Quant Analytics: from scratch to pro · Quant foundations: returns, risk and clean data · Intermediate
Return distributions and fat tails
The normal distribution is the default model for returns, and it understates how often large moves happen. This lesson covers skewness and kurtosis, shows how rare a five-sigma day would be under the normal, and explains why real markets produce them far more often.
In this lesson
- Compute and interpret skewness and kurtosis
- Work out how rare a k-sigma day is under the normal distribution
- Explain what fat tails mean for risk estimates built on the normal
- Choose measures that do not rely on the normal shape
Read first: Volatility: from daily moves to an annual figure
Most risk numbers start from a picture of how returns are spread. The usual picture is the bell curve, the normal distribution. It is easy to work with and it describes everyday moves well. It describes extreme moves badly, and extreme moves are where the damage happens.
Describing a distribution
Four numbers summarise a set of returns.
- Mean: the centre.
- Standard deviation: the typical spread (see volatility).
- Skewness: whether one tail is longer. Negative skew means large falls are more common than large rises of the same size.
- Kurtosis: how heavy the tails are, as
kurtosis = mean((r - mean)^4) / variance^2. The normal distribution has a kurtosis of 3. Excess kurtosis is the kurtosis minus 3, so the normal scores 0.
A distribution with positive excess kurtosis has fat tails: more of its probability sits far from the centre, and usually more in the middle too, with less in the shoulders.
How rare is a k-sigma day
A sigma is one standard deviation. Under the normal distribution:
| Move | Chance | Roughly once in |
|---|---|---|
| 3 sigma, either direction | 0.27% | 370 days |
| 4 sigma or more down | 0.0032% | 31,574 days (125.3 years) |
| 5 sigma or more down | 0.00003% | 3,488,556 days (13,843 years) |
Years use 252 trading days. Real daily returns show many more moves of this size than the normal allows. That is the central evidence for fat tails. Mandelbrot argued the point from cotton prices in 1963, and Cont's survey lists heavy tails among the stylised facts found across markets.
Worked example: Black Monday
On 19 October 1987 the S&P 500 fell by about 20% in one day, according to a Federal Reserve staff history. See the Black Monday case file. Suppose, for illustration, that the index's daily volatility had been 1%. A 20% fall is then 20 standard deviations. The normal gives that a chance of roughly 10^-89, which is not a number that means anything. The conclusion is not that the day was impossible. It is that a normal model with a constant volatility cannot describe it.
Worked example: kurtosis from ten returns
Take ten returns in per cent: four of +1, four of -1, one of +5 and one of -5. The mean is 0.
- Second moment (variance): (8 x 1 + 2 x 25) / 10 = 5.8.
- Fourth moment: (8 x 1 + 2 x 625) / 10 = 125.8.
- Kurtosis: 125.8 / 5.8^2 = 3.74, so the excess kurtosis is 0.74.
Without the two large returns every value is plus or minus 1 and the kurtosis is 1.00. Two outliers in ten lift it above the normal's 3. Kurtosis is sensitive to single points, so it is noisy in small samples.
Why tails are fat
Two common reasons. Volatility is not constant: calm and turbulent periods alternate, so returns drawn from a mix of volatilities have fat tails even if each regime were normal. And information and forced trading arrive in jumps. A Student t distribution is a standard stand-in; with 6 degrees of freedom its excess kurtosis is 6 / (6 - 4), which is 3.
What to do about it
- Do not read "within two standard deviations" as "95% of days" without checking the data.
- Prefer measures of loss taken from the actual return history, or from a heavy-tailed model, to figures that assume normality.
- Look at the worst days directly. See tail risk and rare events.
- Be careful with small samples: kurtosis and the extreme quantiles are the least stable statistics.
Common mistakes
- Treating a 5-sigma day as a once-in-millennia event. It usually shows the model is wrong.
- Quoting kurtosis from one calm year.
- Assuming that fat tails and skew affect only the downside. They can affect both.
- Assuming volatility figures bound the loss.
In practice
Check the shape of your returns before relying on a tail estimate: a histogram on a log scale and the five worst days tell more than one number. Analytics on FinUties applies measures like these to the markets it covers, so they are at hand when you read a chart.
How rare is a k-sigma day under the normal?
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| two-sided chance of a 3-sigma day under the normal | 2 * phi(-3) | 0.27% |
| one day in | 1 / p3 | 370 |
| one-sided chance of a day 4 sigma or more below the mean | phi(-4) | 0.0032% |
| days between such falls | 1 / p4 | 31,574 |
| years between such falls at 252 days | n4 / 252 | 125.3 |
| one-sided chance of a day 5 sigma or more below the mean | phi(-5) | 0.00003% |
| days between such falls | 1 / p5 | 3,488,556 |
| years between such falls at 252 days | n5 / 252 | 13,843 |
| a 20% fall when daily volatility is 1% | 0.20 / 0.01 | 20 standard deviations |
| second moment of the ten returns | (8 * 1 + 2 * 25) / 10 | 5.8 |
| fourth moment of the ten returns | (8 * 1 + 2 * 625) / 10 | 125.8 |
| kurtosis of the ten returns | m4 / m2**2 | 3.74 |
| excess kurtosis of the ten returns | kurt - 3 | 0.74 |
| excess kurtosis of a t distribution with 6 degrees of freedom | 6 / (6 - 4) | 3 |
Check yourself
Under the normal, how often does a daily return fall 3 standard deviations or more from the mean, in either direction?
2 x phi(-3) = 0.27%, about one day in 370. With 252 trading days a year, about two-thirds of a day per year.
A return series has an excess kurtosis of 0. Does that prove it is normal?
No. Kurtosis is one summary of the tails. A skewed or lumpy distribution can still have excess kurtosis of 0. Plot the data and look at the extremes too.
A daily volatility of 1% and a one-day fall of 20%: how many standard deviations?
20% / 1% = 20 standard deviations. Under the normal the chance is about 10^-89, far too small to matter. The model, not the day, is wrong.
Ten returns are four of +1, four of -1, one of +5 and one of -5 (per cent). What is the kurtosis, and what would it be without the two large ones?
With the large ones: variance 5.8, fourth moment 125.8, kurtosis 125.8 / 5.8^2 = 3.74. Without them the values are all plus or minus 1 and the kurtosis is 1.00. Two outliers out of ten move it far above the normal's 3.
Key terms
Checked facts
Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).
- Primary checked
On 19 October 1987 the S&P 500 fell by about 20% in a single day.
Sources
- The Variation of Certain Speculative Prices. The Journal of Business 36(4), 394-419 (Benoit Mandelbrot), 1963. Accessed 9 Nov 2026
- Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance 1(2), 223-236 (Rama Cont), 2001. Accessed 9 Nov 2026
- A Brief History of the 1987 Stock Market Crash with a Discussion of the Federal Reserve Response. Board of Governors of the Federal Reserve System, Finance and Economics Discussion Series 2007-13 (Mark Carlson), 2007. Accessed 9 Nov 2026