, a positive result still means you probably don't have the disease (the false positive rate overwhelms the base rate).
## The Gambler's Fallacy and the Hot HandThe Gambler's Fallacy is the belief that past random outcomes influence future ones. "Red has come up 8 times in a row - black must be due!" The roulette wheel has no memory. Each spin is independent. P(red) is always 18/37 regardless of history.
Why does this feel so wrong? Because we have a deeply ingrained sense that the universe is "fair" and should balance out. We expect random sequences to look random to us - alternating frequently between outcomes. But true random sequences contain far more streaks than our intuition expects.
Try this: Write down what you think 20 fair coin flips look like, then actually flip 20 times. Your invented sequence will almost certainly have shorter streaks than the real one. In 20 flips, there's a 50% chance of a run of 5 or more identical outcomes. Most people would look at HHHHHTTHTTH and insist the coin is biased.
The Hot Hand is related but subtly different. For years, psychologists claimed that the "hot hand" in basketball was a fallacy - players aren't actually more likely to make a shot after making several in a row. Recent research (Miller & Sanjurjo, 2018) has shown this was wrong: there is a small hot hand effect in basketball and other skilled activities.
The key distinction: in games of pure chance (roulette, lottery), the gambler's fallacy is always wrong - past outcomes are irrelevant. In games involving skill (basketball, poker), momentum effects can be real but are typically smaller than people believe.
## Conjunction Fallacy and Narrative BiasTversky and Kahneman's famous "Linda problem":
Linda is 31, single, outspoken, and very bright. She studied philosophy and was deeply concerned with issues of discrimination and social justice as a student.
Which is more likely? A) Linda is a bank teller. B) Linda is a bank teller AND is active in the feminist movement.
About 85% of respondents choose B. This is logically impossible - the conjunction of two events can never be more probable than either event alone. "Bank teller AND feminist" is a subset of "bank teller," so it must be less likely.
People choose B because it fits a coherent narrative. The description of Linda makes "feminist bank teller" a more representative story, and we confuse representativeness with probability.
Why this matters for decisions:
When asked to give 90% confidence intervals (ranges they're 90% sure contain the true answer), most people create intervals that contain the truth only 50-60% of the time. We are dramatically, consistently overconfident.
This isn't restricted to laypeople:
The planning fallacy is a special case of overconfidence: we consistently underestimate how long things will take and how much they'll cost. The Sydney Opera House was estimated at $7 million and 4 years; it cost $102 million and took 16 years. This isn't an outlier - it's the norm for large projects.
Corrective strategies:
The good news: probabilistic thinking is a learnable skill. Research on "superforecasters" (Tetlock, 2015) shows that certain practices dramatically improve predictive accuracy:
1. Think in percentages, not words
Replace vague language with numbers. "Likely" means different things to different people (experiments find it ranges from 55% to 90% depending on the person). Say "I'd put this at 70%" instead. The precision forces clarity. This kind of quantitative thinking is also what makes expected value calculations so powerful - they force you to assign numbers to outcomes instead of relying on vague impressions.
2. Update incrementally
When new evidence arrives, don't swing from one extreme to another. Move your probability estimate by an amount proportional to the strength of the evidence. Start from your prior, apply the evidence, arrive at a posterior. This is Bayesian thinking in practice.
3. Keep a prediction journal
Write down your predictions with probability estimates and dates by which you'll know the answer. After enough predictions, you can check: when you said 70%, did the thing happen roughly 70% of the time? If you're poorly calibrated, you can adjust. A more structured way to do the same thing is to forecast on real markets - see how prediction markets work for a tour of the platforms (Metaculus, Manifold, Polymarket) where you can score yourself against thousands of other forecasters.
4. Seek disconfirming evidence
Our natural tendency is to seek confirmation for what we already believe. Actively look for reasons you might be wrong. What would change your mind? If nothing could, your belief isn't based on evidence.
5. Disaggregate complex questions
Instead of estimating one big uncertain thing, break it into components. "Will this startup succeed?" becomes: "Will they build the product? (80%) Will customers want it? (40%) Will they outcompete alternatives? (30%) Will they raise enough capital? (60%)" The product of these (5.8%) is probably more accurate than whatever single number you'd have guessed.
6. Use the outside view first, then adjust
Before applying specific knowledge about your situation, ask: "What usually happens in cases like this?" Base rates first, then specific adjustment. This single habit can eliminate many of the biases discussed in this article.
**Q: Can cognitive biases ever be helpful?** Yes. Heuristics exist because they're fast and usually good enough. The availability heuristic helps you quickly avoid genuinely common dangers. Overconfidence motivates action that would never happen under perfectly calibrated uncertainty. The key is knowing when your fast thinking is helping versus hurting - and switching to slow, deliberate probabilistic reasoning for important decisions. **Q: How long does it take to become well-calibrated?** A range of studies suggest that with deliberate practice (making predictions, tracking outcomes, reviewing your accuracy), meaningful improvement happens within a few months. Superforecasters in Tetlock's research weren't born with special abilities - they developed calibration through consistent practice and willingness to update their beliefs. **Q: Are some people naturally better at probabilistic thinking?** There's some natural variation, but the biggest differentiator is mindset rather than innate ability. People who score well on calibration tests tend to be actively open-minded, numerate, and willing to change their minds. These are learnable dispositions. Intelligence helps, but intellectual humility matters more. **Q: What resources can help me improve my probabilistic reasoning?** Start with 'Thinking, Fast and Slow' by Daniel Kahneman for understanding biases, then 'Superforecasting' by Philip Tetlock for practical calibration techniques. For practice, try Metaculus or Good Judgment Open - platforms where you can make predictions and track your accuracy over time against real outcomes. **Q: Does knowing about these biases make you immune to them?** Unfortunately, no. Knowledge of biases provides some protection but doesn't eliminate them. Even Kahneman himself admitted he remained susceptible to the biases he'd spent decades studying. The most effective defences are structural: checklists, prediction tracking, pre-mortems, and decision frameworks that force you through debiasing steps regardless of how you feel in the moment. --- # Anchoring Bias: How First Numbers Hijack Judgement Source: https://expectedvalue.co.uk/blog/anchoring-bias/ # Anchoring Bias: How First Numbers Hijack Judgement A single irrelevant number can drag your estimate by 30% or more. Why your brain anchors on the first figure it sees - and how to stop it. **Answer** Anchoring bias is the tendency for an initial reference number to disproportionately influence later judgments, even when that number is irrelevant. Classic Tversky + Kahneman demonstration: participants spun a wheel marked 10 or 65, then estimated the percentage of African countries in the UN. Those who saw 10 estimated 25%; those who saw 65 estimated 45%. Practical implication: opening offers in negotiations set the anchor and matter more than they should; asking-price information should be hidden when soliciting independent valuations.In 1974, Amos Tversky and Daniel Kahneman ran one of the most cited experiments in psychology. They spun a rigged wheel of fortune in front of participants - the wheel only ever stopped on 10 or 65 - and then asked a question with nothing to do with the wheel: what percentage of countries in the United Nations are African?
Participants who saw the wheel stop on 10 guessed, on average, 25%. Participants who saw it stop on 65 guessed 45%. A meaningless number, generated in front of them by a wheel they knew was random, shifted their answers by 20 percentage points. That is anchoring bias in its rawest form: the first number you encounter quietly becomes the gravitational centre of every estimate that follows.
The anchoring effect is now one of the most replicated findings in behavioural economics. It shows up in salary negotiations, house sales, courtroom sentencing, retail pricing, restaurant menus, and your own portfolio. Once you know to look for it, you start seeing the anchors everywhere - including the ones you set for yourself.
## What is anchoring bias, in plain English?Anchoring bias is the tendency to rely too heavily on the first piece of numerical information offered (the “anchor”) when estimating an unknown value. After encountering an anchor, people adjust away from it - but they do not adjust enough. The distortion persists even when the person knows the anchor is irrelevant, has been warned about the bias, or sees an obviously absurd anchor.
The mechanism is well-studied. Estimating something uncertain, your brain looks for a starting point; any recently-presented number gets used. From there you adjust, but adjustment is effortful and people stop too early. Tversky and Kahneman called this anchoring and adjustment. Later research added a second mechanism: anchors prime related information from memory, biasing the answer in their direction.
**The one-line definition** Anchoring bias is the systematic tendency for unrelated numbers to drag your estimates of unknown quantities toward themselves - even when you know the anchor is meaningless. ## Why does whoever names a salary first usually win?The old negotiation advice was “never name a number first.” The research now says the opposite. In a much-cited study by Adam Galinsky and Thomas Mussweiler at Northwestern, the side that opened with a specific number got a final price closer to their preferred outcome - by a substantial margin - in both buyer and seller roles.
The reason is anchoring. When a candidate says “I’m looking for £75,000,” the recruiter’s mental search starts at 75. If the recruiter speaks first with “the band tops out at £58,000,” the candidate’s search starts at 58 and quietly accepts the band as the constraint.
Two practical implications:
If the other side anchors first, the most reliable counter is not to negotiate down from their number - it is to refuse the frame: “That’s a long way from where I need to be. I’m looking for £82,000.” Replace their anchor with yours rather than tugging against it.
## How does the listing price anchor your offer?Gregory Northcraft and Margaret Neale’s 1987 study is the cleanest demonstration of anchoring in a professional setting. They walked two groups - undergraduates and experienced estate agents - through the same house with the same information packet. The only thing they varied was the listing price printed on the packet.
The students were nudged by the listing price, as expected. The agents - who explicitly said the listing price had not influenced them - were nudged by almost the same amount. A higher listing anchor produced a higher appraisal across every measure: estimated value, estimated sale price, and lowest acceptable offer. The professionals were anchored and did not know it.
This is why listing strategy matters even when the market “knows” the right number. List too low and offers anchor below market; list aggressively (within reason) and offers cluster nearer the asking price. It is also why buyer’s agents do their own comparable analysis before looking at the listing - once you have read “£525,000” on Rightmove, you cannot unread it.
## Why is your buy price the most expensive anchor in investing?The most damaging investing anchor is the price you paid. It has no information about future returns and does not appear in any rational valuation model. Yet investors anchor on it relentlessly - selling winners early to “lock in the gain” and refusing to sell losers until they “get back to even.” This is the disposition effect, and the purchase price is the anchor at its core.
Terrance Odean’s analysis of 10,000 retail brokerage accounts found investors were roughly 1.5 times more likely to sell a winner than a loser of equivalent size - and the winners they sold went on to outperform the losers they kept. The anchor was costing them measurable returns every year.
Other investing anchors that quietly distort decisions:
The fix is structural: build a sell rule that does not reference your purchase price. Sell when the thesis breaks, valuation crosses a ceiling, or a rebalance is due - not when the share price crosses an arbitrary number you happened to pay.
**A useful prompt** Before any sell or hold decision, ask: "If I owned no shares of this company today, would I buy it at the current price?" If the answer is no, the anchor on your buy price is doing the talking, not the analysis. ## What anchors are retailers using right now?Every retailer’s pricing team understands anchoring - even if they call it something else. The reference price next to the sale price is the anchor; the actual price is the offer. Strip the anchor away and conversion drops, often by a lot.
The classic example is the William-Sonoma bread maker. The retailer launched a model at $279 and sales were unimpressive. They added a larger, more expensive model at $429 alongside it. The $429 unit barely sold - but sales of the $279 model nearly doubled, because shoppers now had a higher anchor to compare it against and $279 suddenly looked reasonable.
Anchoring is most studied with numbers, but it operates anywhere your brain needs a reference point.
The bad news from debiasing research: simply being told about anchoring does not help much. People who have read about the bias still anchor. People who have just been warned in the same room still anchor. The effect is unusually robust against awareness alone.
The good news: a handful of structural techniques do reduce the effect, sometimes substantially.
1. **Generate your own anchor first** - Before reading the listing price, the salary band, or the analyst target, write down your own estimate. Your number, generated independently, becomes the reference you compare the external anchor against - rather than the other way round. 2. **Consider the opposite** - Ask yourself, in writing: "What are three reasons this number might be far too high?" then "What are three reasons it might be far too low?" This forces a wider range and pulls you off the anchor. 3. **Ignore extreme anchors entirely** - When an anchor is implausible (the £1.4M asking price for a £600K house), do not negotiate down from it - refuse the frame and offer your own anchor. Adjusting from a bad anchor still leaves you closer to it than you should be. 4. **Use ranges, not point estimates** - Forecasts expressed as ranges ("between 12 and 18 months") are less anchored than single numbers. Where possible, demand probabilistic ranges from yourself and others. 5. **Separate estimation from disclosure** - Get bids, valuations, or forecasts independently, in writing, before anyone sees anyone else's number. Group estimation after one person speaks is mostly anchoring with extra steps. 6. **Build sell rules without buy-price references** - For investing specifically: write your exit criteria in terms of fundamentals, valuation multiples, or portfolio weights - never in terms of the price you paid. ## Frequently Asked Questions **Q: What is anchoring bias in simple terms?** Anchoring bias is the tendency to lean too heavily on the first number you see when estimating something uncertain. Even random or irrelevant numbers can pull your estimate toward themselves, and you typically do not adjust far enough away from them. **Q: Who discovered the anchoring effect?** Amos Tversky and Daniel Kahneman published the classic anchoring experiments in 1974 as part of their work on heuristics and biases - the same line of research that later won Kahneman the 2002 Nobel Prize in Economics. **Q: Does knowing about anchoring make you immune to it?** No. This is one of the most striking findings in the research: even subjects warned about anchoring, told the anchor is random, or paid for accuracy, still anchor. Awareness helps a little; structural techniques (independent estimates, considering the opposite, range forecasts) help much more. **Q: Should I name the first number in a salary negotiation?** The current research consensus says yes - provided your number is aggressive but defensible. Studies by Galinsky and Mussweiler show the side that anchors first tends to get a final outcome closer to their preferred figure, in both buyer and seller roles. **Q: How is anchoring different from confirmation bias?** Anchoring distorts your numerical estimate from a starting point. Confirmation bias is about selectively gathering and interpreting evidence to support what you already believe. The two often combine - the anchor sets your initial belief, confirmation bias defends it. ## The Bottom LineAnchoring is not a quirk you can think your way out of. It is a feature of how the brain handles numerical uncertainty: grab a starting point, adjust from there, stop early. Once you accept that the bias is structural rather than a personal failing, the playbook gets simpler.
Generate your own number first. Treat extreme anchors as noise to be replaced, not signals to negotiate down from. Build sell rules that do not reference your purchase price. Get independent estimates before group discussion. And when a price tag, salary band, or 52-week high is silently doing the persuading, name it out loud - anchors operate best in the dark.
Related reading: loss aversion compounds the anchoring problem in investing, the sunk cost fallacy shares the same disposition-effect machinery, and base rate neglect is the cognitive cousin that warps probability estimates the way anchoring warps point estimates, and overconfidence bias compounds the damage when the anchored estimate is acted on without calibration.
--- # Risk vs Uncertainty: The Distinction That Matters Source: https://expectedvalue.co.uk/blog/risk-vs-uncertainty/ # Risk vs Uncertainty: The Distinction That Matters One has odds you can calculate. The other doesn't. Treating them the same is how smart people end up surprised. **Answer**Risk and uncertainty are different concepts, formalised by Frank Knight in 1921. Risk applies when the probabilities are known or can be reliably estimated (a fair coin, an insurance pool, a casino game). Uncertainty applies when the probability distribution itself is unknown (a novel technology launch, geopolitical events, pandemic spread). Risk can be priced and hedged; uncertainty (Knightian uncertainty) cannot - it must be navigated with optionality, robustness, and reversibility. Most expensive forecasting errors come from treating Knightian uncertainty as if it were risk.
