Fat tails, crises, and the Student-t
On October 19, 1987, the US market lost 20% in a single session. Under a normal law calibrated on the volatility of the day, that move sat more than twenty standard deviations out: a probability so small that the entire history of the universe should not have been enough to produce it once.
It happened anyway. And 2008, 2020, 1929, 1973 tell the same story at the annual scale: markets throw extremes far more often than the bell curve allows. These are fat tails. For a retiree, that is no statistical curiosity. A retirement plan fails in the tails, exactly where the normal law is not looking.
This page takes the subject in order. The phenomenon first: where the tails come from, and how they are measured. Then the standard tool for modeling them, the Student-t distribution and its df parameter. Then what tails really change in a withdrawal plan. And last, the honest limits of that modeling. By the end, df will have stopped being an esoteric knob. It is the number that decides how often your simulator is allowed to produce 2008.
Why the normal law is tempting, and where it breaks§
The normal law is not naive. It follows from the central limit theorem: a sum of many small independent effects tends toward a Gaussian. If returns were the sum of thousands of small independent pieces of news, they would be normal. And in the middle of the distribution they nearly are: ordinary years (−10% to +20%) roughly follow the bell.
The trouble is at the extremes, and it is huge. The standard measure is kurtosis. The normal law has a kurtosis of 3. Monthly stock returns typically come in at 6 to 12, daily returns far higher. Extreme months and extreme years are several times too frequent for the bell.
The deep reason comes down to two mechanisms that break the theorem's assumptions. The first: the effects are not independent. Volatility comes in clusters, and one big move announces others. This is the volatility clustering that GARCH models capture. Extremes then feed on themselves, through panic, margin calls and forced selling, everyone selling because everyone else is selling. The second: the market changes regime. The return distribution of a calm market and the one of a crisis are not the same law. And mixing two Gaussians with different volatilities mechanically produces a fat-tailed distribution.
So the tails are not a quirk. They are the statistical signature of a fact: markets are a social system with feedback loops, not an urn (Market regimes (growth × inflation, sticky bears) and why they matter).
One reassuring fact: tails get thinner as you aggregate over longer horizons. Daily returns are wildly non-Gaussian. Monthly returns much less so, annual returns less again, because stacking twelve months smooths part of the extremes. Ten-year returns are almost respectable. A retirement plan lives at the monthly and annual scale, the rhythm of withdrawals. At those scales the tails are very real. That is where the model does its work.
The Student-t and the df parameter, in plain terms§
One family of distributions was built for exactly this problem: the Student-t. It is a bell with three parameters, the center μ, the scale (tied to σ), and the degrees of freedom df, which set how fat the tails are. At infinite df, the Student-t is the normal law. As df comes down, the center barely moves but the tails get heavier. Below df 4 they turn genuinely wild. This is the foundation of the central model in serious simulators: monthly Student-t draws, compounded into years (Historical windows, bootstrap, parametric: the three families of models).
What df changes in practice, at identical annual volatility (σ = 12%, μ = 4% real; the orders of magnitude are those of a "catastrophic" year at −30% real, close to a 3 standard deviation event):
| df | The world it describes | How often a −30% real year | Over a 45-year retirement |
|---|---|---|---|
| 30+ (about normal) | The world of math textbooks | ~1 year in 400 | Probably never |
| 10 | Moderate tails | ~1 in 150 | Maybe once |
| 5 (the sensible default) | The world of real monthly data | ~1 in 40 | Once or twice |
| 3 | A catastrophe-prone world | ~1 in 20 | Two or three times |
Compare the df 5 row with the df 30 row: the same σ, and a factor of ten on how often catastrophes happen. That is why two simulators both showing "12% volatility" can hand you unrelated verdicts. It all sits in the law of the tails, which most commercial tools do not even document; they are Gaussian without saying so (Simulator traps: independence, US bias, survivorship).
Where does the df value come from? You do not guess it, you estimate it from the data. The method is two steps. Measure the kurtosis of the portfolio's monthly returns. Then invert the Student-t's theoretical relation, kurtosis ≈ 3 + 6/(df − 4), to get the matching df. Months with a kurtosis of 7 to 9 give df around 5 to 6. Plain equity funds land near df 4 to 6, while a heavily diversified portfolio with defensive sleeves can climb to 8 to 12. A tool that can read the history of your holdings runs that computation for you and leaves the value editable for sensitivity tests. So a df of 5 is not a cautious opinion. It is the number that comes out of the real monthly data of most global equity portfolios.
What tails change in a withdrawal plan§
From the model to the plan. Fat tails hit a withdrawal plan in a precise pattern, and understanding it heads off two opposite mistakes.
