Skip to content
On this page
  1. Key Takeaways
  2. What It Is
  3. The Intuition
  4. How It Works
  5. Worked Example
  6. Common Mistakes
  7. Frequently Asked Questions
  8. Sources
  9. Disclaimer
← All concepts
Quant MethodsAdvanced5 min read

Fat Tails Kurtosis: Why Crashes Happen More Than Models Predict

Fat tails describe a return distribution where extreme outcomes occur far more often than a normal distribution would predict. Kurtosis is the statistical measure that quantifies how fat those tails are.

Key Takeaways

  • Excess kurtosis for S&P 500 daily returns typically runs 15 to 40, meaning extreme days are orders of magnitude more frequent than Gaussian.
  • The October 1987 crash was roughly a minus 22 standard deviation event under a Gaussian fit, with a probability effectively zero by that model.
  • Scaling Gaussian volatility by 1.5x does not replicate fat tails; it just widens the whole distribution including the center.
  • Risk models using Gaussian VaR at 99 percent routinely understate actual equity losses by 30 to 100 percent; Student-t or EVT corrections are the fix.

Key Takeaways

  • Excess kurtosis for S&P 500 daily returns typically runs 15 to 40, meaning extreme days are orders of magnitude more frequent than Gaussian.
  • The October 1987 crash was roughly a minus 22 standard deviation event under a Gaussian fit, with a probability effectively zero by that model.
  • Scaling Gaussian volatility by 1.5x does not replicate fat tails; it just widens the whole distribution including the center.
  • Risk models using Gaussian VaR at 99 percent routinely understate actual equity losses by 30 to 100 percent; Student-t or EVT corrections are the fix.

What It Is

Kurtosis is the fourth standardized moment of a distribution. It measures the relative weight of the tails compared to the center. The normal distribution has a raw kurtosis of 3. Statisticians usually quote excess kurtosis, defined as raw kurtosis minus 3, so the normal reference sits at zero.

A distribution with excess kurtosis greater than zero is leptokurtic, meaning it has a sharper peak and fatter tails than the normal. Daily equity returns are consistently leptokurtic. Long samples of S&P 500 daily returns produce excess kurtosis estimates in the 15 to 40 range, and single-stock returns can go much higher.

The formula uses the fourth central moment scaled by the squared variance:

kurtosis = E[ (r - mu)^4 ] / sigma^4
excess_kurtosis = kurtosis - 3

Where r is the return, mu is the mean, and sigma is the standard deviation.

The Intuition

If daily returns really were normally distributed with a 1 percent standard deviation, a 5 standard deviation move would be expected roughly once every 14,000 years of trading. A 10 standard deviation move would be effectively impossible across the lifetime of the universe.

Reality disagrees. The S&P 500 has produced many moves beyond 5 sigma since 1950. The Black Monday crash on October 19, 1987 was roughly a minus 22 standard deviation event under a Gaussian fit. Under that assumption, its probability is less than one in 10 to the 100th power. The only honest conclusion is that the assumption is wrong.

Mandelbrot made this case in 1963 for cotton prices. Fama reinforced it for equities in 1965. Cont's 2001 survey elevated fat tails to a universal stylized fact of asset returns. Taleb popularized the broader point that rare, extreme outcomes dominate long-run P&L.

How It Works

You estimate kurtosis from a return series by computing the sample fourth moment:

sample_kurtosis = (1/N) * sum( (r_i - mean) / sigma )^4
excess = sample_kurtosis - 3

Because fourth powers amplify outliers, estimates are noisy. A single crash day can move the estimate by several points in a 10-year sample. Practitioners often report rolling estimates alongside full-sample ones.

Three standard responses to fat tails show up across the profession:

  1. Student-t distribution. Parameterized by degrees of freedom nu, it interpolates between Cauchy (nu = 1, infinite variance) and normal (nu = infinity). Daily equity returns typically fit a Student-t with nu between 3 and 6, which matches observed kurtosis well.
  2. Historical simulation for Value-at-Risk. Rather than assume a distribution, sort the empirical returns and read off the quantile. This keeps the real tails rather than imposing Gaussian thin ones.
  3. Extreme Value Theory (EVT). Uses the generalized Pareto distribution to model the tail beyond a threshold. Standard in bank capital models and catastrophe insurance.

Worked Example

Take 10 years of daily S&P 500 returns, roughly 2,520 observations. A typical sample has standard deviation 1.0 percent and excess kurtosis about 12.