Risk is when you don't know what will happen, but you do know the odds. Uncertainty is when you don't even know the odds. A roulette wheel has risk. A new technology launching into a new market has uncertainty. Treating the second like the first - pretending you can attach a probability to genuinely unknowable outcomes - is the source of most spectacular forecasting failures, from the 2008 financial crisis to nearly every confident prediction about geopolitics.
The distinction is more than a semantic quirk. It changes how you should size bets, build portfolios, plan strategy, and decide what to ignore. This guide unpacks where the idea came from, why it matters, and how to operate sensibly when you're in true uncertainty rather than measurable risk.
## What is Frank Knight's distinction between risk and uncertainty?The economist Frank Knight introduced the formal split in his 1921 book Risk, Uncertainty, and Profit. Knight argued that risk applies to situations where outcomes follow a known probability distribution - coin flips, dice rolls, mortality tables, equipment failure rates calculated from large maintenance datasets. You don't know the next outcome, but you know the odds, and over enough trials those odds will hold.
Uncertainty, in Knight's sense, applies to situations where the probability distribution is itself unknown. Will this start-up succeed? Will this election result reshape trade policy? Will this drug pass Phase 3 trials? You can guess. Experts can guess better than novices. But there is no underlying frequency you can sample to learn the true odds, because the situation is essentially unique.
Knight's insight: profit, in a competitive market, is the reward for bearing genuine uncertainty, not measurable risk. Risk can be priced and insured. Uncertainty cannot - and bearing it is what separates entrepreneurs from clerks.
**The shorthand version** Risk = roulette. You can calculate the house edge. Uncertainty = launching a new product into a market that does not yet exist. There is no edge to calculate, only judgement. ## Examples: Risk vs Uncertainty in the WildThe cleanest way to internalise the distinction is through paired examples - two superficially similar situations, one of which is risk and one of which is uncertainty.
An insurance company writing 100,000 home policies operates in pure risk. The frequency of fires, floods, and break-ins is well-documented across decades of claims data. Premiums are priced from those frequencies plus a margin. The business is mathematics applied to large numbers.
An entrepreneur launching a new home-insurance product targeting renters in a market with shifting climate patterns operates in uncertainty. They can borrow base rates from adjacent markets, but the actual distribution of outcomes for this product, in this market, at this moment is genuinely unknown.
A blackjack card counter is taking risk. Card decks have known compositions; the math is fixed; the only edge comes from tracking what's been dealt. A skilled counter knows their long-run expected value to several decimal places.
A poker tournament player faces uncertainty layered on top of risk. The cards have known probabilities, but the opponents' strategies, tilt levels, fatigue, and table dynamics do not. Two equally skilled players can disagree wildly about how to play the same hand and both be defensible.
Investing in a global equity index over 30 years is closer to risk than to uncertainty. Long-run equity returns have a well-studied distribution. Year-to-year returns are noisy, but the 30-year base rate is stable enough to plan around.
Investing in a single early-stage start-up is uncertainty. Most fail. A few return 100x. The distribution is fat-tailed, sample sizes are small, and any individual investment's outcome is genuinely unknowable in advance.
## Fat Tails and Black SwansNassim Taleb's contribution to this conversation is a sharp warning: many situations that appear to be risk are actually uncertainty in disguise, because the probability distribution we're using is the wrong one.
Pre-2008, mortgage-backed securities were modelled with risk-style distributions calibrated on a few decades of US housing data. The models implied that a nationwide drop in house prices was a once-in-thousands-of-years event. The distribution was wrong - partly because the underlying market structure had changed (subprime origination, securitisation incentives), partly because the historical sample didn't include the conditions that produced the crash. What looked like measurable risk was actually deep uncertainty about a fundamentally novel system.
Taleb calls these blow-ups Black Swans: outliers, impactful, and rationalised after the fact. The deeper point is that thin-tailed risk models applied to fat-tailed reality are a recipe for periodic disaster. The 1987 stock market crash was a 20-standard-deviation event under the prevailing models. The 2008 credit crisis, similar. Long-Term Capital Management's collapse, similar. None of these should have been possible. They keep happening anyway.
The lesson is not to abandon probability - it's to be brutally honest about which regime you're in. If your distribution is calibrated on a short, calm sample of a system that has since changed structure, you don't have risk. You have uncertainty wearing a lab coat.
## Rumsfeld's FrameworkDonald Rumsfeld's much-mocked 2002 press conference quote turns out to be a cleaner taxonomy than the headlines suggested. He distinguished:
The implicit fourth category, sometimes added by analysts, is unknown knowns - things we know but pretend we don't, the uncomfortable truths buried in footnotes that nobody wants to act on. The 2008 crisis had plenty of these.
The practical use of the framework is asking, before any consequential decision: which category dominates here? If it's known knowns, just execute. If it's known unknowns, run the numbers and size the bet appropriately. If it's unknown unknowns, the right move is usually to limit your downside rather than optimise your expected return, because the expected return calculation is unreliable.
## Implications for InvestingThe risk-uncertainty distinction reshapes how you should approach markets in three concrete ways.
The Kelly Criterion tells you how much to bet given an edge. But Kelly assumes you know the true probability. When you're in uncertainty rather than risk, your edge estimate has its own large error bar, and most practitioners advocate fractional Kelly - sizing at half or a quarter of full Kelly to absorb the parameter error. The deeper the uncertainty, the smaller the fraction.
Standard portfolio theory diversifies away idiosyncratic risk - the known unknowns. But it leaves you exposed to common-mode failures (everything goes down together) and to risks the historical sample never contained. Robust diversification means owning assets that respond to fundamentally different drivers, including some that look uncorrelated for the wrong reasons in calm periods.
One-quarter earnings forecasts: closer to risk. Ten-year sector forecasts: deep uncertainty. The further out the forecast, the more it depends on the structure of the world remaining stable. Most ten-year forecasts that look precise are uncertainty dressed as risk. The rational response is wider confidence intervals and less leverage on the central case.
## Implications for Business StrategyStrategy is largely the art of operating in uncertainty rather than risk, which is why importing financial-engineering precision into strategic planning so often misfires.
When uncertainty is high, preserving optionality is worth more than picking the optimal path. Pilot programmes, staged investments, reversible decisions, and product platforms that can pivot all preserve options. Big-bang commitments that depend on a single forecast being correct destroy them. The five-year strategic plan with a single revenue trajectory is usually worse than a one-year plan with explicit branch points.
Before committing to a strategy, run a pre-mortem: imagine the project has failed two years from now, and write the obituary. The exercise surfaces unknown unknowns by giving the team permission to voice doubts that would otherwise be filtered out as off-narrative. It's one of the cheapest, most effective tools for navigating uncertainty.
Taleb's other contribution is that forecasters who don't bear the consequences of their forecasts converge on confident-sounding nonsense. When evaluating any prediction in your business, ask: what does the predictor lose if they're wrong? If the answer is nothing, downweight the prediction.
## How to Operate Under True UncertaintyIf you accept that much of life is uncertainty rather than risk, the operating principles change. A practical checklist:
If this distinction was useful, three other posts on this site will sharpen it further:
Prediction markets are exchanges where people trade contracts that pay out based on whether real-world events happen. A typical contract pays $1 if the event occurs and $0 if it does not. The current price - somewhere between $0 and $1 - is the market's collective estimate of the probability.
If you understand expected value, you already understand most of what makes these markets tick. A contract trading at $0.62 implies the market thinks there is a 62% chance of the event happening. If you believe the true probability is 70%, buying "Yes" at $0.62 has a positive expected value of $0.08 per contract. That is the entire game, scaled up.
This guide is for readers who already think in probabilities. We will skip the popular-science explanations and get to the parts that matter: how the prices form, what the platforms actually offer, where the track record is genuinely impressive, where it falls apart, and how to use markets as an input to your own decisions. UK readers - there are some important caveats for you, which we will get to.
If the relationship between price and probability still feels foreign, our guide to probability vs odds walks through how decimal, fractional and American odds map onto exactly this kind of implied-probability thinking.
## What a Prediction Market Contract Actually IsA prediction market contract is a binary option on a future event. The simplest form looks like this:
If the Yes contract is trading at $0.43, the market is implicitly saying there is a 43% probability of the cut. Buy 100 Yes contracts at $0.43 and you have spent $43. If the event happens, you receive $100, a $57 gain. If it does not, you receive $0, a $43 loss.
This is mathematically identical to an even-money bet at decimal odds of 1/0.43 ≈ 2.33. The two formats are interchangeable; prediction markets just present the probability directly. For readers comfortable with expected value calculations, the EV of buying at $0.43 when you believe the true probability is p is simply: EV = (p × $1) − $0.43. Positive whenever p > 0.43.
**The core insight** A prediction market price is not a forecast made by one analyst. It is the equilibrium between thousands of traders who all stand to lose money if they are wrong. That financial skin in the game is what makes the price meaningful. ## Why Prices Behave Like ProbabilitiesThere are three independent reasons prediction market prices tend to track real probabilities.
1. Arbitrage forces consistency. If "Trump wins" trades at $0.55 and "Trump loses" trades at $0.50, the sum is $1.05. A trader can sell both sides, lock in a $0.05 risk-free profit per pair, and the prices must converge until that opportunity disappears. The same logic applies across related contracts: if "Republican wins presidency" trades higher than the sum of all individual Republican candidates, arbitrageurs will compress the gap.
2. Informed traders push prices toward truth. Anyone with a genuine information edge is incentivised to trade until the price reflects what they know. If you know something the market does not, you can extract money from people who are wrong. This is the same mechanism that supposedly keeps stock markets near fair value.
3. Diverse opinions average out noise. Even when individual traders are biased or poorly informed, their errors often cancel. This is the wisdom of crowds effect - the average of many independent estimates is usually closer to the truth than any single estimate.
None of this means prices are exactly right. It means they are usually a better starting point than your gut, and often better than expert opinion. Importantly, the same forces that keep stock prices roughly efficient apply here, with one big asterisk: prediction markets are far smaller and thinner, so the efficiency is weaker.
## The Major Platforms The landscape has consolidated rapidly over the last few years. Five platforms cover almost everything most people need to know about. Their differences matter - particularly which ones accept real money, which accept which nationalities, and how they resolve disputes. ## Polymarket: The Liquidity Leader Polymarket is built on the Polygon blockchain and settles in USDC, a US dollar stablecoin. It hosts most of the largest-volume contracts in the world: presidential elections, Fed decisions, geopolitical events, sports, and longer-tail prediction questions. Liquidity is concentrated in the headline contracts - you can move six-figure positions in major election markets with limited slippage. The catch is regulatory. Polymarket settled with the CFTC in 2022 and formally restricts US users; in practice, US-based traders sometimes use VPNs, which is against the terms of service and creates real legal risk. UK users sit in a grey area: trading on a non-FCA-regulated overseas crypto-settled venue is not explicitly illegal for retail users, but it is not protected and any earnings would still be liable for UK tax. ## Kalshi: The Regulated US Option Kalshi is the only CFTC-regulated prediction market exchange operating in the US. It runs as a Designated Contract Market under the same regulatory regime as commodity futures exchanges. That means real US dollar settlement, federally regulated custodians, and clear tax treatment. Kalshi's contract menu is more conservative than Polymarket's - it had to fight long legal battles for the right to list event contracts on US elections, and as of 2025 court decisions have generally favoured the exchange. The flagship economic and political contracts have decent liquidity, but obscure questions can be very thin. Kalshi does not currently accept UK retail customers. For US-based readers who want to trade real money on prediction markets without regulatory risk, Kalshi is effectively the only legitimate venue. ## PredictIt: Academic Roots, Capped Stakes PredictIt is run as a research project by Victoria University of Wellington under a CFTC no-action letter. It caps each user at $850 per market and charges a 5% fee on profits plus a 5% withdrawal fee. The combination of small caps and high fees means prices can drift noticeably from fair value - a 2-3% mispricing can persist because the arbitrage profit does not justify the friction. For researchers and individual traders willing to live with thin contracts, PredictIt remains a useful sandbox. For anyone trying to trade size, the caps make it impractical. ## Manifold and Metaculus: Forecasting Without Money Manifold is a play-money prediction market - users trade with "mana" that has no cash value (with limited charity-redemption mechanics). What it lacks in financial stakes it makes up for in breadth: thousands of active markets covering everything from AI benchmarks to neighbourhood gossip. Metaculus is structurally different. It is not a market at all but a forecasting tournament where users submit probability estimates that aggregate into a community forecast. Forecasters are scored on accuracy over time, building reputation. The platform has produced striking calibration data: over thousands of resolved questions, the community forecast tends to be well-calibrated, meaning events forecast at 70% happen roughly 70% of the time. Neither platform offers real-money trades, but both produce probability estimates that are cited by serious researchers and journalists. They are excellent for learning to forecast without putting money at risk. ## The Track RecordThe honest answer is quite well, with caveats. The strongest evidence comes from elections, where prediction markets have a long history of being compared head-to-head with polls and pundits.
Iowa Electronic Markets, run by the University of Iowa since 1988, demonstrated that small real-money markets often outperformed national polls in the final week of US presidential elections. A widely cited Berg, Forsythe, Nelson and Rietz paper found that IEM prices beat polls 74% of the time across multiple election cycles.
More recently, Polymarket's 2024 US election markets were closer to the actual outcome than nearly all major polling aggregators in the days before voting. The market priced Trump's win at roughly 60% on election eve while polls had the race at a coin-flip. This is one data point, not proof - but it is consistent with a wider pattern of markets outperforming pollsters in tight races.
Metaculus publishes calibration plots showing the community forecast is broadly well-calibrated across thousands of questions, though it tends to be slightly underconfident on extreme probabilities (95% predictions happen closer to 90%, etc.). Tetlock's Good Judgement Project demonstrated that small teams of trained "superforecasters" - using techniques very similar to market participants - outperformed intelligence-community analysts with classified information.
**Important caveat** Prediction market accuracy is best demonstrated for high-volume, well-defined questions with lots of trader attention. For obscure questions with thin liquidity, the prices are much noisier - sometimes worse than a coin flip. The track record is real, but it does not generalise to every question on every platform. ## Where Prediction Markets Fall ShortIf you treat prediction market prices as gospel, you will get burned. The honest weaknesses are:
Thin liquidity in long-tail markets. A market with $5,000 of total volume can be moved by a single retail trader. The price reflects whatever opinions happen to have shown up, not a deep equilibrium. Always check volume and order-book depth before reading much into a price.
Long-shot bias. Just like in horse racing, low-probability outcomes (sub-5% or sub-10%) tend to be systematically overpriced. Traders pay a premium for the chance at a big payoff. If you are an EV-driven trader, this is where the most reliable edge tends to live - selling the long shots, not buying them.
Resolution ambiguity. Markets pay out based on a defined resolution criterion, and ambiguous criteria create disputes. Polymarket's resolution mechanism (UMA's optimistic oracle) has been challenged on several high-stakes markets. Always read the resolution criteria carefully - "Will Russia invade Ukraine in 2024?" sounds simple until you ask "what counts as an invasion?".
Manipulation risk on small markets. A trader with deep pockets can push thin-market prices around for short periods, especially near close. There have been documented cases of suspected manipulation around political markets - though prices typically snap back as arbitrageurs arrive.
