Risk in a Heavy-Tailed World

Here is a gamble whose expected return is +5% every single round, and in which more than eight players out of ten end up losing most of what they had. No bad luck is required. The word "average" is quietly pointing at two different things, and their answers run in opposite directions.

01Tails Are Not Outliers — They Are the Ledger

Measure 120 people, add up their heights, and ask what share the tallest person contributes. About 1.1%. Take 1,200 people and the share drops towards a fraction of a percent. On the question "how much height is there in this room", the tallest person has almost no say.

Now swap heights for wealth, for the loss from a single accident, for the area burned by a single wildfire. Out of 120 draws, the largest can be 13% of the total, and the top three a quarter of it. Double the sample and that share does not politely shrink — it waits for a bigger draw to show up and then hauls the total over to itself again.

That is what a heavy tail (also called a fat tail) actually means. Not "a big number turns up now and then", but: the total is decided by a handful of observations, and that does not ease off as the sample grows.

Which gives you a one-line test: divide the largest item by the sum, then double the sample size and compute it again. If the ratio falls, you are in a thin-tailed world. If it refuses to move, you are in a heavy-tailed one. No distribution needs to be fitted, and nothing needs to look like a straight line on a log-log plot — which, incidentally, is the single most common way this gets misdiagnosed → ref · identifying power laws.

The difference then changes what you do with data all the way down. In a thin-tailed world an outlier is usually noise: a mistyped entry, a sensor glitch. Trim it and what remains is cleaner. In a heavy-tailed world the thing you just trimmed is the thing you were trying to explain — the "typical behaviour" computed after removing the largest few describes a system that has never existed.

Earthquakes are the cleanest calibration available. One step up in magnitude is roughly 31.6× the energy released (10^1.5). So "how big is an average earthquake" is answered almost entirely by the largest few in the record: change the observation window and you get a different answer. That is not imprecision. The question has no stable answer here.

02Why Risk Models Systematically Underestimate

The claim is not that models are sometimes wrong. It is that four independent biases all push the same way: they make danger look smaller.

First, "never seen" is not "very rare". Medical statistics has a crude and excellent rule called the rule of three: if an adverse event has not been observed once in n patients, the 95% upper bound on its rate is about 3/n. A phase-three trial with 3,000 participants and not a single serious adverse event guarantees only that the rate is probably under one in a thousand — and one in a thousand, across a million users, is a thousand cases. Zero observations carry far less information than intuition assumes.

Second, means and volatilities simply do not converge under heavy tails. The chart below does the same thing twice: draw one more sample, recompute the average, and see how far it still sits from the truth.

The thin-tailed line settles after fifty-odd draws. The heavy-tailed line is still being dragged 40% off course by a single new draw at two thousand — and the next drag has no schedule. Which means the standard deviation computed from such data is an unstable number: change the stretch of history and it changes value. Every risk measure built on top of it inherits that.

Third, the most widely used measure is blind to the tail by construction. VaR (value at risk) says: "on 99% of days, the loss will not exceed X". Notice what it says about the remaining 1% — nothing at all. Whether the loss out there is 1.1×X or 50×X, VaR reports the same number. And in a heavy-tailed world, the difference inside that 1% is exactly the difference between a painful day and being out of the game.

Fourth, the largest value you have seen is a biased lower bound. "Worst drawdown on record", "peak traffic since launch", "hundred-year flood" — all of them mean "so far". In a thin-tailed world records get harder to break. In a heavy-tailed one they keep breaking, on no schedule. Designing to the historical maximum assumes you have already seen the worst.

And one mechanism amplifies all four: correlations rise in the tail. Things that normally move independently — different assets, different suppliers, different data centres — move together in the extreme. Diversification fails on precisely the day you needed it.

Stack those and you get the famous line from August 2007, when Goldman Sachs CFO David Viniar explained a week of enormous losses in the firm's funds to the Financial Times: "We were seeing things that were 25 standard deviation moves, several days in a row." Under a normal distribution, a 25-sigma event should not occur once in the age of the universe. So what the sentence really reports is not that markets went mad: when a model announces an absurdly small probability, the thing being falsified is the model, not the world.

03Expected Value Counts Ten Thousand People — You Have One Life

Play a gamble. You start with 100. Flip a fair coin: heads, your money grows by 50%; tails, it shrinks by 40%. The expected return per round is (+50% − 40%) ÷ 2 = +5%. It looks like free money.

