Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Extreme Value Theory for Insurance Losses

Abstract. Extreme value theory provides a mathematically disciplined account of the far upper tails of random losses. Insurance makes the inferential task distinctive: an observed claim may be a payment after deductibles and limits, an event loss may combine hundreds of insured properties, and a portfolio's annual maximum or aggregate total may depend on a changing population of risks. This monograph develops the foundations, assumptions, estimation methods, diagnostics, and decision consequences of extreme value analysis for insurance. It traces the limiting theory of block maxima and threshold excesses to their established primary literature; derives generalized extreme value (GEV) and generalized Pareto (GPD) laws; distinguishes the tail index from distributional scale and exposure; and connects individual-claim quantiles, event return levels, annual aggregate distributions, reinsurance layers, and solvency capital. A transparent synthetic numerical case calculates severity quantiles, expected shortfall, and annual occurrence return levels, with independent arithmetic checks and sensitivity analysis. Statistical uncertainty, threshold bias, sample selection, reporting thresholds, censoring, clustering, and changing hazard–exposure–vulnerability relations are examined as distinct sources of inferential risk. The central analytical result is a source-grounded evaluation: a credible insurance extreme-value estimate is conditional on a defined loss object, observation process, exposure clock, tail regime, dependence structure, and decision target. Extreme-value asymptotics justify candidate mathematical forms under stated conditions; they do not by themselves authenticate long return-period forecasts for a changing insured portfolio. The analysis draws upon foundational probability papers, peer-reviewed statistical methods, actuarial treatments, a selected European capital framework, and major climate-science assessments. All monetary examples are explicitly hypothetical. This research does not introduce a new theory, model, principle, or doctrine, and makes no claim to external peer review or newly collected insurance data.

Keywords: Extreme value theory; insurance losses; block maxima; generalized extreme value distribution; generalized Pareto distribution; peaks over threshold; regular variation; Fréchet tail; tail index; return level; exceedance probability; expected shortfall; value at risk; claim severity; aggregate loss; catastrophe insurance; reinsurance; tail uncertainty; threshold choice; declustering; extremal index; censoring; policy limits; exposure normalization; nonstationarity; climate change; solvency; model validation.

Introduction: Scientific Object and Research Method

The risk carried by an insurer is a contingent obligation whose magnitude and timing may be extraordinarily uneven. Thousands of small claims can coexist with rare, exceptionally costly losses. A flood can damage many properties at once; a liability claim can mature over years; a single insurance event can trigger direct coverage, multiple excess layers, and a sequence of reinstatements. The statistical objective must therefore be specified before a tail model is selected. An individual gross claim, a net payment, an occurrence loss, a yearly maximum event loss, a total annual loss, and the change in an insurer's economic own funds are separate random variables. Their upper tails need not have the same form, even when they originate in the same natural hazard.

Extreme value theory (EVT) asks what can be learned about unusually large observations from the probabilistic behavior of sample maxima and threshold exceedances. Its classical statistical foundations were established in Fisher and Tippett's analysis of limiting sample extrema, Gnedenko's classification of maximum limits, and the threshold work of Balkema and de Haan (1974) and Pickands (1975). The modern actuarial synthesis of Embrechts, Klüppelberg, and Mikosch (1997) situates these methods in the mathematical treatment of insurance and financial losses through Modelling Extremal Events. These works establish the rigorous vocabulary of attraction domains, tail decay, exceedance probabilities, and statistical extrapolation.

The scientific question is sharper than the search for a single very large number. What type of loss is measured? Which insurance exposures contributed to the experience? How were thresholds, data exclusions, and policy terms defined? Are extremes approximately independent over the time and space scale of interest? How sensitive is the 100-year level to tail-shape uncertainty? Which part of an extrapolation is identified by observed records, and which part rests upon assumptions? How does the answer change when an underwriting decision concerns an excess-of-loss layer, rather than an uncapped loss severity?

