Skewness

Skewness is a measure of the asymmetry of a distribution around its mean, formally the third standardized moment: the average of the cubed deviations from the mean, divided by the standard deviation cubed. A perfectly symmetric distribution has skewness of zero, a distribution with a long tail stretching toward high values (positive or right skew) has a positive value, and one with a long tail toward low values (negative or left skew) has a negative value. Cubing preserves the sign of each deviation while sharply magnifying large ones, so the statistic is dominated by whichever tail reaches farther, and dividing by the standard deviation cubed makes it unitless, so skewness can be compared across variables measured on any scale.

The direction of skew is named for the tail, not the bulk of the data, which is the source of a great deal of confusion. A right-skewed distribution has most of its observations clustered at the low end with a thin tail trailing off toward high values: household income, home prices, city populations, file sizes, insurance claims and waiting times are all familiar examples, and the exponential distribution has a skewness of exactly 2. Left skew is rarer in nature, appearing in quantities bounded above, such as scores on an easy exam where many students cluster near the maximum, or the age at death in developed countries. Karl Pearson, who gave the field much of its early vocabulary, proposed simple skewness coefficients built from the gap between the mean and the mode, or three times the gap between mean and median, divided by the standard deviation, and these remain useful as rough guides.

Skewness matters because it determines which summary statistics are honest and which statistical tools are valid. In right-skewed data the mean is dragged toward the tail while the median stays near the bulk, so reporting a mean salary or mean house price overstates what a typical person experiences, which is why official statistics so often quote medians for these quantities. Many procedures that assume approximate symmetry, such as t-based confidence intervals on small samples, ordinary least squares with normal errors, and control charts with symmetric limits, lose accuracy under strong skew. The standard remedy is a transformation that compresses the long tail: a logarithm for strongly right-skewed positive data (which turns a lognormal variable into a normal one), a square root for milder cases, or the Box-Cox family, introduced by George Box and David Cox in 1964, which estimates the best power transformation from the data.

Three pitfalls recur: the mean-versus-median rule of thumb, the myth that zero skewness implies symmetry, and over-reading small samples. The textbook rule that the mean exceeds the median for right-skewed data and falls below it for left-skewed data holds often but not always; Paul von Hippel's 2005 article in the Journal of Statistics Education documented counterexamples, including discrete and multimodal distributions, and corrected the rule's overly confident presentation. A skewness of zero only says that the third moment balances, and asymmetric distributions can be constructed with exactly zero skewness. Finally, sample skewness is a noisy estimate: for normal data its standard error is roughly the square root of 6 divided by n, so with 30 observations a value around 0.45 in magnitude is well within chance variation. Commonly quoted bands, such as absolute skewness under 0.5 being approximately symmetric and above 1 being highly skewed, are conventions, not theorems.

In practice skewness is a diagnostic used before modeling and as a risk descriptor afterward. Data scientists compute it per feature to decide which columns need a log or power transform, and the Jarque-Bera test combines sample skewness with kurtosis to test normality, flagging departures in either direction. In finance, negative skewness in returns means occasional large losses outweigh frequent small gains, a risk profile that variance alone cannot capture, and portfolio managers watch it alongside the standard deviation. In machine learning, skewed targets are often modeled on the log scale so that errors are relative rather than absolute, and skewed class distributions motivate resampling or reweighting. Software packages differ slightly in their bias corrections (the sample-adjusted formula used by spreadsheets versus the plain moment ratio), so the value reported by one tool may not match another's exactly.

Read skewness as the direction and degree in which one tail outruns the other: it tells whether the mean can be trusted as a typical value, whether symmetric tools and a plain average are safe, and when a transformation, a median, or a tail-aware risk measure should replace them, always remembering that a single number summarizing asymmetry is noisy in small samples and cannot capture a distribution's whole shape.