Skip to content
Menu
Finabase Learn

Portfolio and risk · Portfolio theory

Fat tails: why extreme market moves are more common than models say

What fat tails, skewness and kurtosis mean, measured on SPY: three months beyond three standard deviations where a bell curve expects one every 60 years.

On this page
  1. The idea
  2. How it works
  3. Using it
  4. Where it fails or misleads
  5. Compared with its neighbours
  6. How Finabase does it
  7. Frequently asked questions
  8. Summary
  9. Check your understanding

A distribution has fat tails when extreme outcomes occur more often than a bell curve with the same average and standard deviation would predict. Market returns have them. Days and months that the normal distribution rates as once-in-a-lifetime events turn up every few years, and the largest of them account for a large share of all the losses and gains an investor ever experiences.

This matters because most of the standard tools of risk measurement, including standard deviation, the Sharpe ratio: return per unit of risk, calculated on real funds and the simple forms of Value at risk (VaR): the loss a portfolio should rarely exceed, either assume a bell curve or work best when returns follow one. They describe ordinary periods well and extraordinary ones badly. This lesson measures the tails of SPY's returns, introduces the two statistics that describe them, and sets out what follows for anyone managing a portfolio.

The idea

The normal distribution, the bell curve, is thin at the edges. Its probabilities drop away extremely fast: an outcome 2 standard deviations from the average occurs about one time in 22, 3 standard deviations one time in 370, 4 standard deviations one time in 16,000 and 5 one time in 1.7 million. A fat-tailed distribution drops away more slowly. The middle looks much the same. The difference is all at the edges, where the consequences are.

SPY monthly returns in the tails: observed and what a bell curve expects, 1993–2026 Observed months Expected under a normal distribution 3 0.5 Below −3 SD 14 8.6 −3 to −2 SD 6 8.6 +2 to +3 SD 0 0.5 Above +3 SD 404 months of total returns. SD = standard deviation (4.25% a month) from the average (0.95%).
Number of SPY monthly total returns in each tail, March 1993 to October 2026 (404 months), and the number a normal distribution with the same average and standard deviation would produce.

The figure counts SPY's months in the four outer regions and compares each count with what a bell curve predicts. The left side has more than expected, about twice as many. The right side has fewer. The tails of monthly stock returns are fat on the side that hurts.

How it works

Two statistics describe the shape of a distribution beyond its average and spread.

  • Skewness measures lopsidedness. Negative skew means the large moves are mostly losses: many modest gains and a few severe falls. Positive skew is the reverse.
  • Excess kurtosis measures the weight of the tails. Because each z is raised to the fourth power, a single observation 5 standard deviations out contributes 625, as much as 625 ordinary observations. A positive value means fatter tails than a bell curve.

A worked example: SPY by month

The data: 404 monthly total returns, March 1993 to October 2026. The average was 0.95% and the standard deviation 4.25%.

Region Observed months A bell curve expects
Below −3 SD (under −11.8%) 3 0.5
−3 to −2 SD 14 8.6
+2 to +3 SD 6 8.6
Above +3 SD (over +13.7%) 0 0.5

Skewness was −0.55 and excess kurtosis 0.97. The worst month, October 2008, lost 16.5%, which is 4.1 standard deviations below average. A bell curve gives that a probability of about 1 in 50,000 months, or once in four thousand years. The best month, April 2020, gained 12.7%, only 2.8 standard deviations above.

SPY monthly returns against a normal distribution, 1993–2026 −5 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 A normal distribution would lie on this line Where the month would fall if returns were normal (standard deviations) Where it actually fell (standard deviations) Each dot is one month, ranked from worst to best. Red: more than half a standard deviation off the line.
Each of 404 SPY monthly returns, in standard deviations from the average, against the position it would have if returns were normally distributed. March 1993 to October 2026.

The second figure makes the same comparison point by point. Sort the months from worst to best and plot each against where a bell curve would put a month of that rank. Normal data would lie along the diagonal. In the middle, SPY's months do. At the bottom left they bend away downwards: the worst months were worse than their rank implies. At the top right they bend slightly below the line: the best months were less extreme than expected.

By day, the tails are fatter

Over shorter intervals the departure from the bell curve is larger. The data: 500 daily returns from closing prices, November 19, 2024 to November 17, 2026.