Run it 4,000 times, sixty rounds each:

The average across 4,000 paths reaches 38× the stake. The median path holds 4% of it. 83% of players finish below where they started and 55% finish below a tenth. And the reason that handsome average keeps rising is that the top 1% of paths — forty of them — hold 93% of all the wealth in the ensemble.

Both numbers are right; they average over different things. "+5% per round" averages sideways: ten thousand people standing at the same instant, their money summed and divided by headcount — the ensemble average. What you actually live through averages downwards: one person, round after round — the time average.

Going downwards, money multiplies, it does not add. So the relevant quantity is the geometric mean: √(1.5 × 0.6) = 0.9487, which is −5.1% per round. In the same gamble, "expected +5%" and "−5% every round" are both true statements.

When are the two averages equal? Physics has a name for that condition: ergodicity. Boltzmann's assumption was that the long-run behaviour of one gas molecule matches the behaviour of many molecules at one instant. For a gas that broadly holds, because molecules do not drop out. For money it fails, because you can hit zero — and zero is an absorbing barrier: nothing follows it, so every later round and every bit of its positive expectation vanish with you. Expected value is counting rounds you are no longer present for.

The repair is not abstinence; it is sizing. Stake only a fraction f of what you have and the same gamble changes character immediately:

All in (f = 100%): −5.1% per round ｜ half (f = 50%): exactly break-even ｜ f = 25%: +0.62%, the optimum ｜ f = 10%: +0.40%

Position size decides whether this gamble is slow suicide or a compounding machine, and the expected return never changed at any point. That optimum has a name: the Kelly criterion (1956), which maximises the logarithmic growth rate — that is, the time average. Almost nobody bets full Kelly in practice: the drawdowns are brutal, and it assumes your probability estimates are correct, an assumption usually more dangerous than the gamble.

If you want to know what your own path looks like, the honest method is to run the rules ten thousand times and look at the whole distribution rather than at one number produced by a formula → ref · Monte Carlo & ensemble simulation.

04Where This Breaks Down

Heavy-tail language is very easy to wear as a universal coat. Five limits follow, two of which are enough to overturn most of what you have just read.

First, not everything is heavy-tailed. Quantities built by adding many small bounded contributions — heights, body temperature, measurement error, a day's commute — are thin-tailed, the central limit theorem applies as advertised, and the mean and standard deviation are good rulers. Heavy tails come from specific mechanisms: multiplicative growth, preferential attachment (rich get richer), contagion coupling, winner-take-all. The test is whether the mechanism is present, not whether the plot looks straight. And most things that look straight fail a proper test → ref · identifying power laws. Applying heavy-tail thinking where it does not belong has a real price: you hoard defences against disasters that will never arrive, and spend the budget on fear.

Second, the ergodicity argument has a hard precondition: you cannot pool and you cannot reset. For an entity that can pool a large number of near-independent risks — an insurer, an index fund, an institution with continuing outside capital — the ensemble average really is the right object, because it genuinely is all ten thousand paths at once. So "never use expected value" is wrong. The right question is not whether expected value is legitimate but "in this particular case, am I the path or the ensemble?"

Third, one formula, two rival justifications — do not treat them as one thing. In 1738, working on the St Petersburg paradox at the Petersburg Academy, Daniel Bernoulli proposed that people weigh not money but its utility, which grows roughly with the logarithm of wealth: a psychological explanation. Ole Peters has recently defended the same logarithm from the time average: not a matter of preference at all, but a property of repeated multiplicative processes. Economists have not accepted the substitution — Doctor, Wakker and Wang replied directly in Nature Physics in 2020, arguing that it repackages an old problem. You need not take a side, but you should know where the two justifications diverge: exactly at the question of whether risk can be pooled — the psychological account lets preferences change, the dynamic account says preference has nothing to do with it.

Fourth, heavy tails are not fatalism. What you give up is point prediction, not intervention. The distribution itself is structural and can be edited: per-bet caps, stop-losses, modular firebreaks, insurance, splitting one large bet into many small ones — each of these truncates the tail. Saying "it is unpredictable anyway" and doing nothing moves a conclusion about prediction onto a question about structure, where it does not belong.

Fifth, and hardest to hold to: do not use "black swan" as a retroactive excuse. An honest heavy-tail claim has to be sayable in advance: events of what size, arriving about how often, absorbed by what structure. If you cannot say those three things beforehand, saying "heavy tail" afterwards is only a more dignified name for the failure.