The research method combines five operations. First, the probability theory of maxima and conditional excesses is stated with its mathematical assumptions. Second, related statistical methods are compared using accessible original papers, publisher records, and authoritative scientific syntheses. Third, key equations are derived rather than treated as slogans. Fourth, numerical examples are explicitly synthetic and independently recalculated with a script using fixed input parameters. Fifth, the consequences are interpreted for claim distributions, frequency–severity aggregation, catastrophe occurrence models, reinsurance, and regulation. Scientific assertions about climate change are taken from assessed physical science, not inferred from insurance loss ratios.

The conceptual perimeter is univariate tail inference. The interaction of several insured losses is considered insofar as it changes the univariate event or portfolio loss constructed for analysis. Formal multivariate extremal dependence, copula identification, and the full theory of heavy-tailed ordinary severity distributions are related subjects with their own research dossiers. The present text makes their interfaces explicit without duplicating their entire fields. Sources are used in their appropriate roles: a limiting theorem provides a mathematical condition, a statistical article offers an estimator or diagnostic, a climate assessment supplies evidence about hazards, and a legal instrument states a jurisdiction-specific regulatory rule. None substitutes for the others.

Block Maxima: The Probabilistic Foundation

From single losses to extreme observations

Let X1,…,XnX_1,\ldots,X_n be independent, identically distributed random variables with cumulative distribution function FF. When a block contains nn observations, its maximum is Mn=max⁡(X1,…,Xn)M_n=\max(X_1,\ldots,X_n), so for every real xx,

Pr⁡(Mn≤x)=F(x)n.\Pr(M_n\leq x)=F(x)^n.

This identity is exact under independence. If the observations in one year are individual insurance claims, then the maximum individual claim during that year is the block maximum. If instead an observation is an entire storm's insured event loss, the annual maximum is the largest storm loss. Changing the observation changes both the distribution FF and the interpretation of the block. In a dependent sample, the exact formula F(x)nF(x)^n generally fails, even if every observation shares the same marginal FF.

A naive comparison of annual maxima from two portfolios can also be meaningless when their numbers of policies differ greatly. More insured locations generate more opportunities for extreme payouts even if the per-risk conditional severity law remains identical. Exposure enters through the number and composition of trials, the frequency of events, and the monetary consequences. This point is especially important when records aggregate business expansion, rising insured property values, and changing hazard intensity into one historical dollar series.

Max-stability and domains of attraction

The classical limit problem seeks sequences an>0a_n>0 and bnb_n such that (Mn−bn)/an(M_n-b_n)/a_n converges in distribution to a nondegenerate law GG. Fisher and Tippett (1928) established an early classification, and Gnedenko (1943) characterized the domains of attraction underlying the three traditional extreme-value types. If a suitable nondegenerate limit exists for independent identically distributed maxima, its distribution belongs to the generalized extreme value family. The statement is conditional: many useful real-world samples are dependent, non-identically distributed, and finite, and the limiting law alone does not determine the quality of a finite-sample approximation.

The max-stability property of a limiting extreme-value distribution means that the maximum of mm independent draws from that distribution, suitably shifted and scaled, remains in the same family. This is a structural characterization of the law of extremes, not a requirement that observed insurance claims themselves be GEV distributed. The role of normalization is fundamental: raw maxima may diverge as block size increases, while normalized maxima converge to a useful limit.

A useful distinction is the difference between the asymptotic theorem and the statistical model. The theorem concerns the limit under a growing block size. The model fits a three-parameter GEV law to finite block maxima. As Coles (2001) develops in his statistical introduction to extremes, the inferential result depends on whether the blocks are sufficiently large, sufficiently homogeneous, and approximately independent for the intended use. The fact that annual blocks are customary does not confer any special mathematical sufficiency upon 12 months of observations.

The block as an insurance design choice

Blocks may be defined by calendar year, underwriting year, catastrophe season, or a prespecified exposure unit. Calendar-year annual maxima simplify comparisons with capital and reporting horizons, but they may divide a single extended catastrophe across two reporting years. A wildfire season can cross calendar boundaries. Long-tail liability claims may be assigned according to accident, report, or payment year. Every choice creates a specific stochastic object and data-generating protocol.

