Kurtosis

Kurtosis is built from deviations from the mean that have been divided by the standard deviation and then raised to the fourth power. In words: subtract the mean from each observation, divide by the standard deviation, raise each result to the fourth power, and average those values across the sample. Because the deviations are standardized first, the statistic is unitless and unaffected by the scale of the data, so returns measured in percent or in decimals give the same answer.

The fourth power is what gives the measure its character. A standardized deviation of 3 contributes 81 to the sum, while a deviation of 0.5 contributes only 0.0625. Observations near the mean are therefore almost invisible to the calculation, and a handful of extreme outcomes dominate it. Kurtosis is best read as a statement about how much of the total variance is generated by rare, large deviations rather than by ordinary ones.

A distribution with positive excess kurtosis is leptokurtic: more probability mass sits in the extreme tails than a normal curve with the same mean and variance would allow. Because total probability is fixed at one, that extra tail mass has to be paid for elsewhere, and it is drawn from the shoulders, the intermediate regions roughly one to two standard deviations from the centre. The result is a shape with more mass in the middle and in the tails but thinner flanks, which is why describing kurtosis purely in terms of the height of the peak misleads.

Zero excess kurtosis is mesokurtic, matching the normal benchmark. Negative excess kurtosis is platykurtic: tails are thinner and outcomes cluster in a broader, flatter band. The continuous uniform distribution is the standard platykurtic example. Excess kurtosis has a hard floor of minus two, but no upper bound.

Most risk infrastructure built on the normal distribution understates the probability of large losses. Value at Risk computed from a mean and a standard deviation alone, position sizing rules that assume returns scale smoothly, and stress tests calibrated to two-standard-deviation moves all break down when returns are leptokurtic. Daily equity, bond, credit and currency returns typically show substantial positive excess kurtosis, and the effect is strongest at high frequency, weakening as returns are aggregated to monthly or annual horizons.

Markets price this explicitly. The implied volatility smile in options reflects demand for deep out-of-the-money protection that a lognormal model would consider almost worthless. Hedge fund strategies that sell tail risk, such as short-volatility or carry trades, generate smooth returns with modest measured standard deviation and severe negative skew combined with high kurtosis, which is precisely why Sharpe ratios computed on such series overstate quality.

At Level I you should be able to convert between kurtosis and excess kurtosis, classify a distribution from a reported figure, and state the consequence: with positive excess kurtosis, probabilities of extreme outcomes and the frequency of large losses are higher than a normal approximation implies. Kurtosis is also an input to formal normality tests, where sample skewness and sample excess kurtosis are combined into a single statistic and compared against the values a normal distribution would produce.

Read kurtosis alongside skewness rather than in isolation. Kurtosis treats both tails together and cannot tell you which side the extremes fall on; a symmetric fat-tailed series and a negatively skewed crash-prone series can report similar values. Also treat sample estimates cautiously: because the statistic depends on fourth powers, it is highly unstable in small samples and a single outlier can move it dramatically.