Symbol Excess kurtosis Days beyond 3 SD (bell curve: 1.4) Days beyond 4 SD (bell curve: 0.03) Worst day Best day
SPY 21.3 6 5 −5.85% (5.5 SD) +10.50% (9.8 SD)
QQQ 11.2 6 3 −6.49% (4.6 SD) +12.00% (8.4 SD)
NVDA 6.3 7 2 −16.97% (6.0 SD) +18.72% (6.6 SD)
TSLA 4.5 6 4 −15.43% (4.4 SD) +22.69% (6.4 SD)
GLD 4.7 3 3 −10.27% (6.6 SD) +6.36% (4.0 SD)
KO 1.8 6 1 −4.44% (3.9 SD) +5.00% (4.2 SD)
TLT 0.9 3 1 −3.02% (4.2 SD) +2.59% (3.7 SD)

Every one of the seven had at least one day beyond 4 standard deviations. A normal distribution expects 0.03 such days in 500, or one in about 60 years. SPY had five.

On April 9, 2025, SPY rose 10.5% in a day, 9.8 standard deviations above its average. Under a bell curve an event of that size has a probability too small to express meaningfully: it should not occur in the lifetime of the universe. It happened, three trading days after the period's worst day (April 4, −5.85%). The correct conclusion is that the bell curve is the wrong description, not that something impossible took place.

The most cited example is October 19, 1987, when the Dow Jones Industrial Average fell 22.6% in a single session, roughly twenty times its typical daily move at the time.

A few days carry the result

SPY, November 19, 2024 to November 17, 2026 (price only) Return
All 500 days +24.5%
Without the 5 worst days +56.9%
Without the 5 best days +0.8%
Without both +26.9%

Ten days out of 500, 2% of the period, account for 12.8% of all the movement in either direction. Removing the five best erases the two-year gain. Removing the five worst more than doubles it. This is often presented as an argument for or against trying to time the market. The table's last row says more: the best and worst days arrive together, in the same turbulent weeks, so avoiding one set means missing the other.

Using it

  1. Do not convert standard deviations into odds for large moves. "A 4-sigma event" is a description of size. Treating it as a 1-in-16,000 chance understates the frequency by a factor of hundreds.
  2. Expect losses beyond the worst on record. The largest loss in a sample is a floor on what can happen, not a ceiling. Each of the worst days in the table was the worst in that sample until it occurred.
  3. Prefer measures that look into the tail. Historical Value at risk (VaR): the loss a portfolio should rarely exceed and expected shortfall use the actual distribution. Maximum drawdown (MDD): the worst fall from a peak, and the way back records what was experienced. Stress tests ask about a specific scenario without assigning it a probability.
  4. Be wary of leverage and of selling insurance. Borrowing magnifies tail losses and can force a sale at the worst moment. Strategies that collect small, steady premiums (selling options, lending to weak borrowers) have smooth records because their losses live in the tail.
  5. Size positions for survival. A portfolio that can absorb a loss two or three times its worst historical month without forcing its owner to sell is robust to this problem. One that cannot is relying on the bell curve.

Where it fails or misleads

  • Kurtosis is itself unstable. One day drives it. SPY's daily excess kurtosis of 21.3 falls sharply if April 2025 is left out. The statistic confirms that tails are fat and gives no reliable measure of how fat.
  • Fat tails do not make everything unpredictable. Volatility clusters: calm follows calm and turbulence follows turbulence. Much of the excess in the tails comes from mixing quiet and violent periods, and current volatility says something about which kind of period it is.
  • Longer horizons look more normal, not safe. Monthly returns are closer to a bell curve than daily ones, and annual closer still. The left tail remains heavier at every horizon examined here.
  • No alternative distribution is agreed. Mandelbrot proposed distributions with infinite variance in 1963; later work favours tails that are heavy but finite. Any fitted tail rests on a handful of observations.
  • "Black swan" is overused. The term refers to events outside the range of what a model allowed for. A 10% market fall is an ordinary event in a fat-tailed world and should be in every plan.

Compared with its neighbours

Statistic What it describes Value for a bell curve SPY monthly, 1993–2026
Mean The centre Any 0.95%
Standard deviation Typical distance from the centre Any 4.25%
Skewness Lopsidedness: which side has the larger moves 0 −0.55
Excess kurtosis Weight of the tails relative to a bell curve 0 0.97
Worst observation The edge of what has happened so far About 3 SD in 400 draws 4.1 SD

How Finabase does it

Finabase does not report skewness, kurtosis or tail statistics, and none of its indicators or studies assume or test for a distribution of returns.

Several things on the site are built around a bell curve and inherit its limits. Bollinger Bands on the chart are drawn 2 standard deviations either side of a 20-period average; prices close outside them more often than the 5% a normal distribution implies. An option chain's implied volatility is a standard deviation under the standard pricing model, and the fact that far out-of-the-money puts trade at higher implied volatilities than at-the-money options is the market's own correction for fat tails. See Volatility and the VIX: historical vs implied volatility explained.

