Advanced Loss Distributions and Tail Behavior
Analyze advanced loss distribution properties and tail behavior for Exam MAS-II.
Tail Classification
Loss distributions are classified by tail weight. Light-tailed distributions (exponential, normal) have moment generating functions that exist everywhere and their tails decay exponentially fast. Heavy-tailed distributions have MGFs that do not exist for all positive t, meaning their tails decay more slowly than any exponential. Subexponential distributions (Pareto, lognormal, Weibull with shape parameter tau less than 1) satisfy P(X_1 + X_2 > x) ~ 2*P(X_1 > x) as x approaches infinity, meaning the sum of two independent copies is driven by the maximum rather than by many moderate values.
This classification matters profoundly for insurance applications. For light-tailed distributions, aggregate losses are well-approximated by normal distributions via the central limit theorem, and risk measures are relatively easy to estimate. For heavy-tailed distributions, the central limit theorem converges slowly, extreme observations dominate aggregate behavior, and risk measures are difficult to estimate reliably. Property and casualty losses (especially liability and catastrophe exposures) frequently exhibit heavy-tailed behavior.
Regularly Varying Distributions
Regularly varying distributions form an important subclass of heavy-tailed distributions. A distribution is regularly varying with index negative alpha if P(X > x) ~ L(x) * x^{-alpha} as x approaches infinity, where L(x) is a slowly varying function (meaning L(tx)/L(x) approaches 1 as x approaches infinity for any positive t). The Pareto distribution with survival function (theta/(theta + x))^alpha is the canonical example, where L(x) is asymptotically constant.
The tail index alpha determines which moments exist: the k-th moment E[X^k] is finite if and only if k is strictly less than alpha. A Pareto with alpha = 3 has a finite mean and variance but infinite third moment. A Pareto with alpha = 1.5 has a finite mean but infinite variance. This has practical implications for estimation: if the variance is infinite, the sample mean converges to the true mean very slowly, and confidence intervals based on normal approximations are unreliable.
Subexponential Distributions
The subexponential property P(X_1 + X_2 > x) ~ 2*P(X_1 > x) has a striking interpretation: if two independent, identically distributed subexponential random variables sum to a large value, it is almost certain that one of them is large rather than both being moderately large. This extends to n independent copies: P(X_1 + ... + X_n > x) ~ n * P(X_1 > x). The practical consequence is that for a portfolio of heavy-tailed risks, aggregate extreme losses are driven by a single catastrophic event rather than by the accumulation of many moderate losses.
All regularly varying distributions are subexponential, but the reverse is not true. The lognormal distribution is subexponential but not regularly varying. The Weibull distribution with shape parameter less than 1 is also subexponential. The exponential distribution (Weibull with shape equal to 1) is the boundary case and is neither heavy-tailed nor subexponential. Distributions with shape parameter greater than 1 (such as the Weibull with tau greater than 1) are light-tailed.
Mean Excess Function and Diagnostics
The mean excess function e(d) = E[X minus d | X > d] is the expected overshoot above a threshold d, given that the threshold is exceeded. It is a powerful diagnostic for tail behavior. For the exponential distribution, e(d) is constant (the memoryless property). For light-tailed distributions, e(d) generally decreases toward zero. For heavy-tailed distributions, e(d) increases, reflecting the fact that the further into the tail you go, the larger the expected additional loss.
The mean excess plot graphs the empirical mean excess function against d. A linear, increasing mean excess plot is characteristic of a generalized Pareto distribution (GPD) tail, which suggests that a GPD model is appropriate for exceedances above a suitably chosen threshold. A decreasing mean excess plot suggests a light-tailed or bounded distribution. An erratic plot with no clear trend may indicate insufficient data in the tail or a mixture of different loss mechanisms.
The Hill Estimator
The Hill estimator estimates the tail index alpha from the k largest order statistics. Given ordered observations X_{(1)} <= X_{(2)} <= ... <= X_{(n)}, the Hill estimator is alpha_hat = k / (sum from i=1 to k of ln(X_{(n-i+1)} / X_{(n-k)})). This estimator assumes that the tail is regularly varying and uses only the k largest observations, treating the k-th largest as an implicit threshold.
Choosing k involves a bias-variance trade-off. A small k uses fewer observations, producing high variance but low bias (since only the most extreme observations, which are most representative of the true tail, are used). A large k includes more observations, reducing variance but potentially introducing bias if observations far from the tail are included. The Hill plot, which graphs alpha_hat against k, helps identify a stable region where the estimate is relatively constant across a range of k values. This stable region suggests an appropriate choice of k.
QQ Plots for Tail Assessment
Quantile-quantile (QQ) plots compare the empirical quantiles of the data against the theoretical quantiles of a reference distribution. For tail assessment, exponential and Pareto QQ plots are particularly useful. An exponential QQ plot graphs the log of the order statistics against the expected exponential quantiles. If the data follows an exponential distribution, the plot is linear. If the upper tail curves upward, the data has a heavier tail than the exponential, suggesting a Pareto or similar heavy-tailed model.
A Pareto QQ plot graphs log(X_{(n-i+1)}) against log((n+1)/i) for i = 1, ..., n. Linearity in the upper portion of this plot indicates a Pareto-type tail, and the slope estimates 1/alpha. These diagnostic plots complement the mean excess plot and provide visual confirmation of tail behavior before fitting a parametric model.
Common Pitfalls
Students sometimes confuse the definitions of heavy-tailed and subexponential. All subexponential distributions are heavy-tailed, but not all heavy-tailed distributions are subexponential (though counterexamples are rare and unlikely to appear on the exam). Another common error is misinterpreting the tail index: a smaller alpha means a heavier tail (fewer finite moments), not a lighter one.
When using the Hill estimator, be careful about the convention for the estimator formula. Some references estimate alpha directly, while others estimate 1/alpha (the extreme value index). Check which version your study materials use and be consistent. Also remember that the Hill estimator applies only to regularly varying (Pareto-type) tails. Using it on data from a lognormal or Weibull distribution can produce misleading results.
Study Tips
For Exam MAS-II, understanding tail behavior is crucial for risk measure calculation, reinsurance pricing, and capital modeling. Know how to identify, estimate, and model heavy tails. Practice classifying distributions as light-tailed, heavy-tailed, subexponential, or regularly varying. Work through Hill estimator calculations and mean excess plot interpretations. Be prepared to connect tail properties to practical consequences: why heavy tails make VaR non-subadditive, why aggregate losses are dominated by the maximum, and why standard estimation methods may fail when moments do not exist.