Why do real markets have fatter tails than the bell curve?
Almost every risk model ever built starts, somewhere in its foundations, with the normal distribution — the bell curve. It's mathematically beautiful, it needs only a mean and a standard deviation, and it makes everything else tractable.
It is also, in the tails, badly wrong about markets — and wrong in the dangerous direction.
What the bell curve promises
Under a normal distribution, extreme moves are astronomically rare. Measured in standard deviations (sigma) from the mean:
| Move | Expected frequency |
|---|---|
| 3 sigma | about once every 3 years |
| 4 sigma | about once every 125 years |
| 5 sigma | about once every 13,000 years |
| 6 sigma | roughly once in 4 million years |
These are one-sided frequencies at 252 trading days a year. Read the last two rows again: the model says a 5-sigma down day should not occur in recorded human history.
What markets actually do
On 19 October 1987, the Dow Jones Industrial Average fell 22.6% in a single session. Against a normal distribution fitted to the preceding period, that is on the order of a 20-standard-deviation event — a probability so small it has no meaningful physical analogue. It happened anyway.
October 2008 produced several daily moves near or beyond ±9% on the S&P 500 — within weeks of each other. March 2020 produced a cluster of double-digit swings. Under a normal model, moves of that size clustering that closely are effectively impossible. Under the actual historical record, they are a recurring feature.
In August 2007, Goldman Sachs' then-CFO David Viniar described the firm's quantitative funds' losses to the Financial Times: "We were seeing things that were 25-standard deviation moves, several days in a row." The sentence is often quoted as a joke about model failure. It is better read as a diagnosis: if your model says the last three days were impossible, the days were fine and the model was wrong.
Why the tails are fat
Three mechanisms, none exotic:
Volatility clusters. Markets are not equally turbulent over time. Calm periods produce small moves; stressed periods produce large ones. A single fixed standard deviation averages these regimes together and describes neither.
Behaviour is correlated. The bell curve arises when many small, independent influences add up. Market participants are not independent. Fear synchronises them, and synchronised selling produces moves no independent-agent model can generate.
Leverage forces action. Falling prices trigger margin calls, which force selling, which pushes prices lower, which triggers more margin calls. This feedback loop has no analogue in a coin-flipping model — and it is precisely what turns a bad day into a historic one.
The measurement
Statisticians quantify tail-fatness with kurtosis. A normal distribution has a kurtosis of exactly 3 (an excess kurtosis of 0). Daily equity index returns routinely show excess kurtosis well above zero — often several times the kurtosis of 3 that a bell curve carries, depending on the window and the asset. Individual stocks are typically fatter still.
There is a simpler diagnostic you can run yourself: count the 3-sigma days. The bell curve says a 500-day window should contain roughly one down day beyond −3 sigma. Most real windows contain several. Every one you find in excess is a day your parametric model considered nearly impossible.
What this does to your numbers
Fat tails mean parametric VaR understates the deep tail — the more extreme the confidence level, the worse the understatement. It means Expected Shortfall computed under a normal assumption is too small. And it means that when someone reports a "6-sigma event," the correct first reaction is not awe at the improbability. It's a question about the model.
In the data
The tail lives in the daily record. Six sessions of the S&P 500 fund in March 2020:
Five moves in a row of between 5% and 11%, each far beyond three standard deviations of a calm year. Now the same weeks as monthly bars:
One bar, down 13%. A month adds up about twenty-one daily moves, so the individual crash days vanish into it, and a decade that showed several extreme days can show almost none once resampled. Kurtosis measured on monthly data describes monthly data, not the days the portfolio actually lived through.
Try it now
- The full history is below, as daily candles. Pick a calm year and measure a dozen ordinary sessions to get a feel for its typical daily move; three times that is roughly your −3σ line. Now count the sessions in that year that fell further. A normal distribution predicts about one in three years of trading.
- Do the same count in a crisis window — February to April 2020 will do — and then in a genuinely calm one such as 2017. The difference between the two counts is volatility clustering, made visible: the tail days are not sprinkled evenly through history, they arrive together.
- Now switch the chart to Monthly and look for those same days. They are gone, averaged into bars that look unremarkable. Write one sentence describing what your count implies about a normal-distribution VaR on this data, and one about what a monthly series would have told you instead. Both are observations about model fit — neither is a claim about what happens next.