The analyst must also determine whether the block maximum is a gross occurrence loss, a net retained loss after treaty recovery, or an inflation-adjusted normalized loss. Reinsurance and coverage limits can introduce upper support bounds or discontinuities. Insurance inflation is not necessarily identical to consumer price inflation, because construction-cost growth, litigation severity, repair technology, and exposure mix can change independently. An empirical series created by relabeling each historical extreme into current dollars needs an explicitly documented valuation procedure and sensitivity checks.

In a portfolio with NtN_t claims in period tt, an annual maximum Mt=max⁡1≤i≤NtXt,iM_t=\max_{1\leq i\leq N_t}X_{t,i} requires a convention for years with no claims. If NtN_t is stochastic, even conditional independence among severities does not imply that MtM_t has the same law as a maximum with a fixed count. For Poisson counts with rate λ\lambda and severity distribution FF, the positive-threshold maximum obeys Pr⁡(M≤x)=exp⁡[−λ{1−F(x)}]\Pr(M\leq x)=\exp[-\lambda\{1-F(x)\}] when xx exceeds the defined zero-loss baseline. This exact frequency–severity result will later connect a claim-level tail to annual occurrence return periods.

Estimation, fit, and effective sample size

For block maxima m1,…,mBm_1,\ldots,m_B, a stationary GEV can be fitted by maximum likelihood, subject to its parameter-dependent support. The resulting estimates are not automatically well behaved in small samples or at extreme shape values. Profile likelihood, return-level plots, quantile plots, residual diagnostics, and alternative block lengths help reveal instability; they cannot manufacture information absent from the observed maxima. Coles (2001) presents the statistical approach and the distinction between diagnostics and proof.

An analyst with 30 annual maxima possesses at most 30 block observations even if the underlying claim database contains millions of individual claims. The number of observations available to fit the tail is thus a design consequence. Blocks discard all observations except the largest per block, which may waste informative intermediate extremes. This loss of information motivates the peaks-over-threshold approach developed in the foundational threshold literature and, for statistical modeling, in Davison and Smith's high-threshold exceedance framework.

Generalized Extreme Value Laws and Tail Geometry

The unified GEV family

For location μ\mu, scale σ>0\sigma>0, and shape ξ\xi, the GEV cumulative distribution is

G(x;μ,σ,ξ)=exp⁡{−[1+ξ(x−μσ)]−1/ξ},1+ξ(x−μ)/σ>0.G(x;\mu,\sigma,\xi)=\exp\left\{-\left[1+\xi\left(\frac{x-\mu}{\sigma}\right)\right]^{-1/\xi}\right\},\quad 1+\xi(x-\mu)/\sigma>0.

The continuous limit at ξ=0\xi=0 is G(x)=exp⁡{−exp⁡[−(x−μ)/σ]}G(x)=\exp\{-\exp[-(x-\mu)/\sigma]\}. The sign of ξ\xi governs the tail class. For ξ>0\xi>0 (Fréchet type), the right tail decays polynomially and has no finite upper endpoint. For ξ=0\xi=0 (Gumbel type), the domain includes many unbounded light-tailed laws, including the normal and exponential under appropriate normalization. For ξ<0\xi<0 (reverse Weibull type), the limiting distribution has a finite right endpoint μ−σ/ξ\mu-\sigma/\xi. The support restriction is as important as the formal equation: a fitted negative shape parameter implies a specific endpoint within the chosen parametric model.

The terms Fréchet, Gumbel, and Weibull are names for limiting families; they are not causal explanations of insurance claims. A positive estimated GEV shape can suggest that a sufficiently homogeneous underlying loss variable has a heavy tail, but a finite record, rounding, a moving portfolio composition, or dependence can also influence the estimate. Tail extrapolation therefore requires mathematical suitability and evidence from the observation process.

Regular variation and moment existence

For a nonnegative loss severity XX with survival function F‾(x)=1−F(x)\overline F(x)=1-F(x), regular variation with index −α-\alpha means that for every t>0t>0,

lim⁡x→∞F‾(tx)F‾(x)=t−α,α>0.\lim_{x\to\infty}\frac{\overline F(tx)}{\overline F(x)}=t^{-\alpha},\qquad \alpha>0.