Sample-size problems on rare events. Prediction markets are excellent at the headline questions because there is decades of data on elections, sports and macro events. They are much weaker at one-off, novel questions where neither the traders nor the platforms have base rates to anchor on. This is closely related to base rate neglect - without good base rates, neither humans nor markets do well.
Legal status varies by jurisdiction. US users are restricted on most non-Kalshi platforms. UK users have no fully regulated equivalent (more on this below). Tax treatment is unclear in most jurisdictions and may shift as regulators catch up.
## Prediction Markets vs Sports Bookmaker OddsSports betting odds and prediction market prices are mathematically equivalent - both encode an implied probability - but the structure is different. A traditional sportsbook sets the odds and takes the other side of every bet, building in a margin (the "vig" or "overround") so the implied probabilities sum to more than 100%.
A betting exchange is far closer to a prediction market. Smarkets and Betfair Exchange in the UK are exactly this: peer-to-peer markets where users back and lay outcomes against each other, and the exchange charges a small commission on net winnings instead of building margin into the price. The lay/back prices on Betfair Exchange are the closest legitimate equivalent UK readers have to a Polymarket-style prediction market.
The practical implications:
If you want to apply prediction-market thinking from the UK without the legal greyness of Polymarket, Betfair Exchange and Smarkets are the practical answer.
## How to Actually Use Prediction MarketsMost readers should not try to make a living trading prediction markets. The realistic uses are more mundane and more valuable.
1. As a probability source for your own decisions. When you need to estimate the chance of a macro event - a recession, an election outcome, a geopolitical development - the relevant prediction market gives you a starting probability calibrated by people with money on the line. Use it the way you would use a base rate.
2. As a forecasting tool. Submitting forecasts to Manifold or Metaculus, or simulating trades in real markets, is one of the best ways to develop calibration. Track your predictions over time and compare to actual outcomes. The discipline of writing down a probability before the event happens - and being scored on it - exposes biases you did not know you had. This is essentially Bayesian thinking in practice.
3. As occasional positive-EV trades. If you have genuine information or analysis that the market lacks, prediction markets give you a way to monetise it. The bar is high - you need to be confidently right where the market is wrong, and you need enough liquidity to size the bet meaningfully. The same logic that governs position sizing in any edge-based strategy applies here: even with a real edge, betting too much can blow you up.
## Worked Example: An EV Trade on a Hypothetical MarketImagine a Polymarket contract: "Will the Bank of England cut rates at the November meeting?" The Yes contract is trading at $0.30. You have read the latest MPC minutes carefully, know two of the dovish committee members publicly favour a cut, and your model puts the probability at 45%.
Setting up the EV calculation:
EV per contract = (0.45 × $1.00) + (0.55 × $0.00) − $0.30 = $0.15
That is a 50% expected return on the dollars at risk, which sounds great. But translate it into a Kelly-style position size before celebrating.
The Kelly fraction for a binary bet is: f = (bp − q) / b*, where b is the net odds (here, $0.70 of profit per $0.30 staked, so b = 0.70/0.30 ≈ 2.33), p is your probability of winning (0.45), and q is the probability of losing (0.55).
f = (2.33 × 0.45 − 0.55) / 2.33 = (1.05 − 0.55) / 2.33 ≈ 0.21*
Full Kelly says stake 21% of your bankroll on this single trade. Most experienced traders use fractional Kelly - typically a quarter to a half - to absorb the inevitable miscalibration in your probability estimate. Quarter Kelly here is about 5% of bankroll. That is a reasonable size for a single market view, and it acknowledges that your 45% might really be 38%. EV thinking only works if you stay around long enough for the maths to play out, which means sizing positions so that you survive being wrong several times in a row.
**Reality check** Most retail traders systematically overestimate their edge. If your model says the market is mispriced by 15 percentage points on a high-profile question, the most likely explanation is that your model is wrong, not that thousands of traders are. ## The 2025 Context The 2024 US election cycle was a watershed moment for prediction markets. Polymarket's election volume reached unprecedented levels, with cumulative volume across major political contracts running into the billions of dollars over the year. Across the broader crypto-prediction-market space, total market capitalisation is reported to have reached approximately $64 billion at peak - though these figures vary by source and should be treated as rough orders of magnitude rather than precise statistics. The regulatory picture also moved. Kalshi's long fight with the CFTC over event contracts produced favourable court rulings in 2024-2025, broadly establishing that CFTC-regulated event contracts on elections can be offered legally to US retail. Polymarket's restrictions on US users remain, but the wider regulatory hostility appears to be softening. For traders, the practical takeaway is that real-money prediction markets are now a permanent and growing part of the financial landscape, not a fringe curiosity. Liquidity is deep enough on flagship contracts that the prices genuinely matter for forecasting purposes. For UK readers in particular, the world has not changed much: Polymarket remains accessible but unregulated, Kalshi is unavailable, and the betting exchanges remain the cleanest legal route to similar trades. ## Frequently Asked Questions **Q: Are prediction market prices actually probabilities?** Approximately, yes. The market price is the equilibrium between buyers and sellers with money on the line, which makes it a useful estimate of probability. It is not exact: there is long-shot bias (low probabilities tend to be overpriced), liquidity premia, and short-term noise. For high-volume, well-defined contracts th …[truncated for length] --- # Correlation vs Causation: A Probabilistic Thinking Guide Source: https://expectedvalue.co.uk/blog/correlation-vs-causation/ # Correlation vs Causation: A Probabilistic Thinking Guide Why your brain confuses the two - and how to think clearly about cause and effect. **Answer**Correlation is when two variables move together; causation is when one variable causes the other to change. A correlation can exist without causation - the canonical example is ice cream sales and drowning deaths (both rise in summer because of the confounder 'warm weather', not because ice cream causes drowning). Tyler Vigen's Spurious Correlations archive documents thousands of these. To establish actual causation: randomised controlled trials are the gold standard; instrumental variables, regression discontinuity, and Granger causality help when RCTs aren't possible. Common pitfalls: confounders, reverse causation, selection effects, Simpson's paradox.
Correlation is when two things move together. Causation is when one of them is the reason the other moves. They look identical from the outside, and that's why almost every confidently held belief about cause and effect is wrong.
Ice cream sales and drowning deaths are tightly correlated. So are the number of pirates in the world and global temperature. So are nations that consume more chocolate and the number of Nobel laureates they produce. None of these is a causal relationship - but if you only had the data and a hunch, you'd swear they were.
This guide is about how to spot the difference. We'll cover spurious correlations and why they're so common, Simpson's paradox (where a trend reverses when you split the data), confounding variables, and the actual methods researchers use to establish causation. By the end, you'll have a working mental toolkit for separating “A and B happen together” from “A causes B.”
**The headline distinction** Correlation describes a pattern in observations. Causation describes a mechanism by which one thing produces another. You can prove a correlation with data alone. You almost never can prove causation without an intervention or a very careful argument. ## What does correlation actually mean?A correlation is a measurable, repeatable association between two variables. When one moves, the other tends to move with it (positive correlation) or against it (negative correlation). Statisticians quantify this with a correlation coefficient - usually labelled r - that ranges from -1 to +1.
Real-world correlations almost never hit ±1. A coefficient of 0.7 is considered very strong; 0.4 is moderate; 0.1 is barely there. But here's the catch: even a perfect correlation tells you absolutely nothing about why the variables move together. Correlation is descriptive. It says “this pattern exists.” It does not say what causes the pattern.
Two variables can correlate strongly because:
Of those five, only the first is genuine causation. The other four account for the overwhelming majority of correlations you'll encounter in life, science, and journalism.
## What is a spurious correlation?Tyler Vigen's Spurious Correlations project compiled hundreds of these. A few favourites:
None of these are causal. They're statistical noise - patterns that emerge when you trawl enough variables looking for matches. With sufficient data, you can find a strong correlation between almost any two unrelated time series, especially if both happen to be trending in the same direction over the period you measure.
This is the first lesson of correlation vs causation: the existence of a correlation, on its own, is not strong evidence of anything causal. It's a starting hypothesis at best.
**The garden of forking paths** If you compare 100 unrelated variables against each other, you'd expect five strong-looking correlations to appear by pure chance at the standard 5% significance threshold. That's not statistical malpractice - it's just maths. Researchers who fish through data looking for “significant” results almost always find some, even when none truly exist. ## What are the most famous spurious correlations?The Tyler Vigen archive and decades of statistical writing have surfaced specific examples that make the principle vivid:
The pattern across all ten: any two variables that move together over a finite time window can produce a spectacular-looking correlation coefficient. Whether they're causally linked is a separate question that requires more than the correlation evidence to answer.
## What is a confounding variable?Ice cream sales and drowning deaths are correlated - both go up in the summer. A naive analysis might suggest that ice cream causes drowning, or that drowning causes ice cream sales. Obviously neither is true. The real cause is a third variable: hot weather. Hot days drive both ice cream consumption and the likelihood of swimming in the first place.
This third variable is called a confounder. It causes both observed variables, creating a correlation between them that has nothing to do with one causing the other.
Confounders are everywhere in observational data:
- **Coffee drinkers live longer than non-drinkers** - Confounder: people who can afford and tolerate coffee tend to be healthier overall. - **Children with bigger feet have better vocabularies** - Confounder: age. Older kids have bigger feet and bigger vocabularies. - **Universities with smaller class sizes produce better grades** - Confounder: smaller classes are often at wealthier institutions with better-prepared students. - **Vitamin supplement users are healthier than non-users** - Confounder: people who buy supplements are also more likely to exercise, sleep well, and avoid smoking.The presence of a plausible confounder doesn't disprove a causal claim - but it forces you to control for it before taking the correlation seriously. Good observational research lists every plausible confounder and adjusts for it statistically. Bad research ignores them and reports the headline number.
## When does reverse causation flip the answer?Sometimes two things really are causally linked, but the direction is the opposite of what intuition suggests.
People who use stand-up desks report better focus. Does the desk cause better focus? Possibly. But it's at least as plausible that people who already have good focus and discipline are more likely to buy stand-up desks in the first place. Without a controlled experiment, you genuinely can't tell which way the arrow goes.
Other classic reverse-causation traps:
Reverse causation is particularly sneaky because it produces real correlations and feels like a causal mechanism. The only reliable defence is to ask: “If I imagine the arrow going the other way, does that story also fit the evidence?” If yes, you don't have a causal claim.
## What is Simpson's paradox?Simpson's paradox is one of the strangest things in statistics. It's a situation where a trend appears in several groups of data but reverses when the groups are combined - or vice versa.
The classic example is from a 1973 admissions audit at UC Berkeley. Looking at overall admissions, men were admitted at a noticeably higher rate than women, suggesting bias against women. But when researchers split the data by department, the pattern flipped: most departments were actually slightly biased in favour of women. The aggregate looked discriminatory because women happened to apply more often to highly competitive departments with low acceptance rates for everyone.
The same paradox appears repeatedly:
- **Treatment A looks better than Treatment B overall** - But Treatment B is better in every patient subgroup. The difference comes from how patients were assigned. - **A baseball player has a higher batting average each year than another** - But over a career, the second player has the higher average. The mix of years explains the reversal. - **A school's average test score drops year-on-year** - But every demographic subgroup is improving. The composition of the student body changed.Simpson's paradox is the strongest argument for never trusting an aggregate statistic without seeing the breakdown. The lesson is simple: a correlation that holds at one level can vanish or invert at another. Always ask, “What groups are inside this average, and what happens when I look at them separately?”
**Which level is the ‘real’ answer?** There's no universal rule. Sometimes the aggregate is right (it captures the population-level effect that matters for policy). Sometimes the subgroup is right (it captures the actual mechanism). The point is that you need to think about which question you're trying to answer before picking which level of data to trust. ## How do researchers actually establish causation?Establishing genuine causation is hard. It typically requires one of these approaches, in roughly increasing order of reliability:
1. **Bradford Hill criteria** - A checklist for evaluating observational evidence: strength of association, consistency across studies, specificity, temporal sequence (cause precedes effect), dose-response relationship, biological plausibility. Used in epidemiology when experiments aren't possible (smoking and lung cancer was established this way). 2. **Natural experiments** - Situations where some external factor (a policy change, a natural disaster, a lottery) randomly assigns people to different conditions. Researchers can then study the consequences as if it were a controlled experiment, even though no one designed it. 3. **Instrumental variables** - A statistical technique that uses a third variable correlated with the cause but not directly with the outcome to isolate the causal effect. Common in economics - for example, using rainfall to study how farm income affects local school attendance. 4. **Regression discontinuity** - When eligibility for an intervention depends on crossing a threshold (a test score, an income cutoff), comparing people just above and just below the line approximates a randomised experiment. 5. **Randomised controlled trials (RCTs)** - The gold standard. Randomly assign people to treatment or control, measure the outcome. Random assignment eliminates confounders by construction. Used in medicine, increasingly in policy and tech (A/B testing is essentially this).If you're reading a study that claims X causes Y and none of these methods is described - just a correlation in observational data - be sceptical. The study may still be useful as a hypothesis generator, but it almost certainly hasn't proved causation.
## How can you spot correlation-vs-causation errors in real life?You're not going to run an RCT next time you read a news headline. But you can ask these five questions, and they'll filter out most weak causal claims:
1. **Is there a plausible confounder?** - What third variable could be driving both the cause and the effect? If you can think of one easily, the claim probably hasn't accounted for it. 2. **Could the arrow point the other way?** - Try the reverse causation story. Does it also fit the evidence? If yes, you can't tell direction from this data alone. 3. **Is the effect size suspiciously large?** - Tiny interventions rarely produce huge outcomes. If a single weekly habit allegedly doubles your income, lifespan, or happiness, the claim is almost certainly oversold. 4. **Where does the data come from?** - Self-reports, social media, online surveys, observational studies - all are vulnerable to selection bias. Random samples and intervention studies are far more trustworthy. 5. **Has it been replicated?** - A single study, no matter how dramatic, is rarely conclusive. Wait for replication, especially in psychology, nutrition, and social science where reproducibility rates are low. ## Frequently Asked Questions **Q: Can correlation ever prove causation?** On its own, no. A correlation can be perfectly consistent with a causal relationship, but it's also consistent with confounding, reverse causation, coincidence, and selection effects. Causation requires either a controlled experiment or a careful argument that rules out the alternatives. **Q: If correlation isn't causation, why do scientists use observational studies at all?** Because experiments are often impossible - you can't randomly assign people to smoke for 30 years or live in different countries. Observational studies are how we generate hypotheses and gather circumstantial evidence. The Bradford Hill criteria exist precisely to evaluate this kind of evidence systematically when randomised trials aren't an option. **Q: How strong does a correlation need to be before it's worth taking seriously?** Strength matters less than mechanism. A correlation of 0.3 with a clear causal pathway and replication across studies is more meaningful than 0.9 with no plausible mechanism. That said, weak correlations (under ~0.2) in noisy real-world data are often just statistical artefacts. **Q: What's the difference between confounding and Simpson's paradox?** Confounding is a hidden cause of both variables that creates a misleading correlation. Simpson's paradox is what happens when a confounder is so strong that it actually reverses the apparent direction of an effect when you slice the data differently. Simpson's is a special, dramatic case of confounding. **Q: How does this apply to investing?** Strongly. Most market “edges” that look like causation in backtests are spurious correlations from a small sample, confounded by overall market direction, or the result of survivorship bias. See our expected value and probabilistic thinking guides for how to evaluate investment claims more rigorously. **Q: Are A/B tests the same as randomised controlled trials?** Yes, in the statistical sense. A well-designed A/B test randomly assigns users to variants and measures the outcome - exactly the same logic as an RCT. They're the most reliable causal tool most product teams have. The catch is that they only work for short-term effects on the variables you measure; long-term and second-order effects are still observational. ## The Probabilistic Bottom LineProbabilistic thinkers don't reject every correlation - they treat correlations as evidence with a weight that depends on the design behind them. A randomised trial moves the needle a lot. A natural experiment moves it some. A messy observational study with obvious confounders barely moves it at all.