For the raw material, a symbol's History tab lists daily closes and percent changes, which can be sorted to find the largest moves, and the indices page shows the VIX, which rises when the market expects larger moves.

Frequently asked questions

What does "fat tails" mean in plain terms?

That very large gains and losses happen more often than the standard bell-curve model says. The "tails" are the far left and right ends of a chart of how often each size of return occurs.

Are stock returns normally distributed?

Approximately, in the middle. Not at the extremes. Daily returns depart from the normal distribution strongly, monthly returns moderately, and at every horizon the largest losses exceed what the model allows.

What is the difference between skewness and kurtosis?

Skewness says whether the extremes are mostly on one side. Kurtosis says how heavy the extremes are on both sides together. Stock indexes typically show negative skewness and positive excess kurtosis.

What is a black swan?

Nassim Taleb's term for an event that lies outside what past experience suggested was possible, has a large impact, and is explained afterwards as if it had been predictable. Fat tails are the statistical expression of the idea.

If standard deviation understates extremes, is it useless?

No. It is a good measure of ordinary variability and is the right input for comparing how much two investments usually move. Its weakness is confined to statements about rare, large events.

Summary

Market returns have fat tails: extreme outcomes are more frequent than a normal distribution with the same standard deviation predicts, and for stock indexes the excess is on the downside. In 404 months SPY had three losses beyond 3 standard deviations where a bell curve expects one every 60 years, and in 500 days it had five moves beyond 4 standard deviations. A handful of days carried the period's result, and the best and worst arrived together. Standard deviation remains a sound measure of the ordinary. For the extraordinary, use historical losses, expected shortfall and stress tests, and build portfolios that survive a loss larger than any yet seen.

Check your understanding

4 questions on this lesson. Each answer comes with the reason, and nothing is scored.

  1. What does it mean to say that a return distribution has fat tails?

    1. Its average return is higher than normal
    2. Extreme outcomes occur more often than a bell curve with the same standard deviation predicts
    3. Most returns are far from the average
    Show the answer

    B. Extreme outcomes occur more often than a bell curve with the same standard deviation predicts The middle of the distribution looks ordinary. The difference is at the edges, where large gains and losses are more frequent than the normal distribution allows.

  2. In 404 months, SPY had 17 months worse than −2 standard deviations and 6 better than +2. Which statistic describes that imbalance?

    1. Negative skewness
    2. Positive kurtosis
    3. Low volatility
    Show the answer

    A. Negative skewness Skewness measures lopsidedness. More and larger extremes on the loss side give a negative value; SPY’s monthly skewness was −0.55.

  3. A stock index falls by an amount equal to 5 standard deviations of its daily returns. How should the bell-curve probability of that (about 1 in 3.5 million days) be treated?

    1. As accurate: the event was nearly impossible
    2. As evidence that the bell curve is the wrong model for large moves
    3. As proof of a data error
    Show the answer

    B. As evidence that the bell curve is the wrong model for large moves Moves of that size occur every few years in real markets. The normal distribution understates their frequency by orders of magnitude, so its odds for extremes should not be used.

  4. Over two years, removing SPY’s five best days cut its return from +24.5% to +0.8%, and removing its five worst raised it to +56.9%. Removing both left +26.9%. What does the last figure suggest?

    1. The best and worst days tend to occur in the same turbulent periods, so avoiding one group usually means missing the other
    2. Extreme days do not matter
    3. Timing the market is easy
    Show the answer

    A. The best and worst days tend to occur in the same turbulent periods, so avoiding one group usually means missing the other Volatility clusters. The largest gains and losses arrive close together, which is why stepping aside to avoid the falls generally forfeits the rebounds.

Sources and further reading

  1. Benoit Mandelbrot, “The Variation of Certain Speculative Prices,” Journal of Business, vol. 36, no. 4 (1963)
  2. Eugene F. Fama, “The Behavior of Stock-Market Prices,” Journal of Business, vol. 38, no. 1 (1965)
  3. Rama Cont, “Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues,” Quantitative Finance, vol. 1, no. 2 (2001)
  4. Nassim Nicholas Taleb, The Black Swan: The Impact of the Highly Improbable (Random House, 2007)
  5. Report of the Presidential Task Force on Market Mechanisms (the Brady Report), January 1988 — the October 19, 1987 decline
  6. Finabase chart data — SPY monthly bars and dividends, March 1993 to October 2026, and daily closes of SPY, QQQ, NVDA, TSLA, KO, GLD and TLT, November 19, 2024 to November 17, 2026; statistics computed by Finabase