The tail index is commonly written ξ=1/α\xi=1/\alpha. This condition describes a ratio limit, not an assertion that the empirical loss distribution equals a Pareto law at every size. Slowly varying factors can materially affect finite-threshold behavior, which is why second-order convergence and threshold choice matter. The actuarial treatment in Embrechts, Klüppelberg, and Mikosch (1997) and Beirlant and colleagues' Statistics of Extremes explain the connection between regular variation, high quantiles, and risk aggregation.

If the upper tail behaves as Cx−1/ξCx^{-1/\xi} and XX is nonnegative, a moment of order p>0p>0 is finite in the regularly varying idealization when p<1/ξp<1/\xi; at equality the slowly varying component may decide the boundary case, so the strict inequality is the dependable general criterion. In a pure Pareto or GPD tail, an expected severity is infinite when ξ≥1\xi\geq 1 and a variance is infinite when ξ≥1/2\xi\geq 1/2. A finite insurer contract limit can, of course, produce finite payments even when the hypothetical gross-loss distribution lacks a moment. Consequently, gross-loss, contractual-claim, and net-retention moments must be studied separately.

Tail indices and actuarial measures

Tail shape governs how rapidly high quantiles diverge with increasing confidence. Consider two models calibrated to the same exceedance probability and scale at an observed threshold. If the shape parameters differ, their near-threshold predictions may be similar while their remote-tail quantiles diverge substantially. Since a high return-level estimate depends exponentially on the logarithm of the extrapolation factor, small changes in ξ\xi can make large differences in a 100- or 500-year claim assessment.

For a continuous severity loss XX, the value at risk at probability pp is VaR⁡p(X)=inf⁡{x:F(x)≥p}\operatorname{VaR}_p(X)=\inf\{x:F(x)\geq p\}. For integrable XX, expected shortfall may be defined as ES⁡p(X)=(1−p)−1∫p1VaR⁡v(X) dv\operatorname{ES}_p(X)=(1-p)^{-1}\int_p^1\operatorname{VaR}_v(X)\,dv. This quantile-integral formulation handles atoms more carefully than the casual phrase “the mean loss conditional on exceeding VaR.” For a continuous distribution at its quantile, the latter agrees with the former. Artzner and colleagues' analysis of coherent risk measures clarifies the limitations of quantile-only risk summaries. A heavy-tailed model with infinite mean also makes the unbounded-loss ES infinite.

Insurance capital calculations may require distributional information beyond a marginal severity tail. Frequency, cross-line aggregation, investments, the balance-sheet valuation horizon, and reinsurance collectibility can all determine the distribution of economic losses. No theorem relating individual maxima to GEV laws converts a stand-alone severity quantile directly into a regulatory solvency capital requirement.

Nonstationary and covariate-based GEV inference

A stationary GEV assumes fixed μ\mu, σ\sigma, and ξ\xi across fitted blocks. Where changes in hazard, exposure, or claims handling are material, models may let selected parameters depend on time or measured covariates, such as μt=β0+β1t\mu_t=\beta_0+\beta_1t or log⁡σt=γ0+γ1zt\log\sigma_t=\gamma_0+\gamma_1 z_t. These are statistical specifications requiring substantive justification, stable covariate definitions, support checks, and diagnostics. Introducing too many smooth trend coefficients into a small set of annual extremes can fit noise and obscure structural changes.

Katz, Parlange, and Naveau (2002) review statistics of extremes in hydrology, including maximum likelihood with covariates and the analysis of trends in extremes. Their subject illustrates the inferential issue: a physical extreme may change over time while the loss produced by that extreme changes for additional reasons. In insurance, the distinction between the hazard distribution and the monetary loss distribution is particularly consequential. One can estimate the probability of heavy rainfall correctly and still misestimate the probability of a large insured loss because the location, reconstruction cost, policy terms, or vulnerability of insured structures has changed.