Most causal claims you'll meet - in the news, on social media, in management books, in folk wisdom - are based on the weakest type of evidence. That doesn't make them all wrong. It does mean you should hold them loosely. Be willing to update if better evidence arrives, and be willing to discard them if the underlying study was a fishing expedition.
The goal isn't to become a sceptic of everything. It's to allocate your belief in proportion to the actual strength of the evidence - which is exactly what probabilistic thinking is for.
## Continue the Series --- # The Signal and the Noise Review (Nate Silver, 2012) Source: https://expectedvalue.co.uk/review/signal-and-noise-silver-review/ # The Signal and the Noise Review (Nate Silver, 2012) **Answer**Nate Silver's The Signal and the Noise (2012) is the most practical book in the probabilistic-thinking trio (alongside Tetlock's Superforecasting and Taleb's Fooled by Randomness). Its central contribution is making Bayesian reasoning accessible with worked examples across genuinely different domains - weather, poker, baseball, economics, politics, terrorism. Silver's central claim: forecastability varies dramatically by domain, and recognising which side of the line a question sits on is more important than the specific forecasting technique.
This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
Where to buy: The Signal and the Noise on Amazon.co.uk →
## What the book is aboutThe Signal and the Noise organises around a single claim: prediction is genuinely possible in some domains and genuinely impossible in others, and the difficult cognitive work is recognising which is which. Silver walks through 11 case studies - weather, baseball, poker, the 2008 financial crisis, US elections, climate, terrorism, chess, economic indicators, earthquakes, infectious-disease epidemics - and shows that 'forecasting' looks very different across them.
Weather forecasting is genuinely well-calibrated and improves slowly with computational power. Baseball forecasting (Silver's own PECOTA model) is accurate enough to be commercially useful. Election forecasting is harder but tractable when modelled properly. Earthquake forecasting at the prediction level (when, where, how big) is not currently possible despite decades of effort. Terrorism prediction is so badly-conditioned that historic intelligence-community claims about it are mostly false confidence.
The unifying analytical tool throughout is Bayes' theorem - the formal rule for updating probability estimates with new evidence. Silver's exposition of Bayes for non-statisticians is probably the best in popular nonfiction.
## Why Bayes' theorem mattersBayes' rule is intuitively simple but operationally counter-intuitive. The rule: posterior probability = prior probability × likelihood of new evidence / marginal probability of evidence. In practice: you start with a base rate, observe new evidence, and update the base rate in proportion to how surprising the evidence is given each possible hypothesis.
The cancer-screening example Silver works through is the canonical demonstration. We covered the same example in our probabilistic thinking guide - a 99%-accurate test for a 1-in-1000 disease, where a positive result still gives ~9% true-positive rate because the base rate dominates. Silver works the maths through patiently, then shows that the same Bayesian logic applies to almost every forecasting question once you frame it as 'prior + new evidence → posterior'.
## Where it sits in the probabilistic-thinking canonRead together with the other two books in the trio.
Taleb's Fooled by Randomness (2001) diagnoses the problem - survivorship bias and fat tails distort how we judge success in finance. Silver's book extends this beyond finance into 11 specific forecasting domains, with the diagnostic being domain-specific rather than universal.
Tetlock and Gardner's Superforecasting (2015) prescribes the calibration habits that beat naive intuition. Silver's book is more practical-tool-focused - here's how Bayes works, here's how PECOTA predicts baseball, here's why the polling-aggregation approach worked in 2008 and 2012.
If you only read one, Superforecasting has the best practical framework. The Signal and the Noise has the best Bayesian exposition. Fooled by Randomness has the best diagnostic frame. Read them in publication order (2001 → 2012 → 2015) for the cleanest narrative arc.
## What about the 2016 election?The Signal and the Noise was published in 2012, before Silver's most-discussed forecasting moment. FiveThirtyEight's 2016 model gave Donald Trump a roughly 30% probability of winning the presidency on election eve - higher than most other forecasters but lower than the actual outcome warranted.
Two readings of this. The defensive reading: 30% probability is not a confident prediction either way; a 30% event happens nearly one time in three, and the 2016 outcome was within the predicted distribution rather than a model failure. The critical reading: Silver's book confidently claims the political-prediction domain is tractable, and the 2016 result raised real questions about whether the polling-aggregation methodology has the variance-handling the framework requires.
Both readings have merit. The book's broader points about Bayesian reasoning, weather forecasting, and the variation of forecastability across domains aren't undermined by the 2016 case. The specific confidence in political forecasting is the most-aged section.
## Who should read it 1. **Anyone who wants to actually use Bayes' theorem in everyday reasoning** - Silver's exposition is the most accessible practical introduction to Bayesian thinking in popular nonfiction. Worth reading the Bayes-specific chapters (3-4) even if you skim the rest. 2. **Anyone curious how different domains compare on forecastability** - The 11 case studies show real variation - weather is well-calibrated; earthquakes are genuinely unpredictable; political forecasting is tractable but harder than weather. Useful intuition for which domains to treat as forecastable. 3. **Readers who want a counterweight to pure Taleb-style scepticism** - If Fooled by Randomness has convinced you that all forecasting is bunk, Silver's domain-by-domain treatment is the corrective. Some forecasting works; the trick is knowing where. 4. **Skip if 544 pages feels long for a single argument** - The book is the longest of the three in the probabilistic-thinking trio. Some sections (chess, earthquakes) feel padded relative to their analytical contribution. Strategic skipping is fine. ## Frequently asked questions **Q: What's The Signal and the Noise about?** Nate Silver's 2012 book on the practical mechanics of probabilistic forecasting across 11 domains (weather, poker, baseball, the 2008 financial crisis, US elections, climate, terrorism, chess, economic indicators, earthquakes, infectious disease). The central claim is that forecastability varies dramatically by domain and Bayes' theorem is the right analytical tool for getting from base rates to calibrated forecasts. **Q: Did the 2016 election damage The Signal and the Noise's credibility?** Partially. FiveThirtyEight gave Trump ~30% probability of winning in 2016 - higher than most forecasters but lower than the outcome warranted. 30% events do happen, so this isn't a clean model failure, but the book's confidence in political-prediction tractability is the most-aged section. The broader Bayesian framework and the domain-variability claims are unaffected. **Q: How does it compare to Superforecasting?** Complementary. Superforecasting is prescriptive - here are the calibration habits that produce better forecasts. The Signal and the Noise is practical-tool-focused - here's how Bayes works in 11 specific domains. Read together for the fullest picture; if you only read one, Superforecasting has the more directly actionable framework. **Q: Is the Bayes exposition really that good?** Yes - the cancer-screening worked example (Chapter 8) is the most accessible Bayesian-reasoning introduction in popular nonfiction, and the subsequent chapters apply the same framework to baseball, poker, and economic indicators with consistent rigour. Worth reading even if you skip the more political sections. **Q: Should I read the 2012 original or wait for a newer book?** Read the original. Silver has written subsequent books (notably On the Edge in 2024 about risk-takers) but The Signal and the Noise remains the canonical introduction to his thinking. The 2012 examples have aged but the analytical framework hasn't. For more current applications, follow FiveThirtyEight (now owned by ABC News) and Silver's Silver Bulletin substack. ## Related guides **Affiliate disclosure** We may earn a commission on Amazon UK purchases made via the link above. The disclosure doesn't change the editorial assessment - this is a research-led review based on the book's content and Silver's documented FiveThirtyEight forecasting record. See our Affiliate Disclosure. **Verdict:**Nate Silver's The Signal and the Noise sits between Taleb's Fooled by Randomness (diagnostic) and Tetlock's Superforecasting (prescriptive). Where Taleb argues that most forecasting is bunk and Tetlock argues that calibration is learnable, Silver argues that forecastability varies dramatically across domains - weather is genuinely predictable to a useful horizon, terrorism is not, and most everyday economic and political forecasting sits somewhere in between.
The book's biggest contribution is making Bayesian reasoning accessible to readers without statistical training. Silver works through Bayes' theorem with concrete examples (the cancer-screening case, the lab-test case, the baseball PECOTA forecasting case) in a way that demystifies the maths without dumbing it down.
Two caveats. First, the 2012 publication date predates Silver's 2016 election misfire, which raised hard questions about his confidence in the political-prediction sections. The book's claims are mostly defensible - 2016 was a rare-event outcome at the edge of the predicted distribution - but reading it without that context misses an important corrective. Second, the prose is more data-journalism-flat than Kahneman or Taleb; readers who prefer punchy aphorism may bounce.
Recommended as the practical applications complement to Tetlock + Taleb.
--- # Fooled by Randomness Review (Nassim Taleb, 2001) Source: https://expectedvalue.co.uk/review/fooled-by-randomness-taleb-review/ # Fooled by Randomness Review (Nassim Taleb, 2001) **Answer**Fooled by Randomness is Nassim Taleb's 2001 book establishing survivorship bias as the central distortion in how we judge success in finance, business, and other competitive domains. The argument: the people who get interviewed, written about, or imitated are systematically the ones who got lucky, and the 999 similar candidates who didn't survive are invisible. Best Taleb entry point - more readable than The Black Swan, foundational to the rest of the Incerto series.
This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
Where to buy: Fooled by Randomness on Amazon.co.uk →
## What the book is aboutFooled by Randomness's central claim is that competitive domains with high variance (finance, business, sport, entrepreneurship) systematically over-attribute success to skill and under-attribute it to luck. The mechanism is survivorship bias: the surviving fund manager, the successful entrepreneur, the bestselling author are visible because they survived, while the 999 similar people who didn't survive aren't. We see the visible cases and infer skill; the maths says many of them were lucky.
Taleb's case studies come mostly from finance - his own world before he became a public intellectual. The book sketches characters (the trader who blew up after years of success, the analyst who never blew up because he was structurally cautious) to make the point that 'good outcomes followed by bad outcomes' is what survivorship-biased markets look like from inside, not 'skill followed by mysterious decline'.
## The core ideas worth knowing - Survivorship bias. We see the visible survivors and infer their traits caused the survival. The non-survivors are invisible. Most success narratives are survivorship-biased. - Alternative histories. Outcomes draw from a distribution. A specific path that happened (the 10x return) isn't the same as the median path that could have happened given the same starting conditions. - Rare events. Catastrophic losses in fat-tailed distributions are larger and more frequent than the normal-distribution intuition suggests. Insuring against rare events is rational even when EV says otherwise. - Mediocristan vs Extremistan. Some domains (height, weight, IQ) have bounded distributions where extreme outcomes are impossible. Others (wealth, book sales, market returns) are unbounded, where one extreme outcome can dominate the whole distribution. - The narrative fallacy. Humans construct after-the-fact explanations for outcomes that were largely random. Most business-success stories are narrative fallacy applied to survivorship-biased samples. ## Where the book is strongestThree things put Fooled by Randomness ahead of most popular finance writing.
Readability. Taleb writes in extended anecdote and aphorism rather than economic exposition. The argument is carried by characters - some real, some composite - rather than by formal proofs. That makes the book accessible to readers who'd bounce off a dry treatment of the same material.
Durability. Twenty-five years on (2001 publication), the core ideas haven't been falsified or superseded. If anything, the 2008 financial crisis and the post-pandemic asset-price moves both vindicated Taleb's central case that fat-tailed distributions produce surprises the normal-distribution mindset can't handle. The book has aged well.
Entry point. As Book 1 of the Incerto, Fooled by Randomness introduces the vocabulary - survivorship bias, alternative histories, rare events - that the rest of the series develops. Reading the Incerto out of order (starting with The Black Swan or Antifragile) means missing the foundational case that those later books build on.
## Where it falls shortThree honest caveats.
The tone is divisive. Taleb is famously combative in print, and Fooled by Randomness sets the template. Readers who find the prose abrasive often struggle to reach the analytical core. The defensible position is that the tone is doing work - it's anti-narrative, anti-skill-attribution, and intentionally rude to the journalists and analysts Taleb thinks have made the survivorship bias worse. Whether that justifies the reading friction is a personal call.
Light on quantitative method. The book recognises patterns - survivorship bias, fat tails, narrative fallacy - but doesn't give you a measurement framework or a decision rule. For 'how do I actually calibrate my forecasts', Tetlock's Superforecasting is the better companion text. Fooled by Randomness is diagnostic; Superforecasting is prescriptive.
Sharper successor exists. The Black Swan (2007) extends the central ideas with more rigour and a tighter argument. Some readers prefer to skip Fooled by Randomness and start there. The case for reading FbR first is that it's shorter, more readable, and establishes the vocabulary; the case against is that The Black Swan reaches the same conclusions more efficiently. Either order works.
## Who should read it 1. **First-time Taleb readers** - Fooled by Randomness is the right Incerto entry point - shorter, more accessible, and establishes the vocabulary the later books build on. Reading the series out of order works but means missing the foundational case. 2. **Anyone in finance or active investing** - The survivorship-bias case is most directly applicable to fund-manager and trader evaluation. The book changes how you read most performance attribution after you've absorbed it. 3. **Anyone in any field with high outcome variance** - Founders, sales people, sports professionals, journalists, authors - any domain where the visible survivors are a tiny biased fraction of the original cohort. The framework transfers beyond finance. 4. **Skip if you want a calibration framework** - FbR is diagnostic - 'here's what's wrong with how you think about success'. For prescriptive 'here's what to do differently' content, Superforecasting is the better fit. ## Frequently asked questions **Q: Is Fooled by Randomness the same book as The Black Swan?** No - Fooled by Randomness is the 2001 first book in Taleb's Incerto series; The Black Swan is the 2007 sequel. FbR focuses on survivorship bias and the role of luck in success; The Black Swan focuses on rare high-impact events ('black swans') and the limits of probability-based reasoning when distributions are fat-tailed. They share vocabulary and worldview but argue different specific points. **Q: Which Taleb book should I read first?** Fooled by Randomness, in most cases. It's shorter, more readable, and establishes the vocabulary the rest of the Incerto develops. Read The Black Swan second, Antifragile third, Skin in the Game fourth (the Incerto's later books). Some readers prefer The Black Swan as a starting point because its argument is tighter; either order works. **Q: Is Fooled by Randomness still relevant in 2026?** Yes. Twenty-five years on, the central ideas - survivorship bias, fat tails, alternative histories, narrative fallacy - haven't been falsified and have arguably been vindicated by both the 2008 financial crisis and post-pandemic asset-price moves. Some specific market case studies have aged but the framework remains current. **Q: How does Fooled by Randomness compare to Superforecasting?** Complementary. FbR is diagnostic - 'here's what's wrong with how you reason about success and rare events'. Superforecasting is prescriptive - 'here are the habits that improve your forecast calibration'. Read FbR first to internalise the survivorship-bias frame; read Superforecasting second to learn what to do about it. **Q: I find Taleb's tone abrasive - is the book worth reading anyway?** Mostly yes, but you can skim the more personality-driven sections (especially the early chapters where Taleb is settling scores with named individuals from his trading days). The analytical core - survivorship bias, fat tails, narrative fallacy - sits in Chapters 3-7 and is worth reading even if you skip the rest. If the tone is a complete deal-breaker, jump to The Black Swan (slightly more measured) or read our behavioural-finance explainers for the underlying ideas in different prose. ## Related guides **Affiliate disclosure** We may earn a commission on Amazon UK purchases made via the link above. The disclosure doesn't change the editorial assessment - this is a research-led review based on the book's content and the wider Incerto literature. See our Affiliate Disclosure. **Verdict:**Fooled by Randomness is the book that established Nassim Taleb's intellectual brand. Published in 2001, it makes the case that success in finance, business, and other competitive domains is far more shaped by luck than the people involved (or the journalists who write about them) acknowledge. Survivorship bias is the central organising idea: the fund manager who beat the market for ten years looks impressive until you remember the 999 similar managers who didn't survive to be interviewed.
