When do skewness and kurtosis change a decision that mean and standard deviation alone would drive?

## The premise A mean and a standard deviation are a complete description of a distribution only if you have already committed to the normal family. For everything else, they are two summaries out of infinitely many, and the two they capture are the ones least connected to rare events. The cleanest way to make this concrete is to construct two distributions with identical mean and identical standard deviation but different third and fourth standardised moments. One can be symmetric with light tails, bounded, and almost never far from centre. The other can be symmetric with heavy tails, mostly hugging the middle but occasionally producing a value six or eight standard deviations out. Every summary table built on mean and SD reports them as the same variable. Every decision that depends on the far tail treats them completely differently. ## The test I actually apply Ask what the loss function looks like as a function of the realised value. **Roughly linear payoff** — the total cost across many draws, average handle time across a quarter, aggregate spend. Here the mean does most of the work and the standard deviation tells you how precisely you know it. Shape matters mainly through estimation noise: heavy tails mean your estimate of the mean converges more slowly, so you need more data than the usual formulas suggest. But the decision itself is not shape-driven. **Convex or threshold payoff** — a capacity buffer that either holds or fails, a solvency margin, an alert threshold, a queue that collapses past a limit, an SLA breach. Here almost all of the expected loss is concentrated in the tail, and the first two moments are close to silent about it. This is the regime where skewness and kurtosis change the answer. **Asymmetric payoff** — where being wrong high and wrong low cost different amounts. Skewness, specifically, decides which side you should be conservative on. ## Why three-sigma sizing fails on heavy tails "Mean plus three standard deviations" is an implicitly normal rule: under a normal, that band covers about 99.87% of the upper tail, so the buffer is breached about one time in 750. On a distribution with substantial positive excess kurtosis, the same nominal band can be breached one time in 50 or one time in 20. The standard deviation has not lied — it correctly summarises the average squared deviation — but the mapping from standard deviations to probabilities is a property of the shape, not of the spread, and that mapping is exactly what excess kurtosis is telling you has changed. The error compounds because the failure is silent: the buffer looks conservative on paper, and the shortfall only reveals itself on the rare day it matters. ## What I do instead of adding a third and fourth number to the dashboard The temptation is to put skewness and kurtosis into the standard reporting template. I generally resist that, for three reasons. They are noisy statistics that invite over-reading; most consumers of a dashboard cannot convert a kurtosis of 4 into an action; and their value is diagnostic rather than operational. A better division of labour: - **Use the shape statistics as a trigger, not as an output.** Compute them once during characterisation of a metric. If the excess kurtosis is large, that is a signal that the metric needs a different treatment, and the finding gets written down as a property of the metric. - **Replace the parametric summary with a direct tail statement.** Rather than mean plus three SD, state the empirical exceedance level and the size of the exceedances when they happen. That is what the decision actually consumes. - **Simulate rather than derive.** For a convex payoff, resample from the observed distribution and evaluate the loss directly. This sidesteps the need to summarise shape into a scalar at all. - **Record the sample size and the vintage.** A shape characterisation from a period that contained no stress event is a characterisation of a calm regime, not of the process. ## The organisational failure mode The recurring pattern is a team that has standardised on a mean-and-SD template because it is uniform and easy to automate, applied uniformly to metrics whose shapes are wildly different. The template is not wrong for the well-behaved metrics; it is silently wrong for the heavy-tailed ones, and nobody notices because the template gives an answer either way. The lead's job is to know which metrics in the portfolio are in the second category and make sure those have a different treatment, rather than to demand that every metric carry four moments. ## What to say when challenged A reasonable pushback is that adding shape complexity is over-engineering. The honest answer is that it usually is, and the discipline is knowing where it is not: name the small set of decisions in your system whose cost is dominated by rare large values, and confine the extra rigour to those. Everywhere else, two moments and a plot are enough.