Mean Excess Plots and Tail Diagnostics

Before fitting any tail-risk model — extreme value theory, a Pareto tail, a fat-tailed VaR model — you need to answer a more basic question first: does this data even have a fat tail? Two loss distributions can have identical variance and yet behave completely differently 3 standard deviations out, one decaying quickly to nothing and the other still producing occasional catastrophic losses. The mean excess plot is the standard, model-free diagnostic for telling these apart before you commit to any particular tail model.

An analogy: how much worse does it get, once it's already bad?

Ask a friend "given that you're already running late, how much later on average do you end up?" For someone whose lateness is mild and self-correcting (they rush once they notice), the answer stays roughly constant no matter how late they already are — 10 minutes late or 30 minutes late, the additional expected delay is about the same, maybe 5 more minutes. For someone whose lateness spirals (missing one connection cascades into missing the next), the expected additional delay grows the later they already are — being 30 minutes late predicts a much bigger additional delay than being 10 minutes late. The mean excess function asks exactly this "given we're already past this threshold, how much more on average?" question of a loss distribution, and how the answer changes with the threshold is the tail diagnostic.

The mechanics, one piece at a time

For a random loss $X$ and a threshold $u$, the mean excess function is

$$e(u) = \mathbb{E}[X - u \mid X > u],$$

the expected amount by which $X$ exceeds $u$, given that it does exceed $u$. In plain English: pick a loss threshold, look only at the observations that blew past it, and average how far past it they went. The mean excess plot graphs the sample estimate of $e(u)$ against $u$ across a range of thresholds. Its slope is the diagnostic: a roughly flat mean excess plot indicates an exponential-like tail (memoryless — losses beyond a big threshold aren't systematically worse, in expectation, than losses beyond a smaller one). An upward-sloping plot indicates a heavy (Pareto-like) tail, where the expected overshoot keeps growing as the threshold rises — the tail gets proportionally worse, not just larger. A downward-sloping plot indicates a thin, bounded tail (like the normal or a distribution with a hard cutoff), where extreme excesses actually shrink in expectation as the threshold rises, because there's little probability mass left out there to draw from.

Worked example 1: reading a mean excess plot by hand

Suppose you compute mean excess at thresholds $1M, $2M, $3M, $4M for a portfolio's historical daily losses, getting $e(1M) = 0.4$, $e(2M) = 0.9$, $e(3M) = 1.6$, $e(4M) = 2.5$, all in millions of dollars. The mean excess is roughly doubling as the threshold increases by $1M each time and growing faster than linearly — a clear upward-sloping signature. This tells you, before fitting any formal model, that losses beyond $4M are not rare-but-tame events; the tail is heavy, and a normal-distribution-based VaR model (which implicitly assumes a downward-sloping mean excess) would badly underestimate the risk of losses beyond your historical sample's largest observed value.

Worked example 2: comparing two strategies' tails

Strategy X's mean excess at thresholds $100K, $200K, $300K is $45K, $48K, $44K — essentially flat. Strategy Y's mean excess at the same thresholds is $45K, $95K, $210K — clearly accelerating upward. Both strategies might show similar 95th-percentile VaR from a standard historical calculation, since the flat-vs-rising behavior only shows up further out in the tail than the 95th percentile typically reaches. But Strategy Y is carrying meaningfully more catastrophic-tail risk than X — a risk that a VaR number computed at a conventional 95% or 99% confidence level, sitting inside the region where the two still look similar, would completely miss.

What this means in practice

The mean excess plot is the standard first step before applying peaks-over-threshold extreme value theory: its shape tells you whether a heavy-tailed model is even warranted, and roughly where the tail "begins" — the threshold above which the plot turns reliably linear-upward is a natural choice for where to start fitting a generalized Pareto tail. It's cheap, model-free, and catches the common mistake of assuming a well-behaved (thin) tail purely because standard deviation or a normal-model VaR looks unremarkable.