The book is readable, often funny, and rich in market-anecdote case studies. Twenty-five years on, its central ideas have been so widely absorbed that newer readers may find the argument obvious - which is exactly the kind of compliment Taleb's books pay themselves into. Where it falls short is in quantitative method: the conceptual frame is sharp but the practical 'what should I do differently' content is thinner than in The Black Swan (2007), the next book in the Incerto series.
For most readers, Fooled by Randomness is the right Taleb entry point. It's shorter, more accessible, and frames the survivorship-and-luck core that the rest of the Incerto develops further. Recommended.
--- # Superforecasting Review (Tetlock & Gardner, 2015) Source: https://expectedvalue.co.uk/review/superforecasting-tetlock-gardner-review/ # Superforecasting Review (Tetlock & Gardner, 2015) **Answer** Superforecasting is the most rigorously-researched book in the probabilistic-thinking canon. It reports on the Good Judgment Project, a 4-year IARPA-funded forecasting tournament that identified 'superforecasters' - the top 2% of volunteer forecasters who consistently outperformed CIA analysts on short-horizon political and economic questions. The book's practical contribution is the 10-commandment framework for improving forecast calibration. Best for anyone whose work involves estimating uncertain outcomes (analysts, investors, founders, policy).This page contains affiliate links. If you buy through them we may earn a small commission at no extra cost to you - it helps keep the site running. See our affiliate disclosure.
Where to buy: Superforecasting on Amazon.co.uk →
## What the book is aboutBetween 2011 and 2015, the Intelligence Advanced Research Projects Activity (IARPA - the US intelligence community's equivalent of DARPA) ran the Good Judgment Project: a competitive forecasting tournament where teams of volunteers made probability estimates on hundreds of geopolitical and economic questions. The tournament was structured as research - the forecasters' calibration could be measured against ground truth, and the top performers identified statistically.
The headline finding: a small group of volunteer forecasters (the top 2%, dubbed 'superforecasters') consistently outperformed serving CIA analysts on the same questions, by a meaningful margin. Across all questions, superforecasters' aggregate calibration was 30% better than the best alternative methodology IARPA could field. The book is built around explaining what the superforecasters did differently, with case studies from the tournament archive.
## What the superforecasters actually doTetlock and Gardner's 10-commandment framework (Chapter 12 + appendices) captures the practical core. The shorter version is that superforecasters habitually:
- Break vague questions into clearer sub-questions with measurable answers. - Anchor on base rates from comparable historical cases before considering the specific case. - Express probability estimates as percentages rather than vague verbal labels ('likely', 'very probable'). - Update beliefs in proportion to the strength of new evidence, neither over-reacting nor anchoring stubbornly. - Recognise their own biases - particularly over-confidence, anchoring on initial estimates, and motivated reasoning. - Work in teams when possible - aggregated forecasts beat individuals at the calibration margin. - Treat forecasting as a practice with feedback loops, not a one-shot prediction to be defended later.None of these habits is exotic. The contribution of the book isn't novelty - it's that Tetlock's team measured the contribution of each habit against forecast accuracy across thousands of forecasts, so the claim that these habits matter rests on data rather than plausibility.
## Where the book is strongestThree things set Superforecasting apart from the typical decision-making book.
Rigour. The Good Judgment Project was a competitive research programme with measurable outputs. Most decision-making books synthesise existing literature; Superforecasting reports on a specific research project where the central claims were tested against tournament results. That difference shows in the confidence of the practical claims.
Actionability. The 10-commandment framework is implementable by individuals without a research team. You can start practising it on personal forecasts - GDP next quarter, election outcomes, the next round of layoffs at your employer - and measure your calibration improvement over weeks. Few decision-making books offer that level of operationalisation.
Honest scope. Tetlock is careful not to overclaim. The findings apply to short-horizon (months to 18 months) forecasts on political and economic questions where ground truth is measurable. He doesn't extend the claims to long-range forecasting or to stockmarket-tip territory. That restraint makes the rest of the claims more credible.
## Where it falls shortThree caveats worth noting before buying.
The narrative-heavy first third can feel slow if you came for the practical material. Tetlock and Gardner spend time on the history of forecasting research, the Good Judgment Project's design, and several extended case studies before the 10-commandment framework arrives in Chapter 12. Readers who prefer the practical content can jump to Chapter 11-12 and back-fill the context as needed.
The findings explicitly cap at 12-18 month horizons. The book doesn't claim that superforecasting transfers to long-range forecasting (5-10 year predictions), and it's careful not to suggest the framework works for asset-pricing or any environment where market prices already incorporate forecasters' aggregate beliefs.
Organisational forecasting is under-covered. The book is mostly about individual forecasters and small teams. How to embed superforecasting habits inside a larger organisation - corporate planning, government policy units, intelligence analysis - is a separate practical question the book gestures at but doesn't fully address.
## Who should read it 1. **Anyone whose work involves estimating uncertain outcomes** - Investors, analysts, founders, product managers, policy researchers, intelligence analysts - all professions where forecast quality affects performance. The 10-commandment framework is directly applicable. 2. **Anyone who's read Thinking, Fast and Slow and wants the next step** - Kahneman's book is diagnostic (here are the biases that distort your forecasts). Superforecasting is prescriptive (here are the habits that improve your forecasts). Read in that order. 3. **Anyone leading a team that makes recurring forecasts** - The team-level findings (aggregated forecasts beat individuals; calibrated probabilities beat verbal labels; feedback loops matter) are usable without the full research context. Worth reading and putting on a team's recommended-reading list. 4. **Skip if you want long-range or markets-based forecasting** - The book is honest about scope. For long-range thinking, Tetlock's earlier Expert Political Judgment (2005) is closer; for markets, Taleb's Fooled by Randomness or The Black Swan address the asset-pricing question this book deliberately avoids. ## Frequently asked questions **Q: What is Superforecasting about?** Philip Tetlock and Dan Gardner's 2015 book reports on the Good Judgment Project - a 4-year IARPA-funded forecasting tournament that identified 'superforecasters', the top 2% of volunteer forecasters who consistently outperformed serving CIA analysts on short-horizon geopolitical and economic questions. The book extracts the habits and practices that distinguished superforecasters from average forecasters into a practical 10-commandment framework usable by individuals and small teams. **Q: Is Superforecasting still relevant in 2026?** Yes. The book is now 10 years old, so some specific contemporary examples have aged (the 2014 Crimean annexation case, for instance). The 10-commandment framework and the underlying research findings remain current because the cognitive habits that produce better forecasts haven't changed. The Good Judgment Open platform that grew out of the original project is still running and absorbs new research findings. **Q: How does Superforecasting compare to Thinking, Fast and Slow?** Kahneman's book is diagnostic - it catalogues the cognitive biases that distort human judgment. Superforecasting is prescriptive - it identifies the habits that improve forecast calibration. They're complements, not substitutes. Read Thinking, Fast and Slow first to understand the biases; read Superforecasting second to learn what to do about them. **Q: Can I become a superforecaster?** Per Tetlock and Gardner's central claim, yes - calibration in short-horizon forecasting is a learnable skill rather than a fixed trait. The book's research showed measurable improvement in volunteer forecasters who explicitly trained on the 10-commandment framework. The Good Judgment Open platform (free, online) is the recommended starting point for practising on real questions with feedback. **Q: Does Superforecasting say anything about investing?** Deliberately not much. Tetlock is explicit that the Good Judgment Project's findings don't extend to asset markets, where prices already incorporate the aggregate of forecasters' beliefs and beating that aggregate is structurally harder. For investing applications, Taleb's Fooled by Randomness and The Black Swan are closer-fit - or our recency bias and Monte Carlo thinking explainers for the specific failure modes. ## Related guides **Affiliate disclosure** We may earn a commission on Amazon UK purchases made via the link above. The disclosure doesn't change the editorial assessment - this is a research-led review based on the book's published content and the publicly-documented Good Judgment Project results. See our Affiliate Disclosure for the wider position. **Verdict:**Superforecasting is the most rigorous book in the probabilistic-thinking canon. Unlike most decision-making books that synthesise existing research into a narrative, Superforecasting reports on a specific 4-year research programme - the Good Judgment Project, funded by IARPA - that competitively measured what makes some forecasters dramatically more accurate than others. The findings are practical, actionable, and grounded in data rather than anecdote.
The book's central claim is that calibration in short-horizon forecasting (months to 18 months out, on political and economic questions) is a learnable skill rather than a fixed trait. The 'superforecasters' - the top 2% of the Good Judgment Project's volunteer forecaster pool - shared a specific set of habits that the rest of us can practise. That makes Superforecasting useful in a way Kahneman's Thinking, Fast and Slow (more diagnostic) and Taleb's Fooled by Randomness (more philosophical) aren't.
Where it falls short: the narrative-heavy first third can feel slow before the practical material lands, and the findings explicitly cap at ~12-18 month horizons - this isn't a guide to long-range forecasting. For most readers, neither caveat materially detracts. Recommended.
--- # Vanguard vs InvestEngine: UK 2026 Comparison Source: https://expectedvalue.co.uk/compare/vanguard-vs-investengine-uk-2026/Vanguard and InvestEngine are two of the UK's lowest-cost investment platforms, and in 2026 the gap between them comes down to one number: the platform fee. Vanguard now charges a minimum of £4 a month on smaller self-managed accounts, while InvestEngine charges nothing to hold your investments at all. This comparison weighs the fees, the choice of investments, and who each platform actually suits.
**Answer** For a small UK portfolio, InvestEngine is the cheaper platform: it charges no platform fee, while Vanguard now applies a £4/month minimum (£48/year) on self-managed accounts under £32,000. Choose Vanguard if you want its own LifeStrategy funds; choose InvestEngine for whole-market, commission-free ETFs. **TL;DR** InvestEngine wins on cost and ETF choice for most DIY investors. Vanguard makes sense if you want its own LifeStrategy/Target Retirement funds or have a large portfolio that benefits from the £375 fee cap. **Items** - **InvestEngine** - **Vanguard UK Investor** ## How do the platform fees compare?Both platforms are cheap by UK standards, but they price holding your money very differently. Vanguard charges 0.15% a year on your investments, capped at £375 - but from 31 January 2025 it introduced a minimum account fee of £4 a month (£48 a year) on the self-managed Stocks and Shares ISA, SIPP and General Investment Account when your holdings are below £32,000. InvestEngine charges no platform fee at all on its DIY accounts: you pay only the ongoing charge of whichever ETFs you buy, which start from around 0.03% a year.
On top of the platform layer, both pass through the funds' own charges. Vanguard's index funds and ETFs are low-cost; InvestEngine's whole-market ETF menu includes equally cheap options. Tax wrappers matter too - the annual ISA allowance (the tax-free amount you can pay into a Stocks and Shares ISA each year) is the same whichever platform holds it.
## Which is cheaper for a small portfolio?For anyone starting out, InvestEngine has the clear cost edge. On a £5,000 ISA, Vanguard's £4/month minimum works out at £48 a year - just under 1% of the pot swallowed by the platform fee alone, before fund charges. The same £5,000 on InvestEngine costs nothing at the platform level; you pay only the ETF's ongoing charge, often £2-£5 a year in total.
The maths only swings towards Vanguard at the other end. Once a self-managed pot passes £32,000 the percentage fee replaces the minimum, and above £250,000 Vanguard's £375 cap means very large portfolios pay a shrinking percentage. InvestEngine's flat zero, though, stays zero at every size.
## What can you actually invest in?This is where the platforms genuinely differ, not just on price. Vanguard is a walled garden: you can only buy Vanguard's own funds and ETFs. For many investors that is a feature, not a bug - the LifeStrategy and Target Retirement ranges bundle a diversified portfolio into a single fund you never have to rebalance.
InvestEngine is the opposite. It offers ETFs only - no mutual funds, investment trusts or individual shares - but those ETFs span the whole market and many providers, so you can build a portfolio from iShares, Vanguard, Invesco and others side by side. It adds tools Vanguard lacks, including auto-investing Savings Plans and fractional ETF purchases. The trade-off: you assemble and rebalance the portfolio yourself, since InvestEngine's ready-made Managed Portfolios are currently closed to new clients.
## Who should choose which platform? ## Frequently asked questions **Q: Is InvestEngine cheaper than Vanguard?** For most investors, yes. InvestEngine charges no platform fee on its DIY accounts, while Vanguard charges 0.15% a year with a £4/month minimum on self-managed accounts under £32,000. You pay only ETF charges on InvestEngine, which start from around 0.03% a year. **Q: Can I hold a Stocks and Shares ISA on both?** Yes. Both InvestEngine and Vanguard offer a Stocks and Shares ISA and a SIPP, and both are FCA-regulated with FSCS protection. The tax-free ISA allowance is set by HMRC and is identical whichever platform you use. **Q: What's the catch with InvestEngine's zero fees?** InvestEngine only offers ETFs - no mutual funds, investment trusts or individual shares - and its ready-made Managed Portfolios are currently closed to new clients. If you want Vanguard's own LifeStrategy funds or a managed option, that's a reason to look at Vanguard instead. **Q: Did Vanguard increase its fees in 2026?** Vanguard introduced a minimum account fee of £4 a month (£48 a year) on its self-managed ISA, SIPP and General Investment Account for holdings below £32,000. The headline 0.15% rate and £375 cap are unchanged; the minimum mainly affects smaller portfolios.Related: if you want a commission-free trading app rather than a fund platform, see our Trading 212 vs eToro comparison.
**The bottom line** For most DIY investors - especially smaller portfolios - InvestEngine's zero platform fee and whole-market ETF choice make it the stronger pick. Choose Vanguard for its own all-in-one funds or a large, fee-capped pot. --- # Best SIPP UK 2026: 10 Providers Compared Source: https://expectedvalue.co.uk/compare/best-sipp-uk-2026/The best SIPP in the UK for 2026 is the one whose fee model fits your pot size and how you invest - there is no single winner. A SIPP (self-invested personal pension, a do-it-yourself pension wrapper you control) charges two layers of cost: a platform fee to the provider and an ongoing charge (OCF, the annual cost of the fund itself) to the fund manager. Get the platform fee wrong and you can hand over hundreds of pounds a year for nothing.
This guide compares ten of the most popular UK SIPP providers on their 2026 charges, and pairs naturally with our Vanguard vs InvestEngine comparison. Contributions attract tax relief at your marginal rate, and the 2025/26 annual allowance is £60,000, as set out in the gov.uk pension tax rules.