First effect, direct: failure probability rises, the median barely moves. Going from df 30 to df 5 at constant σ leaves the median scenario almost untouched, because ordinary years have not changed. But failure probability swells, typically by 30 to 80% in relative terms depending on the plan. A Gaussian 4% failure becomes 6 to 7% at df 5. Tails do not change your likely life. They change how often the unlikely lives happen, and those are exactly the ones you plan against (Failure probability: reading it, choosing it, and not letting it run you).
Second effect, subtler: the interaction with sequence. A −30% year does not cost the same depending on when it lands. In year 2 it cuts the plan down for good (Sequence of returns risk: the retiree's real enemy). In year 30 it is cosmetic. Fat tails raise the odds that the fragile window contains a catastrophe. The danger is the product of the two risks. The practical corollary: the defenses of the fragile window (glide path, buffer, early income) are also the best anti-tail defenses, at no extra cost (Glide paths: the bond tent, rising equity, and the fragile window, The cash buffer: size, cost, and what it is really for).
Third effect: diversification delivers less than advertised. The Gaussian case for diversification (average correlations reduce σ) misses one fact about crises: in the tail, correlations rise. In 2008 everything fell together except long government bonds and the yen. Diversification works, but it works less well in exactly the scenarios you bought it for. Unless you own assets whose decorrelation survives a crisis: government duration, gold, systematic managed futures (Defensive assets: the map and the roles, Managed futures and trend following: the diversification that works in a crisis, All-weather portfolios: Browne, All Weather, Golden Butterfly, Dragon). A portfolio's kurtosis comes down through the choice of building blocks, more than through their number.
Two mistakes are left to avoid. The optimistic one says: "the median does not move, so tails are a detail." No: retirement planning is tail management. The safe withdrawal rate is defined by the worst cases, not by the median (Bengen, the Trinity study, and the birth of the safe withdrawal rate). The catastrophist one says: "then set df 3 everywhere and stop talking about it." No again. Stacking df 3 on top of an already prudent mean and a CAPE anchor is the triple-counted prudence this book keeps warning about (Forward-looking expected returns (Morningstar, Vanguard, the investment banks)). And a permanent df 3 world is not the real world either. Fitting to the data, plus one sensitivity test, beats both postures.
A little history of the idea, to settle the matter§
The history is worth knowing, because it inoculates you against models that are too clean. As early as 1900, Bachelier founded mathematical finance on Gaussian Brownian motion. In 1963, Benoit Mandelbrot studied cotton prices. He showed that the tails were so fat that the variance itself looked barely defined, and proposed wild "stable" laws. Eugene Fama (1965) confirmed it on stocks. The profession, though, needs models it can compute with. So it settled on a pragmatic middle: keep the variance, but fatten the tails (Student-t, mixtures, GARCH). Praetz (1972), then Blattberg and Gonedes (1974), established that the Student-t fits returns remarkably well. The crises of 1987, 1998, 2008 and 2020 settled the cultural debate. LTCM, in 1998, went down precisely for trusting thin tails. "Fat tails" moved from Mandelbrotian heresy to basic fact, popularized by Taleb (The Black Swan). The modeling that won out in decumulation simulators (a fitted Student-t, plus regimes for the sequence) is the direct heir of that compromise: more honest than the Gaussian, more sober than the stable laws, and fitted to your data.
The essentials§
- Markets throw extremes far more often than the normal law allows: monthly kurtosis of 6 to 12 against 3. That is the signature of volatility clustering and regime changes, not an accident.
- The Student-t adds the missing parameter: df, the thickness of the tails. At identical σ, df 5 makes a −30% real year about 10 times more likely than df 30. You estimate df by inverting kurtosis ≈ 3 + 6/(df − 4) on your funds' monthly returns, which typically gives 4 to 6.
- Effect on the plan: the median does not move, failure probability rises, often by 30 to 80% in relative terms. Tails mostly worsen the fragile window, and the anti-sequence defenses are the anti-tail defenses too.
- Limits of the model: it is symmetric (real risk has a negative skew, covered by the sequence stresses and the lost decade) and memoryless (covered by the regimes). Diversification protects less in the tail, except for the blocks whose decorrelation survives a crisis.
- Practical habit: keep the df fitted to your data, run the 3/30 test once to learn your exposure, and if the decision flips, fix it with the structure of the plan, not with the parameter.
Going further§
- Mandelbrot, "The Variation of Certain Speculative Prices" (1963) and The (Mis)Behavior of Markets (2004): the founding act and its popular version.
- Fama, "The Behavior of Stock-Market Prices" (1965); Blattberg & Gonedes (1974) on the Student-t: the academic confirmation.
- Taleb, The Black Swan (2007): the general culture of tails, to be read with a critical eye.
- How df is estimated, and exactly how it is used: Under the hood: how the simulator computes this book, with Using the FIRE simulator for the manual.
- Next in this book: Market regimes (growth × inflation, sticky bears) and why they matter (where the clusters come from) and Making a Monte Carlo relevant: blending, regimes, stress (how tails and regimes combine in the central model).