Threshold Exceedances and Tail Estimation

The excess law and its interpretation

Rather than selecting only one maximum per block, the peaks-over-threshold (POT) method uses observations above a high threshold uu. Define the excess Y=X−uY=X-u conditional on X>uX>u. Its exact conditional distribution is Pr⁡(Y≤y∣X>u)={F(u+y)−F(u)}/{1−F(u)}\Pr(Y\leq y\mid X>u)=\{F(u+y)-F(u)\}/\{1-F(u)\}. Under broad conditions tied to the appropriate max-domain of attraction, this excess distribution is approximately generalized Pareto for sufficiently high uu:

Hξ,σu(y)=1−(1+ξyσu)−1/ξ,y≥0,1+ξy/σu>0.H_{\xi,\sigma_u}(y)=1-\left(1+\frac{\xi y}{\sigma_u}\right)^{-1/\xi},\qquad y\geq0,\quad1+\xi y/\sigma_u>0.

For ξ=0\xi=0, the limiting expression is H(y)=1−exp⁡(−y/σu)H(y)=1-\exp(-y/\sigma_u). The GPD scale σu\sigma_u is threshold-specific, while the shape ξ\xi is invariant across sufficiently high thresholds in an exact GPD tail. Balkema and de Haan's residual-life limit analysis and Pickands' inference using extreme order statistics provide seminal foundations. These asymptotic results justify the candidate family and its limiting interpretation; their approximation accuracy for a particular insurer's uu must be investigated empirically.

The distinction between excess magnitude and exceedance frequency is essential. A GPD conditional distribution describes Y∣X>uY\mid X>u. It does not independently determine how often XX exceeds uu. To obtain an unconditional exceedance probability above a distant level x>ux>u, the analyst also needs pu=Pr⁡(X>u)p_u=\Pr(X>u). Combining the two gives Pr⁡(X>x)≈pu[1+ξ(x−u)/σu]−1/ξ\Pr(X>x)\approx p_u[1+\xi(x-u)/\sigma_u]^{-1/\xi}, within the support. In a policy database, an observed k/nk/n estimates pup_u only if the observations constitute a suitable sampling frame for the defined claim population.

Estimation of the rate and the excess distribution

Davison and Smith (1990) develop likelihood-based models for exceedances over high thresholds, jointly attending to their magnitude, occurrence, regression structure, seasonality, and dependence. A simple estimation scheme uses nn observations, threshold uu, and kk exceedances; estimates pu≈k/np_u\approx k/n; and fits σu\sigma_u and ξ\xi to the positive excesses y1,…,yky_1,\ldots,y_k. Under conditional independence, the GPD log-likelihood for ξ≠0\xi\neq0 is

ℓ(ξ,σu)=−klog⁡σu−(1+1ξ)∑i=1klog⁡(1+ξyiσu),\ell(\xi,\sigma_u)=-k\log\sigma_u-\left(1+\frac{1}{\xi}\right)\sum_{i=1}^k\log\left(1+\frac{\xi y_i}{\sigma_u}\right),

with σu>0\sigma_u>0 and all support terms positive. The zero-shape limit gives an exponential excess likelihood. A frequency model for the number of exceedances may also be necessary when the calendar time, number of policies, or number of claim-generating events changes. A rate measured “per claim” and a rate measured “per policy-year” cannot be interchanged without a count model.

Likelihood inference near the support boundary can be nonregular, and model diagnostics must look for influential extremes. Extreme observations contain much of the tail information yet can strongly alter a fitted ξ\xi. Analysts should report the actual exceedance count, threshold, maximum and second-maximum, estimation method, and uncertainty intervals; a single reported tail-shape estimate without these details is difficult to audit.

Threshold stability, Hill estimates, and diagnostics

The threshold-stability identity for an exact GPD tail states that for v>uv>u, σv=σu+ξ(v−u)\sigma_v=\sigma_u+\xi(v-u), with the same shape ξ\xi. Equivalently, the modified scale σu−ξu\sigma_u-\xi u should be stable under changing uu. This generates an operational diagnostic: fit candidate thresholds and inspect shape and adjusted-scale stability, while keeping track of shrinking kk. A smooth-looking plot at high thresholds can reflect wide uncertainty as readily as true parameter stability.