**Answer** AJ Bell is the best all-round SIPP for most investors, with a capped 0.25% fee and free drawdown. Interactive investor's flat fee wins for pots above roughly £100,000, while InvestEngine charges £0 if you invest only in ETFs. **Fees current as of June 2026** Charges change often. Hargreaves Lansdown cut its SIPP fund fee from 0.45% to 0.35% on 1 March 2026, and Freetrade revised its plans in January 2026. Always confirm the latest charges on the provider's own fee page before opening or transferring an account. **Items** - **AJ Bell** - **interactive investor** - **InvestEngine** - **Hargreaves Lansdown** - **Vanguard** - **Fidelity** - **Bestinvest** - **Moneybox** - **Wealthify** - **Freetrade** ## How we compared these SIPPsThis comparison is based on each provider's published 2026 fee schedule and independent broker-fee tables, not on personal use of every account. We focus on the platform/custody fee (the charge for holding the account), dealing costs for funds and shares, account minimums, the available investment range, and whether the provider offers flexi-access drawdown. Fund OCFs are extra and depend on what you buy.
## Which SIPP has the lowest fees?It depends entirely on pot size and investments. For a portfolio built only from exchange-traded funds, InvestEngine's £0 platform fee is unbeatable. For a typical mixed portfolio under about £80,000, AJ Bell's capped 0.25% and Vanguard's 0.15% are hard to beat. Above roughly £100,000, interactive investor's flat fee pulls clearly ahead. The table below shows the platform fee alone on a £50,000 pot - fund charges are extra.
**Read the fee carefully** The figures above are the platform fee on a £50,000 pot for illustration only. They exclude fund OCFs, dealing charges and any drawdown costs, and percentage fees rise with your pot while flat fees do not. ## Percentage vs flat fees: which is cheaper?This is the single most important decision. Percentage-fee platforms (Vanguard, AJ Bell, Fidelity, Hargreaves Lansdown) charge a slice of your pot, so they are cheap when the pot is small but grow more expensive as it builds. Flat-fee platforms, chiefly interactive investor, charge the same amount whatever your balance.
The crossover sits at roughly £80,000 to £100,000 for capped-percentage rivals. Below it, a percentage fee usually wins; above it, a flat fee saves more every year your pot grows. On a £250,000 SIPP, interactive investor's roughly £180 a year compares with several hundred pounds on a percentage platform.
## How do I transfer an existing pension into a SIPP?Most transfers are started from the new provider's side: you open the SIPP, complete a transfer request naming your old pension, and the provider handles the rest, usually electronically within a few weeks. Cash transfers are simplest; an in-specie transfer moves your existing investments without selling them.
Before moving, check three things: whether the old scheme charges an exit fee, whether you would lose any valuable guarantees (such as a guaranteed annuity rate or protected pension age), and whether a defined-benefit pension is involved, in which case regulated advice is legally required above £30,000. The government-backed MoneyHelper service explains the safeguards.
## Best for / Skip if ## Frequently asked questions **Q: What is the cheapest SIPP in the UK?** For ETF-only investors, InvestEngine charges £0 in platform fees. For mixed portfolios, AJ Bell (0.25%, capped on shares) and Vanguard (0.15%) are cheapest on smaller pots, while interactive investor's flat fee is cheapest above roughly £100,000. **Q: Can I have more than one SIPP?** Yes. There is no limit on the number of SIPPs you can hold, though your total pension contributions across all schemes count towards the same £60,000 annual allowance for 2025/26. **Q: At what age can I access my SIPP?** You can normally take money from a SIPP from age 55, rising to 57 from April 2028. Up to 25% can usually be taken tax-free, subject to the Lump Sum Allowance of £268,275. **Q: Is my money safe in a SIPP?** SIPP providers are regulated by the FCA and eligible investments are covered by the Financial Services Compensation Scheme up to £85,000 per provider. The value of investments can still fall as well as rise. **Q: Do I get tax relief on SIPP contributions?** Yes. Basic-rate relief of 20% is added automatically, and higher or additional-rate taxpayers can claim the rest through self-assessment, subject to the annual allowance and your earnings. **Not financial advice** This article is general information, not personal financial advice. Pension and tax rules change and depend on your circumstances. Your capital is at risk and investments can fall in value. Consider regulated advice before transferring, especially from a defined-benefit pension.Further reading: gov.uk - Tax on your private pension contributions, MoneyHelper - Pensions and retirement. Provider fees were checked against each platform's published 2026 charges and independent broker-fee tables in June 2026.
--- # Kalshi vs Polymarket vs Manifold (UK & US 2026) Source: https://expectedvalue.co.uk/compare/kalshi-vs-polymarket-vs-manifold/Kalshi, Polymarket and Manifold are the three platforms most users mean when they say "prediction markets" in 2026 - but they serve different audiences, run under different legal regimes, and use different settlement currencies. Kalshi is the US-regulated, CFTC-overseen real-money option - now open to traders in roughly 143 countries, though the UK is on its excluded list. Polymarket is the global crypto-settled option (banned to US residents). Manifold is the play-money platform with the most market-creation freedom. This comparison covers the regulatory status, market depth, fees, withdrawal mechanics, and UK / US availability of each, so you can pick the platform that matches your goals.
**TL;DR** Pick Kalshi if you're in the US or one of the ~140 other countries it now serves (the UK is excluded) and want regulated real-money exposure. Pick Polymarket if you're a UK or non-US resident wanting deep liquidity + broad market selection, and you're comfortable with crypto wallets. Pick Manifold if you want play-money forecasting with no regulatory or KYC friction - the best place to develop calibration skills without financial risk. For background on how prediction markets work in the first place, see our prediction markets explainer. **Items** - **Kalshi** - price: 0 - **Polymarket** - price: 0 - **Manifold Markets** - price: 0 ## Why does regulatory status matter so much here?Prediction markets sit at an unusual intersection of derivatives trading, gambling law, and election interference concerns. The regulatory regime each platform operates under is the single biggest differentiator between Kalshi, Polymarket and Manifold - and it determines who can legally use the platform from which jurisdiction.
Kalshi (US-regulated). Kalshi operates as a Designated Contract Market (DCM) regulated by the US Commodity Futures Trading Commission. This is the same regulatory category as the Chicago Mercantile Exchange. Markets are listed as event contracts under CFTC rules. The regulated status brings FDIC-insured cash balances, mandatory KYC, ACH withdrawals, and a 25,000 USD per-market position cap. In October 2025 Kalshi expanded internationally and now accepts members in roughly 143 countries, all trading the same USD order books - but the UK is on its list of excluded jurisdictions (alongside Ireland, France and Italy), so UK residents are still blocked at account creation.
Polymarket (no US; available to UK and most non-US users). Polymarket previously operated unregulated US markets until January 2022, when the CFTC fined them $1.4M and required them to block US users. Since then, the platform has operated as a crypto-native (USDC-settled) prediction market available to non-US residents including UK users. There is no formal regulator overseeing Polymarket - it sits in a grey legal area outside the US that most jurisdictions tolerate but don't license. UK users can legally use Polymarket but the platform is not FCA-authorised.
Manifold (play-money, legal everywhere). Because Manifold's internal currency (Mana) cannot be cashed out to real money, it is treated as a video game or social platform rather than a gambling or derivatives platform. This means Manifold is legal worldwide with no regulatory friction. The trade-off is that there is no financial upside from being right.
## How does market depth and liquidity compare?Liquidity is the practical question that separates the three platforms once you've settled the legal question. The market depth determines how easily you can enter and exit positions without moving the price meaningfully.
Polymarket. By volume traded, Polymarket is the deepest prediction market in 2026 - top markets routinely settle $10M+ of cumulative volume, and large active markets have $500K-$2M of liquidity available within 1% of the mid-market price. The crypto-native settlement and lack of position caps attract larger traders. The 2024 US election cycle alone saw over $3.7 billion of volume on Polymarket according to publicly reported aggregate data.
Kalshi. Liquidity is meaningfully less deep than Polymarket but still credible for retail-scale trading. Top markets see $50K-$500K of typical depth within 1% of mid. The 25,000 USD per-market cap reduces the contribution of whale activity. Kalshi's market-maker programme provides baseline liquidity even on lower-volume markets.
Manifold. Shallowest of the three - top markets have a few hundred thousand Mana of liquidity, but most markets sit at low hundreds of Mana. Spreads can be wide on lower-volume questions. The play-money nature changes the incentive structure: users are forecasting for reputation and calibration practice, not for liquidity provision.
The practical guidance: if you want to deploy more than 5,000 USD or equivalent on a single position, Polymarket is the platform that can absorb it without significant slippage. Kalshi works for positions under 5,000 USD on most markets. Manifold is not the right tool for any size-driven thesis.
## What does each platform charge in fees?Kalshi: trading fees apply on profitable trades (the loser pays nothing). Typical fees are 1-2% of net profit on resolved positions plus a small per-contract fee (~$0.01-0.04). Withdrawals via ACH are typically free; wire transfers carry a $25-35 fee. The all-in cost on a profitable round-trip is meaningfully higher than Polymarket.
Polymarket: no platform fees on trades. The only cost is the bid/ask spread (which can be 1-3% on liquid markets, more on illiquid ones). USDC deposits via on-chain transfer cost gas fees (typically $1-3 on Polygon, the chain Polymarket uses). USDC withdrawals are similarly gas-fee-only. For active traders, this is the cheapest of the three.
Manifold: no fees at all. Mana is a closed-loop currency and Manifold operates as a non-profit - there is no economic incentive for fees. Users get free Mana on signup and can earn more by being calibrated forecasters or by creating popular markets.
The honest framing: Polymarket is the cheapest place to actually trade for active users, but Kalshi's fees are the cost of CFTC regulation and FDIC-insured balances - which is genuinely valuable for risk-averse users. Manifold's free model is structurally different because nothing has financial value.
## How does withdrawal and cash-out actually work?Kalshi (USD via ACH): standard US bank ACH transfer, typically 1-3 business days. KYC verification (passport or driving licence + SSN + proof of address) is mandatory at signup. Wire transfers are available for larger withdrawals at a $25-35 fee. The withdrawal experience is identical to any regulated US brokerage.
Polymarket (USDC via on-chain transfer): withdrawal is an on-chain transaction from Polymarket's smart contracts to your USDC wallet (typically on Polygon, optionally bridged to Ethereum mainnet or other chains). Transactions are confirmed in 5-30 seconds and cost $1-3 in gas. From there, the USDC can be swapped to GBP via a UK-licensed crypto exchange (Kraken, Coinbase) and withdrawn to a UK bank account via Faster Payments. Total cash-out timeline is typically same-day for crypto-fluent users; first-time users may need 2-3 days to set up the off-ramp.
Manifold (no cash-out): Mana cannot be converted to real money. The platform is closed-loop by design. Some Manifold leaderboard prizes (sponsored prizes for forecasting tournaments) pay in real money, but they're occasional events rather than a standard withdrawal route.
## Which prediction market should UK users actually pick?For UK residents in 2026, the practical decision tree:
Want real-money exposure + comfortable with crypto wallets → Polymarket. This is the only legal real-money option for UK users. The USDC settlement mechanism and on-chain custody are an adjustment from a regulated brokerage but are well-documented and reliable. UK users are an explicitly served market segment on Polymarket - the platform's UK user base is substantial.
Want real-money exposure + uncomfortable with crypto → no good option in 2026. Kalshi is US-only. Smarkets and Betfair Exchange offer political markets to UK users but they're styled as sportsbook exchanges rather than prediction markets, with narrower question scope and worse liquidity on most non-sports events. Our prediction markets explainer covers the smaller UK-available options.
Want forecasting practice without financial risk → Manifold. The play-money structure makes Manifold the ideal place to develop calibration skills before deploying real money on Polymarket. Several published research projects use Manifold as a forecasting-skill training environment specifically because of this.
## Frequently asked questions **Q: Can UK residents legally use Kalshi?** No - though not because Kalshi is US-only. Kalshi expanded internationally in October 2025 and now accepts members in roughly 143 countries, but the UK is on its list of excluded jurisdictions (alongside Ireland, France and Italy), so UK residents are still blocked at the KYC stage. The closest UK-legal alternative for regulated real-money prediction-market exposure is the Smarkets and Betfair Exchange political markets, which cover a narrower question set but are FCA-regulated. **Q: Is Polymarket actually legal for UK users in 2026?** Yes for individual users. Polymarket is not specifically licensed by the FCA but operates in a grey area that UK regulators tolerate - similar to most non-UK crypto platforms. The platform's user agreement permits UK residents to register and trade. UK users should declare any prediction-market gambling winnings to HMRC as part of standard gambling-tax accounting (gambling winnings are not taxable for UK individuals, but the line between gambling and trading on prediction markets can be questioned for institutional or full-time activity). **Q: Can I make money on Manifold?** Not directly - Mana cannot be cashed out. However, Manifold occasionally runs sponsored forecasting tournaments with real-money prizes (typically $1,000-$10,000 prize pools), and the platform's reputation rankings can lead to job offers in forecasting or quantitative research roles. The primary value of Manifold is calibration practice, not financial return. **Q: Which platform has the deepest liquidity for political markets?** Polymarket, decisively. The 2024 US election cycle saw $3.7B+ of volume on Polymarket alone according to publicly reported aggregate data, with individual markets like 'Trump wins 2024 election' holding $100M+ of cumulative volume. Kalshi's political markets are smaller (typically $50K-$500K depth per market) due to the 25,000 USD per-market cap. Manifold's political markets are reputation-driven and shallow. **Q: How accurate are these platforms historically?** Prediction markets in general have well-documented forecasting accuracy that often beats expert pundits. The Polymarket and Kalshi 2024 US election markets correctly priced Trump as the favourite weeks before the election outcome. Manifold's calibration scores (probability stated vs frequency of yes-resolution) are typically within 5% of perfect calibration - consistent with the published superforecasting literature. The honest framing: markets are accurate aggregations of available information, not magic crystal balls. **Q: What's the minimum deposit on each platform?** Kalshi: no formal minimum but practical first deposit is $50-$100 via ACH. Polymarket: no formal minimum but practical first deposit is $50-$100 USDC; lower amounts get eaten by gas fees. Manifold: zero - signup grants you free Mana, no deposit required. **Q: Do any of these platforms accept GBP deposits directly?** No. Kalshi accepts USD only. Polymarket accepts USDC only (you can swap GBP → USDC via a UK exchange like Kraken or Coinbase). Manifold uses Mana (play-money, no fiat deposits). For UK users wanting to fund Polymarket, the typical route is: GBP via Faster Payments → Coinbase/Kraken → buy USDC → withdraw USDC to Polygon wallet → deposit to Polymarket. The total round-trip is typically same-day for set-up users. ## Related guides --- # Trading 212 vs eToro: UK 2026 Comparison Source: https://expectedvalue.co.uk/compare/trading-212-vs-etoro-uk-2026/Trading 212 and eToro are two of the UK's most popular commission-free investing apps, and both now offer a Stocks and Shares ISA. They look similar on the headline - free stock trades - but the small print and the feature sets pull them in different directions. This comparison breaks down the real costs, what each does best, and who should pick which.
