GPD Tail Risk in Catastrophe Models

Frequency-severity models are how insurers, reinsurers, and increasingly banks put numbers on catastrophes. The frequency side is comparatively easy — Poisson processes, over-dispersion adjustments, seasonality. The severity side is where models earn their fees and lose their credibility, because the tail is where solvency lives and where the data is thinnest.

The workhorse for the tail is the Generalised Pareto Distribution (GPD) applied above a chosen threshold — the peaks-over-threshold approach from extreme value theory. Under fairly weak conditions, exceedances over a high enough threshold converge to a GPD regardless of the underlying loss distribution. That is the good news. The rest of this note is about the awkward parts.

Why the tail matters disproportionately

For a typical natural-catastrophe portfolio, most of the events sit in a fat body (single-family fire losses, minor wind, small floods) and a small number of tail events drive the 1-in-200 or 1-in-500 loss. Regulatory capital and reinsurance pricing both care almost exclusively about that tail. Getting the body right and the tail wrong produces a model that fits history nicely and understates capital dramatically.

Choosing the threshold — the hardest step

The GPD is only asymptotically true above a "high enough" threshold. Choose the threshold too low and the fit is contaminated by body data; choose it too high and you run out of exceedances and the shape parameter is estimated from noise.

The two workhorse diagnostics:

• Mean excess plot. Above a valid threshold, the mean of exceedances over any higher level should be approximately linear in that level. Look for the point where the plot becomes stably linear.
• Parameter stability plot. Fit the GPD at a range of thresholds and plot the shape parameter ξ against threshold. The right threshold sits inside a region where ξ is roughly constant.

Neither diagnostic gives a single answer. Threshold choice is a judgement call, and it should be documented, sensitivity-tested, and re-visited when the loss history extends. A model that quietly re-picks the threshold each year to keep numbers stable is not really a tail model — it is a smoothing exercise.

The shape parameter is everything

The GPD shape parameter ξ governs how quickly the tail decays.

• ξ < 0: bounded tail, losses cannot exceed some finite maximum. Rare in catastrophe work.
• ξ = 0: exponential tail. Moderate.
• ξ > 0: heavy Pareto-type tail. Every catastrophe portfolio worth modelling ends up here.

A ξ of 0.2 versus 0.4 does not sound like a big difference. On a 200-year return period it can double the loss estimate. This is why ξ needs a genuine confidence interval — not a maximum-likelihood point estimate reported to three decimals — and why parameter uncertainty is part of the reserving story, not an afterthought.

Parameter uncertainty is a first-class risk

The number of tail exceedances is small by construction — that is what makes them tail events. Fitting a two-parameter GPD to 40 or 60 exceedances gives wide confidence intervals on both parameters. Ignoring this uncertainty and reporting a single 1-in-200 number gives a false precision that will collapse the first time a real event lands outside the previous history.

Two practical tools:

• Profile likelihood confidence intervals for ξ and the return level. These are asymmetric and honest; standard-error intervals under-cover in the heavy-tail case.
• Bootstrap-based return-level intervals. Especially useful for communicating tail uncertainty to a non-quant audience.

Climate is now a first-class problem

Historical frequency and severity are no longer a reliable base for the future. Hurricane frequency in the North Atlantic, European windstorm clustering, wildfire severity in the western US, flood frequency in most inland basins — all show trends that break the stationarity assumption baked into a naive GPD fit. Two responses have become standard:

1. Detrend the loss history to a common climate baseline before fitting, then re-project onto a forward-looking baseline.
2. Layer a climate-scenario overlay on the fitted parameters, informed by RCP or SSP scenario ensembles (or NGFS scenarios for banks).

Neither is perfect. Both are better than ignoring the problem.

What a defensible tail model looks like

• Threshold choice justified by mean-excess and parameter-stability plots, and stability tested.
• Shape parameter reported with a profile-likelihood confidence interval, not a point estimate.
• Return-level uncertainty communicated as an interval and probed by bootstrap.
• Stationarity assumption tested and, where broken, replaced with an explicit detrend or overlay.
• Model output reconciled to actual reinsurance placements and any external vendor model, with divergences investigated rather than averaged away.

Catastrophe modelling is one of the very few places in quantitative finance where 90% of the answer sits in the last 5% of the data. That is what makes it hard. It is also what makes shortcuts here — a single threshold, a headline point estimate, a stationary fit — so costly the day a real event lands.