Under a normal distribution fitted to the same mean and variance, the probability of a one-day drop of 4 percent or worse is about 0.003 percent, or roughly one day per 13,000 trading days (50 years). Under a Student-t with nu = 4 fitted to the same data, that probability rises to about 0.2 percent, or roughly one day every two years.

The empirical record is closer to the Student-t estimate. In the past 25 years, the S&P 500 has posted multiple daily losses of 4 percent or worse. A VaR model that assumed normality would underestimate capital needs by an order of magnitude.

Common Mistakes

  • Confusing excess kurtosis with total kurtosis. The normal distribution has raw kurtosis 3 and excess kurtosis 0. If a vendor reports kurtosis of 12 without specifying, ask which convention. Python's pandas kurt() returns excess by default. Excel's KURT function also returns excess. Other tools differ.

  • Assuming higher kurtosis always means more tail risk. High kurtosis can come from one fat tail (large negative skew) or two. A symmetric Student-t with nu = 4 and a highly negatively skewed distribution can have similar kurtosis with very different loss profiles. Always look at skewness together with kurtosis.

  • Approximating fat tails with a normal times a fudge factor. Scaling Gaussian volatility up by 1.5x does not replicate fat tails. It makes the whole distribution wider, including the center. A proper fat-tailed model changes the shape of the tails only. Student-t or EVT handles this correctly; Gaussian scaling does not.

  • Applying Sharpe ratio to non-normal returns. Sharpe assumes returns are well-summarized by mean and variance. A strategy with high Sharpe but deeply negative skew and high kurtosis (typical of short-volatility or carry trades) looks great right up until the crash that wipes it out. Use Sortino, Omega, or tail-risk-adjusted measures when skew and kurtosis matter.

  • Computing Value-at-Risk with a Gaussian assumption on fat-tailed data. Parametric Gaussian VaR routinely understates 99 percent and 99.5 percent losses by 30 to 100 percent on equity portfolios. Historical simulation, Student-t VaR, or EVT corrections are the fix.

Frequently Asked Questions

Q: What are fat tails and kurtosis in simple terms? Fat tails mean that extreme market events happen far more often than a bell curve predicts. Kurtosis is the fourth statistical moment that measures exactly how much heavier those tails are compared to normal.

Q: How do fat tails and kurtosis affect investment decisions? Any risk model using Gaussian assumptions will understate the frequency and size of large losses, leading to undercapitalization, underhedging, and VaR numbers that are regularly breached during exactly the crises they are supposed to capture.

Q: What is a real-world example of fat tails and kurtosis? The October 1987 Black Monday crash was a minus 22 standard deviation event under a Gaussian fit, with a probability less than one in 10 to the 100th power, yet it happened. Ten years of daily S&P 500 returns typically show excess kurtosis around 12, far above the Gaussian zero.

Q: How can investors use knowledge of fat tails and kurtosis? Replace Gaussian VaR with historical simulation, Student-t VaR, or EVT-based tail estimates, and check that any strategy with a high Sharpe ratio does not achieve it through structural negative skew that will eventually produce a catastrophic drawdown.

Q: How are fat tails and kurtosis different from skewness? Kurtosis measures whether both tails are heavier than normal, capturing symmetric tail risk. Skewness measures whether one tail is heavier than the other. A distribution can have fat tails and near-zero skewness, fat tails and strong negative skew, or other combinations, so both moments are needed for a complete picture.

Sources

  1. Mandelbrot, B. (1963). "The Variation of Certain Speculative Prices." Journal of Business, 36(4), 394-419. https://www.jstor.org/stable/2350970
  2. Cont, R. (2001). "Empirical properties of asset returns: stylized facts and statistical issues." Quantitative Finance, 1(2), 223-236. https://www.tandfonline.com/doi/abs/10.1080/713665670
  3. Fama, E. (1965). "The Behavior of Stock-Market Prices." Journal of Business, 38(1), 34-105. https://www.jstor.org/stable/2350752
  4. Taleb, N.N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House. https://www.penguinrandomhouse.com/books/176226/the-black-swan-second-edition-by-nassim-nicholas-taleb/

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.

The IWP Substack

You understand the concept. Now see it applied.

The Investing With Purpose Substack turns ideas like this into research and risk-managed trade plans on real stocks, updated every week.

Read on Substack (opens in a new tab)

Related concepts