**Answer** For UK ISA investing, Trading 212 is the cheaper, simpler choice: commission-free with only a 0.15% FX fee and no withdrawal fee. eToro is commission-free too and uniquely offers copy trading, but its ISA (powered by Moneyfarm) carries a custody fee capped at £45 a year and higher FX costs unless you use a GBP account. **TL;DR** Trading 212 wins on cost and simplicity for most UK investors. Choose eToro if copy trading - mirroring other investors' portfolios - is the feature you actually want. **Items** - **Trading 212** - **eToro** ## How do the fees compare?Both platforms charge zero commission on real stock and ETF trades, so the difference is in the edges. Trading 212 charges no platform fee, no ISA fee and no inactivity fee; its only charge is a 0.15% FX fee when you buy a share priced in a currency other than sterling. eToro is also commission-free on stocks, but its Stocks and Shares ISA (powered by Moneyfarm) carries a custody fee capped at £45 a year, and its currency-conversion cost has historically been around 0.50-0.70% - far higher than Trading 212's - unless you use eToro's newer GBP account or a premium membership tier.
One cost is identical on both: Stamp Duty of 0.5% on UK-listed shares is a government tax, not a platform charge. The tax-free ISA allowance (the amount you can pay into a Stocks and Shares ISA each tax year) is also the same wherever you invest.
## Which is cheaper for an ISA investor?For a hands-on UK ISA investor, Trading 212 is the lower-friction option. There's nothing to pay to hold the account, the 0.15% FX fee is among the cheapest available, and you can take money out without penalty. eToro's £45-a-year ISA custody cap is easy to overlook but counts against smaller pots, and its conversion costs only match Trading 212's if you deliberately run a GBP account.
If you mostly buy UK-listed shares and rarely withdraw, the gap narrows - but Trading 212 still edges it on transparency, with fewer conditions attached to the headline 'free'.
## What does each platform do best?eToro's standout feature is social investing. CopyTrader lets you automatically mirror the trades of other eToro users, and the platform leans into community, crypto and CFD trading. That breadth is genuinely useful if copying experienced investors or trading crypto alongside shares is what you came for - but it also adds complexity and risk that a long-term ISA investor doesn't need.
Trading 212 is narrower and calmer by design. Its Invest and ISA products focus on straightforward commission-free shares and ETFs, with auto-investing 'Pies' that let you set target allocations and drip-feed them. For buy-and-hold investing inside an ISA, that focus is a strength rather than a limitation.
## Who should choose which platform? ## Frequently asked questions **Q: Is Trading 212 cheaper than eToro?** For most UK investors, yes. Both are commission-free on stocks, but Trading 212 has no platform or custody fee and a 0.15% FX fee, while eToro's ISA has a £45-a-year custody cap and around 0.50-0.70% FX unless you use a GBP account or premium tier. **Q: Do both offer a Stocks and Shares ISA?** Yes. Trading 212 and eToro both offer a commission-free Stocks and Shares ISA, and both are FCA-regulated with FSCS protection. The annual ISA allowance is set by HMRC and is the same on either platform. **Q: What is eToro's main advantage over Trading 212?** Social and copy trading. eToro's CopyTrader lets you automatically replicate other investors' portfolios, and it offers crypto and CFDs. Trading 212 has no copy-trading feature and focuses on straightforward commission-free investing. **Q: Is copy trading on eToro a good idea?** It can be useful for learning or diversifying, but copying another investor doesn't remove risk - past performance doesn't guarantee future returns, and you're still exposed to that strategy's losses. Treat it as one tool, not a shortcut to guaranteed gains.Related: if you want a long-term fund platform rather than a trading app, see our Vanguard vs InvestEngine comparison.
**The bottom line** For most UK ISA investors, Trading 212's lower fees, penalty-free withdrawals and simple buy-and-hold focus make it the stronger pick. Choose eToro if copy trading or crypto is genuinely what you want. --- # Thinking, Fast and Slow vs Thinking in Bets Source: https://expectedvalue.co.uk/compare/thinking-fast-and-slow-vs-thinking-in-bets/ **Answer** Read Thinking, Fast and Slow first if you want the foundational psychology of how humans actually make decisions - it's the academic primer that almost every modern decision-making book builds on. Read Thinking in Bets first if you want practical decision-making frameworks you can apply to specific real-world choices (poker-derived but generalisable to business, careers, investing). Most readers benefit from reading both in this order: Kahneman provides the why; Duke provides the how. They cover meaningfully different ground despite both being labelled 'decision making'.Daniel Kahneman's Thinking, Fast and Slow (2011) and Annie Duke's Thinking in Bets (2018) are the two most-recommended books in modern decision-making literature. They are not substitutes - they serve different reader needs. This guide compares them honestly so you can choose which to read first based on what you actually want to get out of the time.
**Note on this comparison** Both books are widely available in UK retail (Waterstones, Amazon UK, most bookshops). Kahneman's book is the more demanding read at ~500 pages; Duke's is shorter and lighter at ~250 pages. If you're a slow reader who finishes one book a year, this comparison matters more than if you're going to read both in sequence. See our individual Thinking, Fast and Slow guide and Thinking in Bets guide for the standalone treatments. **Items** - **Thinking, Fast and Slow (Daniel Kahneman, 2011)** - **Thinking in Bets (Annie Duke, 2018)** ## What each book actually deliversThinking, Fast and Slow gives you a mental model. By the end of it, you understand that the human brain operates in two modes (Kahneman's metaphorical 'System 1' fast-and-automatic and 'System 2' slow-and-deliberate), that loss aversion makes losses feel ~2x as painful as equivalent gains, that anchoring effects can shift your judgement just by exposure to a number, that the planning fallacy means you'll always underestimate how long a project takes, and dozens of related findings. Each is supported by published experimental evidence. The mental model is the deliverable.
Thinking in Bets gives you a decision process. By the end of Duke's book, you have specific tools: the separation of 'decision quality' from 'outcome quality' (you can make the right decision and still get a bad result; you can make the wrong decision and get a good result; quality your decisions need to be judged on the inputs, not just the outputs). The 'truth-seeking group' framework for making better decisions in teams. The 'mental time-travel' technique for stress-testing forecasts. The 'pre-mortem' for surfacing risks. The process is the deliverable.
This is why they're not substitutes. Kahneman tells you what your brain is doing; Duke tells you what to do about it.
## Who should read each first ## Where they overlapBoth books address the gap between optimal decision-making and what humans actually do. The overlap is real:
The difference is depth vs application. Kahneman documents the phenomena exhaustively with experimental evidence; Duke takes the phenomena as given and builds tools to work around them.
## Where they divergeKahneman is academic; Duke is practitioner. Kahneman's chapters are organised by phenomenon (anchoring, the planning fallacy, the affect heuristic); Duke's are organised by problem (how to make better decisions in groups, how to learn from bad outcomes, how to forecast under uncertainty). The structure follows the reader-need difference.
Kahneman doesn't address poker; Duke leans on it heavily. If poker examples don't land for you, Duke's book becomes harder work. Kahneman's examples are domain-agnostic (medical diagnosis, investment, project planning, courtroom decisions).
Kahneman is more cautious about prescriptions; Duke is more confident. Kahneman frequently writes 'we don't yet know how to debias ourselves at scale'; Duke is more willing to prescribe specific techniques. Both positions are defensible - Kahneman is more accurate about the state of the academic evidence; Duke is more useful for someone needing to make a decision this week.
Kahneman has been partially affected by the replication crisis; Duke largely hasn't. Several of Kahneman's pre-2015 priming-effect studies didn't replicate in subsequent independent attempts. Kahneman himself acknowledged this in a 2017 open letter. The two-system framework + loss aversion + anchoring effects have all replicated robustly; some of the more specific findings haven't. Duke's framework, built on practitioner wisdom and game-theory rather than experimental psychology, is less exposed to this issue.
## The 'both' recommendationFor most readers, reading both books in sequence (Kahneman first, Duke second) delivers more than either book alone. Why:
Reading Kahneman first gives you the vocabulary and the mental model. When Duke references loss aversion or outcome bias in passing, you already know what she means and you can evaluate her practical framework critically. You don't have to take her on faith.
Reading Duke second gives you the practical bridge from 'I know about biases' to 'I have a process for handling them'. Many readers finish Thinking, Fast and Slow and don't change their behaviour - the book is intellectually compelling but doesn't tell you what to DO. Duke's book is the missing 'what to do'.
The two together take roughly 30-40 hours of reading for an average UK adult reader. That's a meaningful time investment, but the marginal utility of the second book after reading the first is high. If you're going to invest the time in the first book, the second is the natural completion.
## Frequently asked questions **Q: Are Thinking, Fast and Slow and Thinking in Bets contradictory?** No - they're complementary. Both treat human cognitive biases as real and consequential. Kahneman documents the phenomena; Duke builds practical frameworks for working around them. The two books make different methodological choices (academic vs practitioner) but reach broadly aligned conclusions about how humans should structure decisions. **Q: Can I skip Thinking, Fast and Slow if I read Thinking in Bets?** You can, but you'll get less out of Duke. Several of Duke's references (loss aversion, outcome bias, the affect heuristic) assume reader familiarity with the Kahneman framework. Skipping Kahneman means trusting Duke's summaries of these concepts rather than understanding them in depth. Worth doing if you have very limited reading time; not worth doing if you're going to read non-fiction regularly. **Q: Which book is better for UK readers specifically?** Both are equally good for UK readers - neither is US-centric in its examples. Kahneman is Israeli-American but the examples span global research. Duke is American but the poker tournaments and business cases are internationally interpretable. Both books are available in Waterstones + Amazon UK + most bookshops in paperback and Kindle. **Q: Has Kahneman's book been undermined by the replication crisis?** Partially. Several of his pre-2015 priming-effect studies didn't replicate in subsequent attempts; Kahneman acknowledged this in a 2017 open letter saying the priming chapter would need revisiting. However, the foundational frameworks (System 1/2, loss aversion, anchoring, the planning fallacy) have all replicated robustly. The book remains the canonical decision-making primer despite the partial revisions. **Q: What should I read after both of these?** Several natural next steps. Superforecasting by Philip Tetlock + Dan Gardner is the empirical companion - 'who are the best decision-makers and how do they think?' Tetlock's research provides the second-order evidence. Probabilistic Thinking covers the broader framework these books fit into. Black Swan Events by Nassim Taleb covers the extreme-event blind-spots that Kahneman + Duke under-emphasise. Status Quo Bias covers one specific bias in depth with UK-specific applications. ## Related guides --- # Fooled by Randomness: The Hidden Role of Chance Source: https://expectedvalue.co.uk/product/fooled-by-randomness-taleb/ Nassim Taleb's 2001 book on luck in success and how survivorship bias distorts judgment of skill in finance + business. First Incerto text. **Specifications** - **Author**: Nassim Nicholas Taleb - **Publisher**: Random House (US) / Penguin UK - **Year**: 2001 (revised 2005) - **Pages**: 316 - **ISBN-13**: 978-0812975215 - **Format**: Paperback / Kindle / Audiobook - **Series**: Incerto (Book 1 of 5) **Pros** - Foundational case for the role of luck in success - Highly readable - more memoir + market anecdote than dry treatise - Concepts (survivorship bias, alternative histories, rare events) became part of the broader financial vocabulary - Useful corrective to skill-narratives in finance, business, and self-help **Cons** - Tone is famously combative - some readers find it grating before reaching the analytical core - Light on quantitative method - more conceptual than prescriptive - The Black Swan (2007) sharpens and expands the central ideas; some readers prefer to start there # Fooled by Randomness: The Hidden Role of Chance Nassim Taleb's 2001 book on luck in success and how survivorship bias distorts judgment of skill in finance + business. First Incerto text. **Affiliate disclosure** We may earn a commission if you buy through links on this page - this never changes the price you pay or our editorial verdict. - **Author**: Nassim Nicholas Taleb - **Publisher**: Random House (US) / Penguin UK - **Year**: 2001 (revised 2005) - **Pages**: 316 - **ISBN-13**: 978-0812975215 - **Format**: Paperback / Kindle / Audiobook - **Series**: Incerto (Book 1 of 5) --- # Superforecasting: The Art and Science of Prediction Source: https://expectedvalue.co.uk/product/superforecasting-tetlock-gardner/ Philip Tetlock + Dan Gardner's 2015 book on Good Judgment Project research into what makes forecasters more accurate. Foundational text. **Specifications** - **Authors**: Philip E. Tetlock, Dan Gardner - **Publisher**: Random House (US); Crown / Penguin (UK) - **Year**: 2015 - **Pages**: 352 - **ISBN-13**: 978-0804136716 - **Format**: Paperback / Kindle / Audiobook - **Origin**: Good Judgment Project (IARPA-funded, 2011-2015) **Pros** - Grounded in 4 years of IARPA-funded competitive-forecasting research, not anecdote - Identifies concrete traits and habits of the top 2% forecasters ('superforecasters') - Practical 10-commandment training framework usable by individuals - Honest about limits: short-horizon political/economic forecasts only **Cons** - Some readers find the narrative-heavy first half slow before the practical material lands - Findings cap at ~12-18 month horizons; doesn't claim to extend to long-range - Limited coverage of organisational vs individual forecasting # Superforecasting: The Art and Science of Prediction Philip Tetlock + Dan Gardner's 2015 book on Good Judgment Project research into what makes forecasters more accurate. Foundational text. **Affiliate disclosure** We may earn a commission if you buy through links on this page - this never changes the price you pay or our editorial verdict. - **Authors**: Philip E. Tetlock, Dan Gardner - **Publisher**: Random House (US); Crown / Penguin (UK) - **Year**: 2015 - **Pages**: 352 - **ISBN-13**: 978-0804136716 - **Format**: Paperback / Kindle / Audiobook - **Origin**: Good Judgment Project (IARPA-funded, 2011-2015) --- # The Signal and the Noise: Why Predictions Fail Source: https://expectedvalue.co.uk/product/signal-and-noise-nate-silver/ Nate Silver's 2012 book on probabilistic forecasting across weather, sports, poker, economics, politics. Builds the case for Bayesian reasoning. **Specifications** - **Author**: Nate Silver - **Publisher**: Penguin - **Year**: 2012 - **Pages**: 544 - **ISBN-13**: 978-0143125082 - **Format**: Paperback / Kindle / Audiobook - **Author background**: Founder of FiveThirtyEight; former PECOTA baseball forecaster **Pros** - Practical worked examples across genuinely different domains (weather, poker, baseball, terrorism, economics, politics) - Strongest exposition of Bayesian reasoning aimed at non-statisticians in popular nonfiction - Silver's track record on the 2008 and 2012 US elections gave the book real-world authority at publication - Distinguishes genuinely-predictable domains (weather) from genuinely-unpredictable ones (terrorism) **Cons** - 2012 publication date - some specific examples have aged, particularly the 2008/2012 election cases - Silver's own 2016 misfire on Trump's victory raised questions about the book's confident framing - Some readers find Silver's data-journalism prose flatter than Taleb's or Kahneman's # The Signal and the Noise: Why Predictions Fail Nate Silver's 2012 book on probabilistic forecasting across weather, sports, poker, economics, politics. Builds the case for Bayesian reasoning. **Affiliate disclosure** We may earn a commission if you buy through links on this page - this never changes the price you pay or our editorial verdict. - **Author**: Nate Silver - **Publisher**: Penguin - **Year**: 2012 - **Pages**: 544 - **ISBN-13**: 978-0143125082 - **Format**: Paperback / Kindle / Audiobook - **Author background**: Founder of FiveThirtyEight; former PECOTA baseball forecaster --- # Expected Value: Probabilistic Thinking Decisions Source: https://expectedvalue.co.uk/ # Think in Probabilities, Not Predictions Master the mental models that separate good decision-makers from everyone else. From expected value to cognitive biases, learn the frameworks that drive better outcomes in investing, career, and life.## Core Topics ## Featured Articles ## The Expected Value Approach - **No Prior Maths Required** - Every concept is explained with concrete examples, not abstract notation. If you can multiply, you can use expected value. - **Immediately Applicable** - Each article ends with practical takeaways you can apply to real decisions - from insurance purchases to career changes. - **Evidence-Based** - Grounded in decades of research from behavioural economics, decision science, and probability theory. - **Built to Be Re-Read** - These aren't news articles. They're reference material designed to deepen with each reading. --- # About Expected Value Source: https://expectedvalue.co.uk/about/ # About Expected Value Quantitative thinking for everyone - not just quants. **Answer**"It is remarkable how much long-term advantage people like us have gotten by trying to be consistently not stupid, instead of trying to be very intelligent."
