Extreme Value Theory Fundamentals for MAS-I

Why Extreme Value Theory Matters

Extreme value theory (EVT) provides a rigorous mathematical framework for modeling the behavior of the largest (or smallest) values in a data set. In insurance, extreme losses from catastrophes, large liability claims, or rare but severe events drive the need for adequate reserves and reinsurance. Standard distributional models fitted to the bulk of the data may seriously underestimate tail probabilities. EVT addresses this by focusing specifically on the tail behavior, providing theoretically justified models for extreme events.

The key insight of EVT is that, just as the central limit theorem governs the behavior of sums (and averages), there are analogous limit theorems governing the behavior of maxima. These theorems tell us which distributions are possible limits for normalized maxima, regardless of the underlying distribution. This universality makes EVT applicable even when the underlying loss distribution is unknown.

The Block Maxima Approach

The block maxima approach divides data into non-overlapping blocks (typically years or quarters) and extracts the maximum from each block. The Fisher-Tippett-Gnedenko theorem states that if the normalized maximum M_n = (max(X_1,...,X_n) - b_n) / a_n converges in distribution (for appropriate normalizing sequences a_n > 0 and b_n), then the limit must be one of three types: the Gumbel distribution (Type I), the Frechet distribution (Type II), or the reverse Weibull distribution (Type III).

These three types are unified into the Generalized Extreme Value (GEV) distribution with CDF G(x) = exp(-(1 + xi*(x-mu)/sigma)^{-1/xi}) for 1 + xi*(x-mu)/sigma > 0, where mu is the location parameter, sigma > 0 is the scale parameter, and xi is the shape parameter. When xi > 0, the GEV becomes the Frechet (heavy-tailed). When xi = 0, it becomes the Gumbel (light-tailed, exponential decay). When xi < 0, it becomes the reverse Weibull (bounded upper tail).

Domain of Attraction

The domain of attraction of a GEV type is the set of parent distributions whose maxima converge to that type. Distributions with polynomial tail decay (like Pareto, Burr, and Student's t) belong to the Frechet domain (xi > 0). Distributions with exponential tail decay (like normal, exponential, gamma, and lognormal) belong to the Gumbel domain (xi = 0). Distributions with a finite right endpoint (like uniform and beta) belong to the Weibull domain (xi < 0).

For insurance applications, the Frechet domain is most relevant because loss distributions typically have heavy tails. A positive shape parameter xi means the distribution has a power-law tail, and the larger xi is, the heavier the tail. This is critical for estimating the probability of extreme losses that far exceed historical experience.

Peaks Over Threshold and the GPD

The peaks over threshold (POT) approach is often preferred over block maxima because it uses data more efficiently. Instead of using only one observation per block, it uses all observations that exceed a high threshold u. The Pickands-Balkema-de Haan theorem states that for a sufficiently high threshold, the distribution of excesses Y = X - u, given X > u, is approximately a Generalized Pareto Distribution (GPD).

The GPD has CDF H(y) = 1 - (1 + xi*y/sigma_u)^{-1/xi} for y > 0, where xi is the shape parameter (the same as the GEV shape) and sigma_u is a scale parameter that depends on the threshold. When xi > 0, the GPD is an ordinary Pareto-type distribution. When xi = 0, it reduces to the exponential. When xi < 0, it has a finite upper endpoint at -sigma_u/xi. The expected excess is E[Y | Y > 0] = sigma_u / (1 - xi) for xi < 1.

Threshold Selection

Choosing the threshold u involves a bias-variance trade-off. A higher threshold better satisfies the GPD approximation (less bias) but leaves fewer exceedances for estimation (more variance). A lower threshold provides more data but may introduce bias because the GPD approximation may not hold well for observations only slightly above u.

The mean excess plot is a key diagnostic tool. For a sample x_1, ..., x_n, the empirical mean excess function at threshold u is e_n(u) = (sum of (x_i - u) for x_i > u) / (number of x_i > u). For a GPD with xi > 0, the theoretical mean excess function is linear in u: e(u) = (sigma_u + xi*u) / (1 - xi). If the mean excess plot is approximately linear above some threshold u_0, the GPD is a reasonable model for exceedances above u_0. The threshold should be chosen where linearity begins.

Estimating Tail Probabilities

Once the GPD is fitted to exceedances above threshold u, you can estimate the probability that X exceeds any level x > u. The estimate is P(X > x) = P(X > u) * P(X > x | X > u) = (n_u / n) * (1 + xi_hat * (x - u) / sigma_hat)^{-1/xi_hat}, where n_u is the number of observations exceeding u and n is the total sample size. This formula enables extrapolation beyond the range of observed data, which is essential for estimating probabilities of events more extreme than anything in the historical record.

Return levels are another useful summary. The T-year return level x_T is the level expected to be exceeded on average once every T years. It solves P(X > x_T) = 1/T (or equivalently, in terms of the block maximum distribution, G(x_T) = 1 - 1/T). Return levels provide intuitive risk summaries for stakeholders who may not be comfortable with probability statements.

Common Pitfalls

A frequent mistake is fitting the GEV or GPD to data where the conditions of the limit theorems are not well approximated. The block size must be large enough for the GEV to be a reasonable model, and the threshold must be high enough for the GPD to apply. Using blocks that are too small or thresholds that are too low introduces bias.

Another error is confusing the three GEV types. Remember: xi > 0 corresponds to heavy tails (Frechet), xi = 0 to light exponential-type tails (Gumbel), and xi < 0 to a finite upper bound (Weibull). Insurance losses almost always fall in the xi > 0 or xi = 0 categories; a negative xi estimate from insurance data should prompt you to check the model carefully.

Study Tips

Exam MAS-I covers GEV and GPD fitting, threshold selection using mean excess plots, and tail probability estimation. Memorize the GEV and GPD distribution functions and their parameters. Understand the relationship between the GEV shape parameter and the GPD shape parameter (they are the same xi). Practice interpreting mean excess plots to identify appropriate thresholds. Know how to estimate tail probabilities and return levels using fitted EVT models. Be prepared to explain why EVT is necessary for modeling extreme insurance losses and how it complements standard distributional fitting.