For positive, regularly varying tails, Hill's general tail-estimation approach is a classical estimator of the extreme-value index based on the largest order statistics. With sorted observations X(1)≤⋯≤X(n)X_{(1)}\leq\cdots\leq X_{(n)}, its common form is ξ^H(k)=k−1∑j=1klog⁡[X(n−j+1)/X(n−k)]\widehat\xi_H(k)=k^{-1}\sum_{j=1}^k\log[X_{(n-j+1)}/X_{(n-k)}], for an appropriate heavy-tailed regime. Hill estimation depends sensitively on the chosen kk, and it does not cover every Gumbel or finite-endpoint tail. Drees, de Haan, and Resnick's Hill-plot analysis addresses the interpretation of these estimator traces. Danielsson and coauthors' bootstrap selection of tail fraction illustrates a systematic choice mechanism under its own asymptotic conditions.

Other informative checks include mean residual life plots, empirical-to-fitted quantile plots, exceedance probability plots, profile likelihood of return levels, threshold-by-threshold estimates, and comparisons of competing tail forms. Scarrott and MacDonald (2012), in their review of threshold uncertainty, explain why the low-threshold approximation bias and high-threshold sampling variance form a joint estimation problem. A diagnostic is evidence about fit conditional on a model; it is not evidence that the omitted region beyond the largest observed loss has been uniquely identified.

Censoring, truncation, coverage transformations, and dependence

Insurance data seldom constitute an unmodified sample of gross physical losses. A deductible dd produces a payment function P(X)=max⁡(X−d,0)P(X)=\max(X-d,0) before applying any further limit. A per-claim payment cap LL produces P(X)=min⁡{max⁡(X−d,0),L}P(X)=\min\{\max(X-d,0),L\}, subject to the contract's precise wording. A capped payment distribution has an atom at LL if gross losses above the exhaustion point are recorded at the limit; treating those values as independent continuous uncapped observations distorts the estimated tail. Top coding, left truncation below a reporting threshold, missing open-claim estimates, and unequal incurred-versus-paid development each require explicit likelihood or data adjustments.

Threshold exceedances may also cluster. Consecutive insured event losses can share a storm system, and multiple payments can arise from one catastrophe. The standard independent-Poisson occurrence approximation must then be tested, modified by declustering or an appropriate dependence model, or restricted to independently defined events. Leadbetter, Lindgren, and Rootzén (1983) provide the classic framework for extremes of dependent sequences. Northrop (2015) examines the extremal index and block maxima, where the extremal index quantifies aspects of local clustering under specified stationary-process conditions. It is inappropriate to multiply an arbitrary insurance catastrophe rate by a single estimated extremal index without showing that the relevant mathematical setting holds.

Return Levels and Reproducible Insurance Calculations

Several meanings of a return period

The phrase “100-year event” is incomplete without a specified random variable, unit of time, exposure basis, and interpretation of recurrence. If the distribution of annual maxima is stationary and each year is independent, a TT-year return level zTz_T is commonly defined by Pr⁡(Myear>zT)=1/T\Pr(M_{\mathrm{year}}>z_T)=1/T. The expected waiting time to the next annual exceedance under a geometric waiting-time model is then TT years; this does not mean that the level occurs once every TT years on a schedule. The probability of at least one exceedance in hh independent years is 1−(1−1/T)h1-(1-1/T)^h. For T=100T=100 and h=30h=30, the probability is about 26%, not 30/100 exactly.

An alternative point-process definition uses a threshold crossed by an average of 1/T1/T independent occurrences per year. For a Poisson process with mean annual exceedance count rr, the annual probability of one or more exceedances is 1−e−r1-e^{-r}. Setting r=1/Tr=1/T yields an annual exceedance probability 1−e−1/T1-e^{-1/T}, which is close to but distinct from 1/T1/T. The distinction matters most at modest return periods, for processes that can produce multiple events in one year, and when different vendor models report annual occurrence exceedance probability (OEP) instead of a mean interarrival period.