- Charlie Munger, attributed at various Berkshire Hathaway annual meetings (CNBC compilation)
Expected Value is a UK site about probabilistic thinking, decision-making, and behavioural finance. We translate the academic ideas behind expected value, calibration, Bayesian thinking and cognitive biases into practical frameworks anyone can use for everyday decisions: investing, career choices, betting, insurance, and the dozens of small calls we all make weekly without realising we are doing maths badly.
## Editorial Standards - **Research-driven** - Every article draws from published research, textbooks, and established sources in probability theory, behavioural economics, and decision science. We cite our sources. - **Accuracy over speed** - We verify mathematical examples, check formulas, and test calculations before publishing. If we're unsure about something, we say so. - **No fake expertise** - This is an editorial resource, not a personal blog from a hedge fund manager. We synthesise expert knowledge to make it accessible - we don't claim credentials we don't have. - **Honest about limitations** - Probabilistic thinking is powerful but not magic. We're upfront about where models break down and where uncertainty is genuinely irreducible. ## What This Site Covers ## How We Make MoneyTransparency matters. This site earns revenue through:
Affiliate links: When we mention investing platforms, books, or tools, we sometimes include affiliate links. If you sign up or purchase through one, we earn a small commission at no extra cost to you. This never influences our editorial content - we recommend based on merit alone.
Display advertising: You may see ads on our pages. We aim to keep these unobtrusive.
We don't accept sponsored content, and no company can pay to influence our recommendations or analysis.
**Disclaimer** This site is not financial advice. The ideas here are educational frameworks for thinking - always do your own research and consult a qualified advisor before making investment decisions. --- # Probability Tools & Calculators Source: https://expectedvalue.co.uk/tools/ # Tools & Resources Platforms, tools, and books we actually use - selected for evidence-based thinkers, not affiliate commissions. **Answer**Expected Value publishes free, browser-based calculators for the most common probabilistic decisions: expected value, Kelly criterion bet sizing, Bayesian updates, decision-tree EV scoring, and probability-vs-odds conversions. No sign-up, no data leaves your browser, and the source maths is documented on each tool's page so you can sanity-check what it is doing.
**Affiliate Disclosure** Some links on this page are affiliate links. If you sign up or purchase through them, we may earn a small commission at no extra cost to you. This never influences what we recommend - we apply the same evidence-based criteria to our tool selection as we do to everything else on this site. Full details on our affiliate disclosure page.There are thousands of investing platforms, productivity apps, and books competing for your attention. Most "best of" lists rank them by commission rate, not quality.
We take a different approach. Everything on this page meets three criteria:
If something stops meeting these criteria, we remove it. If we find something better, we add it.
## Interactive calculatorsFree and browser-based - drag the inputs and the maths updates live. Each calculator sits inside the guide that explains it, with the formula shown in the open so you can sanity-check what it does. Nothing is sent anywhere; it all runs on your device.
## Investing PlatformsThe single most important variable in platform selection is fees. A 1% annual fee difference compounds into tens of thousands of pounds over a lifetime of investing. Every platform below charges significantly less than the industry average - that's the minimum bar for inclusion.
Which one suits you depends on your experience level and what you're investing in.
**How to choose**None of these platforms are perfect for everyone. Your choice should depend on your portfolio size, investment style, and which tax wrappers you need (ISA, SIPP, GIA). The expected value calculation is straightforward: pick the platform where total annual costs (platform fee + trading fees + FX fees) are lowest for your situation.
## Decision-Making ToolsReading about probabilistic thinking is a good start. But calibration - the ability to accurately assess how confident you should be - only improves through deliberate practice. These tools help you move from theory to skill.
Sometimes the best decision tool is a well-structured spreadsheet. We're building a library of templates for common expected value calculations:
These will be available as free downloads as we develop them. In the meantime, the concepts behind each are covered in our Fundamentals articles.
## Books Five books that genuinely changed how we think about decisions, risk, and uncertainty. Not a "top 100" list - just the ones we keep coming back to. **Important Disclaimer** This site is not financial advice. Everything here is educational - frameworks for thinking, not instructions for acting. Always do your own research and consider consulting a qualified financial adviser before making investment decisions. --- # How We Review: Our Methodology Source: https://expectedvalue.co.uk/methodology/ # How We Review: Our Methodology **Answer**Every book and tool reviewed on Expected Value is read or used by us in full before publishing. Recommendations are graded on three axes: how rigorous the underlying probability framework is, how usable the ideas are for non-academic readers, and how well the writing holds up to a critical reading. We disclose where we benefit financially from a recommendation (affiliate links, where used) and we do not accept payment to feature a book or tool.
This page explains how we research, write, and verify content on expectedvalue.co.uk. We publish our methodology because the value of the content depends on whether you trust the process behind it.
## What is our editorial approach?We write from a research-led editorial perspective rather than claiming first-party hands-on experience we don't have. For probability and behavioural-economics claims specifically, we cite peer-reviewed behavioural-finance research (Kahneman/Tversky, Thaler, Laibson) and ground examples in real-world domains rather than abstract scenarios.
For every recommendation, we cite the source: independent published reviews, manufacturer specifications, regulatory filings, peer-reviewed research where relevant. If we can't verify a claim, we either omit it or label it explicitly as speculation.
## What will and won't we claim? - We will cite specific independent reviewers, manufacturer documentation, and primary research where relevant. - We will distinguish between confirmed facts, manufacturer claims, and our editorial inference. - We will update articles when new information lands and note the refresh cadence in the editorial footer. - We will not claim first-party testing or hands-on experience we haven't done. - We will not fabricate statistics, study results, or testimonials. - We will not hide affiliate relationships - every affiliate link is disclosed (see Affiliate Disclosure). ## How do we use AI-assisted research?We use large-language-model tools as part of our research workflow - primarily for summarising published evidence, drafting outlines, and cross-referencing claims against authoritative sources. Final editorial judgment, source verification, and the published prose are human-supervised. See our AI Use Disclosure for the wider position.
## What is our source policy?We rank sources roughly as follows when assessing claims:
Where sources conflict, we surface the conflict explicitly rather than picking the more flattering claim.
## How often do we update our content?Reviews and comparisons are refreshed quarterly or when meaningful new product information lands. Industry-analysis posts refresh on major events (new product launches, regulatory changes). Each article carries its most recent refresh date in the editorial footer.
If you spot a factual error, please flag it via our contact page - we correct quickly and disclose the change.
## RelatedSee also: Editorial Policy, Affiliate Disclosure, AI Use Disclosure, Privacy Policy.
--- # AI Use Disclosure Source: https://expectedvalue.co.uk/ai-use-disclosure/ # AI Use Disclosure What AI is used for here, and what it is not.AI tools assist with parts of the editorial workflow at Expected Value. Given the subject matter of this site, it would be a strange editorial stance to use AI tools and not say so. This disclosure explains what AI is and is not used for, so readers can decide how to weigh the content.
Where accuracy matters most - probabilities, p-values, sample sizes, effect sizes, attributed quotes, study findings, regulatory facts - claims are checked against primary sources and dated. If you spot a factual error, particularly a maths error, please flag it via the editorial policy page and we will correct it promptly.
--- # Editorial Policy Source: https://expectedvalue.co.uk/editorial-policy/ # Editorial Policy How we research, write, source, and review the content on this site.Expected Value is an editorial resource on probability, decision-making under uncertainty, behavioural economics, and the practical application of expected-value thinking. Our credibility depends on getting the maths right and citing it correctly. This policy documents how the content is made.
Expected Value is independently operated. We do not accept payment in exchange for favourable coverage, do not allow vendors to review or alter editorial content before publication, and do not adjust positions based on commission rates. Where affiliate links appear, they are disclosed on the affiliate disclosure page. Where we cite a book, a paper, or a course as recommended reading, the recommendation is not contingent on whether a link exists.
This is a numbers site. Reviews and guides are research-led and citation-heavy. We cite primary sources wherever possible - peer-reviewed papers, working papers from named academics (Kahneman, Tversky, Thaler, Taleb, Mauboussin, and the broader behavioural-economics canon), original data sets, regulatory filings, and direct quotations from named authors. Where a claim has no supporting source, we say so plainly rather than asserting it as fact, and we hedge with language that reflects the uncertainty ("approximately", "based on a sample of N", "as reported by X").
We do not quote or paraphrase content from competing decision-science sites or pop-science blog posts without attribution, and we do not republish forum or social-media commentary as our own analysis.
Every published article is reviewed by a human editor for factual accuracy before publication. Probabilities, study findings, and worked examples are the most failure-prone parts of writing in this niche - a single transposed digit or a misread confidence interval changes the conclusion. Every numerical claim is checked against its source before publication and dated. If you spot a factual error - particularly a maths error - please flag it. We will correct it, date the correction, and note what was changed.
AI assists with research synthesis, drafting, and editing. A human editor reviews every published piece, with extra attention to numerical and probabilistic claims. See the AI use disclosure page for the detailed policy.
--- # Affiliate Disclosure Source: https://expectedvalue.co.uk/affiliate-disclosure/ # Affiliate Disclosure ## Affiliate DisclosureExpected Value is a free educational resource covering probability, decision-making, and quantitative thinking. To keep this site running and our content free, we participate in affiliate programmes with companies whose products and services we genuinely believe are useful to our readers.
When you click on certain links on our site and make a purchase or sign up for a service, we may earn a commission at no extra cost to you. These commissions help us cover hosting, research and content creation costs.
Our editorial content is never influenced by affiliate partnerships. We recommend products and services based on our own analysis and research. If a product doesn't meet our standards, we won't recommend it - regardless of the commission it might offer.
We apply the same rigorous, evidence-based thinking to our recommendations that we advocate in our content. Every recommendation is evaluated on its merits.
We may earn commissions from:
This disclosure complies with the Federal Trade Commission (FTC) guidelines on endorsements and the UK Advertising Standards Authority (ASA) requirements for affiliate marketing.
If you have questions about our affiliate relationships, please visit our contact page.
Last updated: April 2026
As an Amazon Associate we earn from qualifying purchases.
--- # Start Here Source: https://expectedvalue.co.uk/start-here/ # Start Here A guided reading order through probabilistic thinking - from the single concept that everything else builds on, to the biases quietly shaping your decisions. **Answer** If you are new to probabilistic thinking, start with our Expected Value formula explainer and the Thinking, Fast and Slow summary, then work through the Availability Heuristic, Confirmation Bias, and Base Rate Neglect explainers. From there, branch into the area you most want to apply it to: investing, career decisions, or sports betting. The Tools page has the calculators you will use most often. ## 1. Foundations ## 2. Decision Frameworks Once you can think in expected value, the next question is operational: given an edge, how do I act on it? Three guides, in order - start with Bayesian updating because it teaches you to revise beliefs as evidence arrives, then move to sizing. ## 3. Cognitive Biases Knowing expected value doesn't immunise you against the biases that distort how you perceive probabilities in the first place. This is where most decision failures come from. Read in any order - each is self-contained - but the first three are the most common, most costly, and the best place to start. ## 4. Probability Traps ## 5. Going FurtherA note on order. The four foundation pieces (Section 1) compound - each one makes the next easier. After that, the order genuinely doesn't matter much. Treat this page as an index: skim the cards, follow whichever line of curiosity feels alive, and come back when you want a different branch.
--- # Contact Source: https://expectedvalue.co.uk/contact/ # Get in Touch Questions, corrections, and ideas are always welcome. --- # Corrections & clarifications Source: https://expectedvalue.co.uk/corrections/ # Corrections & clarifications When we get something wrong, we fix it in the open.Our methodology commits us to verifying claims against primary sources and refreshing comparisons when information changes. When we find a substantive factual error after publishing, we correct it and log it here with the date and what changed.
## Correction log26 June 2026 - Kalshi vs Polymarket vs Manifold. Corrected the description of Kalshi's availability. The page previously implied Kalshi was open only to US residents. Kalshi expanded internationally in October 2025 and now accepts members in roughly 143 countries; however, the UK remains on its list of excluded jurisdictions, so UK residents still cannot open accounts. Sources: Kalshi Member Agreement / Help Centre; press coverage of the October 2025 international launch.
26 June 2026 - Vanguard vs InvestEngine. Corrected the date Vanguard introduced its £4-a-month minimum account fee. The page previously stated 2026; the fee took effect on 31 January 2025 (balances up to £32,000, total account fees capped at £375 a year). Source: Vanguard UK published fees and charges.
26 June 2026 - Trading 212 vs eToro. Corrected eToro's ISA fees. The page previously stated a £5 fee on every ISA withdrawal. The eToro Stocks and Shares ISA, powered by Moneyfarm, does not charge a withdrawal fee; its custody fee is capped at £45 a year. FX figures were clarified to roughly 0.50-0.70%. Source: eToro UK ISA fees page.
Spotted something that looks wrong? Please tell us and we'll review it.
--- # Privacy Policy Source: https://expectedvalue.co.uk/privacy/ # Privacy Policy ## Privacy PolicyLast updated: 22 April 2026
Expected Value (“we”, “us”, “our”) operates the website expectedvalue.co.uk. This page informs you of our policies regarding the collection, use, and disclosure of personal data when you use our site.
We collect minimal personal information. When you subscribe to our newsletter, we collect your email address. When you visit our site, we collect anonymous usage data through analytics tools.
We use Google Analytics and Google Tag Manager to understand how visitors use our site. These tools collect anonymous data including pages visited, time on site, browser type, and referring URLs. This data is aggregated and cannot be used to identify individual visitors. Google Analytics uses cookies to distinguish unique users and sessions.
Our site uses cookies for:
You can control cookies through your browser settings. Disabling cookies will not affect the core reading experience on this site.
If you are a resident of the European Economic Area (EEA) or the United Kingdom, you have the following rights:
To exercise any of these rights, please contact us at hello@expectedvalue.co.uk.
Newsletter subscriber email addresses are retained until you unsubscribe. Analytics data is retained for 26 months as per Google Analytics default settings.
We use the following third-party services that may process your data:
Each of these providers has their own privacy policy governing their use of your data.
Our site is not directed at children under 16. We do not knowingly collect personal information from children.
We may update this privacy policy from time to time. Changes will be posted on this page with an updated revision date.
If you have questions about this privacy policy or your personal data, contact us at hello@expectedvalue.co.uk.
Email delivery: we use Amazon Web Services (Amazon SES) to send our newsletter and any updates you have signed up for. Email is processed in the AWS Europe (Ireland) region. See the AWS privacy notice at https://aws.amazon.com/privacy/. ---