Kurtosis

Kurtosis is the fourth standardized moment of a distribution: the average of the fourth powers of the deviations from the mean, divided by the standard deviation to the fourth power. It is a unitless summary of how much of a distribution's variability comes from rare, extreme values far from the center rather than from moderate deviations near it. The normal distribution has a kurtosis of exactly 3, and because that value is the natural benchmark, most software reports excess kurtosis, defined as kurtosis minus 3, so that the normal sits at zero. Karl Pearson introduced the term in 1905, along with the labels leptokurtic, mesokurtic and platykurtic for distributions with more, equal, or less tail weight than the normal.

Kurtosis measures tail extremity, not peakedness, correcting a textbook description that persisted for about a century. The older story said a high kurtosis means a sharp, narrow peak and a low kurtosis a flat top, and that picture is not wrong for the standard examples, but it misidentifies the cause. Peter Westfall's 2014 article in The American Statistician, titled "Kurtosis as Peakedness, 1905-2014. R.I.P.," showed that kurtosis is determined essentially by what happens in the tails: raising the fourth power makes any observation several standard deviations out contribute enormously, whereas the center contributes almost nothing. Two distributions with identical peaks can have wildly different kurtosis if one has occasional far-out values, and a distribution can look flat-topped yet be heavy-tailed. The practical reading is therefore "propensity to produce outliers."

Some benchmark values make the scale concrete. The normal has excess kurtosis 0, the uniform distribution has -1.2 (it is bounded, with no tails at all, the lowest among the common continuous shapes), the Laplace (double exponential) has +3, and the Student t distribution has excess kurtosis 6 divided by (degrees of freedom minus 4) when the degrees of freedom exceed 4 and an undefined or infinite value otherwise. That last fact is the deepest warning: for sufficiently heavy-tailed distributions the fourth moment does not exist, so there is no true kurtosis to estimate, and a sample kurtosis computed from such data never settles down as the sample grows. Benoit Mandelbrot's 1963 work on cotton prices argued that speculative price changes are far more heavy-tailed than normal models allow, and the evidence that financial returns have strongly positive excess kurtosis became one of the central empirical facts of modern finance.

Kurtosis has direct practical consequences wherever extremes drive outcomes. In finance and insurance, positive excess kurtosis means that moves of five or six standard deviations, essentially impossible under a normal model, occur at noticeable frequency, so value-at-risk and stress-testing models based on normal assumptions understate tail risk. In quality control and anomaly detection, a feature whose kurtosis jumps signals a change in the rate of extreme readings even if its mean and variance are stable. In statistics, tests such as Jarque-Bera combine skewness and kurtosis to assess normality, and many classical procedures are far more sensitive to tail weight than to skew; the validity of tests comparing variances in particular depends heavily on kurtosis. In deep learning, kurtosis of activations or weights is sometimes tracked to diagnose outliers that complicate quantization.

The main pitfalls are over-trusting a single sample value and mixing conventions. Because of the fourth power, the sample kurtosis is dominated by a single observation: one wild data point can move it dramatically, and its sampling variability is large, with a standard error near the square root of 24 divided by n for normal data. Conventions also differ: some tools report raw kurtosis with the normal at 3, others excess kurtosis with the normal at 0, and spreadsheet functions apply a small-sample bias correction, so a reported number must be checked against its definition before interpretation. Finally, kurtosis says nothing about the direction of the tails, which is skewness's job, and a distribution can be symmetric and heavy-tailed at once.

Read kurtosis as a gauge of how heavy the tails are, meaning how readily a distribution produces extreme values: it is the number that warns that rare events are less rare than a normal curve would suggest, it should be interpreted against the benchmark of 3 (or 0 for excess), and it should be trusted only as far as the sample is large and free of a few outliers that, because of the fourth power, can dominate the figure.