A GEV annual-maxima quantile can be derived by setting G(zT)=1−1/TG(z_T)=1-1/T. For ξ≠0\xi\neq0,

zT=μ+σξ{[−log⁡(1−1/T)]−ξ−1}.z_T=\mu+\frac{\sigma}{\xi}\left\{[-\log(1-1/T)]^{-\xi}-1\right\}.

At ξ=0\xi=0, zT=μ−σlog⁡[−log⁡(1−1/T)]z_T=\mu-\sigma\log[-\log(1-1/T)]. All parameters refer to the same annual-maxima distribution, data basis, and stationary or explicitly conditional regime. These are distribution quantiles, and their uncertainty grows as TT greatly exceeds the length of the observed record.

Synthetic claim-severity case

Consider a hypothetical insurer with n=10,000n=10{,}000 independently recorded gross individual claim severities and k=250k=250 claims above u=\1.000$ million. Assume, solely for illustration, that the sampling mechanism identifies a stable claim population, that the exceedance rate is pu=0.025p_u=0.025, and that excess severity is exactly GPD with scale \sigma_u=\0.800$ million and shape ξ=0.25\xi=0.25. These parameters are stipulated, not estimated from a real insurer. The threshold-exceedance identity gives, for x>ux>u,

Pr⁡(X>x)≈0.025[1+0.25x−1.0000.800]−4,\Pr(X>x)\approx0.025\left[1+0.25\frac{x-1.000}{0.800}\right]^{-4},

where monetary values in the bracket are in millions of dollars. Solving for the severity quantile qpq_p at percentile pp produces

qp=u+σuξ[(pu1−p)ξ−1],1−p<pu.q_p=u+\frac{\sigma_u}{\xi}\left[\left(\frac{p_u}{1-p}\right)^\xi-1\right],\qquad 1-p<p_u.

Consequently, the synthetic 99.9th-percentile individual claim is approximately \4.955$ million; the 99.95th percentile is approximately \6.309$ million; and the 99.99th percentile is approximately \10.524$ million. These are claim-count quantiles: one would expect about one exceedance of the 99.9th-percentile level per thousand independent claims, not necessarily once per thousand calendar years. The sample has 10,000 claims but only 250 threshold exceedances available to identify the fitted GPD, and extrapolation to a one-in-10,000 claim tail is much more remote than the observed threshold probability.

Synthetic quantity Definition Result, USD millions
Threshold uu 1.000
GPD scale σu\sigma_u 0.800
Shape ξ\xi 0.250
Claim exceedance probability at threshold k/nk/n 0.025
Claim severity 99.9th percentile q0.999q_{0.999} 4.955
Claim severity 99.95th percentile q0.9995q_{0.9995} 6.309
Claim severity 99.99th percentile q0.9999q_{0.9999} 10.524
Claim severity expected shortfall at 99.9% ES0.999\mathrm{ES}_{0.999} 7.341

The numerical precision shown is arithmetic precision conditional on stipulated parameters; it is not a statement about uncertainty in a fitted insurance model. The figures can be reproduced with the formulas and the independent computation checks recorded in the evidence file. Including the assumed data-generation process prevents the synthetic illustration from being mistaken for a market estimate or a regional catastrophe benchmark.

Annual maxima and catastrophe-frequency interpretation

Now add a second illustrative assumption: the insurer experiences a Poisson mean of λ=1,000\lambda=1{,}000 individual claims per year, independently of severity, with the same conditional severity tail. The assumed annual expected count above the \1$ million threshold is λpu=25\lambda p_u=25. For the year's largest individual claim, the compound-Poisson identity gives

Pr⁡(Myear≤x)=exp⁡{−λpu[1+ξ(x−u)σu]−1/ξ}.\Pr(M_{\mathrm{year}}\leq x)=\exp\left\{-\lambda p_u\left[1+\frac{\xi(x-u)}{\sigma_u}\right]^{-1/\xi}\right\}.

