Extreme Value Theory: GPD & Peaks-Over-Threshold

Almost everything you have learned about a distribution so far has been about its middle.

The mean lives in the middle. The standard deviation measures the typical wiggle around the middle. Even the histogram you stare at is, visually, dominated by the middle, because that is where nearly all of the data sit. And for most questions that is fine. But risk management is not a question about the middle. Risk management is the question of how bad the worst day can be — and the worst day, almost by definition, is a place where you have almost no data. The entire discipline of this chapter exists because the part of the distribution you most need to know is the part you have seen least.

Here is the trap you must feel in your bones before the math arrives. Suppose you have ten years of daily returns, roughly two thousand five hundred observations, and you want the one-in-a-thousand-day loss — the loss you would expect to be exceeded on only about one trading day in four years. You have, at most, two or three observations out there. If you simply sort your returns and read off the empirical quantile, you are estimating a once-in-four-years event from two data points. Worse, you can never read off a quantile beyond the worst loss you happened to observe. The empirical distribution flatly assigns zero probability to anything more extreme than your historical maximum, which is a confident statement that the unprecedented cannot happen — and that statement has bankrupted firms.

So we need a way to extrapolate past the data we have. Not by guessing, but by exploiting a remarkable piece of mathematics that says the shape of the far tail is not arbitrary.

The intuition first. Imagine you stop looking at the whole distribution and instead fix a high threshold — say, the worst five percent of losses — and study only what happens beyond it. You throw away the calm middle entirely and keep only the exceedances: for each loss that crosses the threshold, you record how far past it went. This is the Peaks-Over-Threshold approach. The astonishing result, the Pickands–Balkema–de Haan theorem, is that as you push the threshold high enough, the distribution of those exceedances converges to a single, universal family — the Generalized Pareto Distribution — almost regardless of what the original return distribution was. The body can be a messy, unknowable thing. The tail, once you isolate it, is law-abiding.

Now the formalism. Let your loss be a random variable, and pick a high threshold u. Define the exceedance as the amount by which a loss above u overshoots, that is, the loss minus u, conditional on having crossed u at all. The theorem says this exceedance follows a Generalized Pareto Distribution with two parameters. Its survival function — the probability the exceedance is larger than some value y — is written as one plus xi times y over beta, all raised to the power of negative one over xi. The scale parameter, beta, sets the size of typical exceedances, much as a standard deviation would. The shape parameter, xi — read "ksi" — is the heart of the matter. It controls how heavy the tail is.

When xi is zero, the tail decays exponentially and the GPD collapses to a plain exponential — thin-tailed, well-behaved, the world the normal distribution pretends we live in. When xi is positive, you have a genuinely heavy, power-law tail: extreme losses are far more likely than any Gaussian would admit, and — crucially — if xi is at least one-half, the variance of losses is theoretically infinite, and if xi reaches one, even the mean diverges. Financial return tails very often show xi somewhere between roughly zero-point-one and zero-point-four, which is firmly in heavy-tailed territory. A negative xi describes a tail with a hard upper bound, which is rare for losses but appears in bounded quantities.

Once you have fitted xi and beta to the exceedances — typically by maximum likelihood — the payoff is that you can compute risk measures analytically, beyond your data. The tail Value-at-Risk at a confidence level reads as the threshold u, plus beta over xi, times a particular ratio: the sample size times one minus the confidence level, divided by the number of exceedances, all raised to the power of negative xi, minus one. And Expected Shortfall — the average loss given that you have breached the VaR, the number you actually care about — has an even cleaner closed form under the GPD: it is the VaR divided by one minus xi, plus the quantity beta minus xi times u, again over one minus xi. Notice that one-minus-xi in the denominator. As xi climbs toward one, Expected Shortfall blows up. The shape parameter is not a technicality; it is the dial that says how catastrophic the catastrophe can be.

The one genuine art is choosing the threshold. Set it too low and you contaminate the fit with the body, where the GPD does not yet hold; set it too high and you starve the fit of exceedances and the estimates become noise. The standard discipline is a mean-excess plot — above the right threshold, the average overshoot grows linearly — together with a check that xi is stable as you nudge the threshold. This is exactly the judgment the platform automates, and exactly the judgment you must still audit by eye.

This is the throughline of the whole institute, sharpened. A loss number is only evidence an asset can move violently if it comes from a tail you have actually modeled rather than a quantile you happened to observe. EVT is how you turn three lonely data points in the far tail into a defensible statement about the once-in-a-decade day.