To obtain an annual OEP return level with Pr⁡(Myear>xT)=1/T\Pr(M_{\mathrm{year}}>x_T)=1/T, let aT=−log⁡(1−1/T)a_T=-\log(1-1/T) and solve:

xT=u+σuξ[(λpuaT)ξ−1].x_T=u+\frac{\sigma_u}{\xi}\left[\left(\frac{\lambda p_u}{a_T}\right)^\xi-1\right].

The resulting hypothetical 100-year annual maximum individual-claim level is approximately \20.4$ million. It represents a one-percent annual chance that at least one individual claim exceeds the level under the stipulated Poisson, independent-severity, and stationary assumptions. It is not the 100-year annual total loss, catastrophe event loss, or reinsured-net loss. A high-frequency book may have a much higher annual extreme than a small book with the same per-claim severity law because the number of opportunities differs. Conversely, bundling the same claims into catastrophe events changes the random variable and dependence structure.

If the process is nonstationary, a fixed 100-year label can conceal changing annual probabilities. For a future sequence of conditional annual exceedance probabilities p1,…,php_1,\ldots,p_h, the probability of at least one exceedance, assuming conditional independence, is 1−∏t=1h(1−pt)1-\prod_{t=1}^h(1-p_t). This expression displays the time dependence explicitly. A single historical return period should not be passed off as a stable forecast when exposure or hazard changes substantially.

Tail risk measures and treaty layers

Under an exact GPD excess law and ξ<1\xi<1, the conditional mean excess beyond any q>uq>u is σq/(1−ξ)\sigma_q/(1-\xi), where σq=σu+ξ(q−u)\sigma_q=\sigma_u+\xi(q-u). Thus for a continuous severity distribution and tail quantile qp>uq_p>u,

ES⁡p(X)=qp+σu+ξ(qp−u)1−ξ.\operatorname{ES}_p(X)=q_p+\frac{\sigma_u+\xi(q_p-u)}{1-\xi}.

For the synthetic 99.9th percentile, q_{0.999}\approx\4.955$ million, and the modeled mean claim severity beyond that quantile is approximately \7.341$ million. Expected shortfall is higher than the quantile because it averages the remaining tail. In a model with ξ≥1\xi\geq1, the unbounded theoretical mean excess is infinite, so reporting a finite ES requires an explicitly bounded exposure or alternative tail structure. McNeil and Frey's (2000) tail-risk estimation study, although concerned with financial time series, offers a relevant methodological warning: volatility dynamics and tail behavior must both be modeled when conditional risk measures are sought. Its empirical results should not be transferred directly to insurance severities without verifying the different data process.

An insurer is also concerned with layer loss. For a hypothetical occurrence loss XX and excess-of-loss retention aa with layer width ww, the ceded payment is C(X)=min⁡{(X−a)+,w}C(X)=\min\{(X-a)_+,w\}, omitting special clauses, reinstatements, and aggregate features. Its expected layer loss is expressible by the survival function:

E[C(X)]=∫aa+wPr⁡(X>x) dx.\mathbb{E}[C(X)]=\int_a^{a+w}\Pr(X>x)\,dx.

This identity follows from the layer-cake representation for a nonnegative bounded payment. If a≥ua\geq u and the GPD approximation applies over the layer, the integral can be evaluated analytically; for ξ≠1\xi\neq1,

E[C(X)]≈puσu1−ξ{[1+ξ(a−u)σu]1−1/ξ−[1+ξ(a+w−u)σu]1−1/ξ},\mathbb{E}[C(X)]\approx\frac{p_u\sigma_u}{1-\xi}\left\{\left[1+\frac{\xi(a-u)}{\sigma_u}\right]^{1-1/\xi}-\left[1+\frac{\xi(a+w-u)}{\sigma_u}\right]^{1-1/\xi}\right\},

with the appropriate logarithmic limit at ξ=1\xi=1. This formula describes expected layer loss per underlying severity observation. To obtain expected annual ceded loss, it must be combined with annual claim frequency and any occurrence or aggregate dependencies. If the layer corresponds to a catastrophe occurrence rather than each claim, the event-loss distribution must be used. The formula is exact under the stipulated GPD tail and cover transformation; its application to actual treaty pricing must also account for expenses, uncertainty loads, reinstatements, and contract conditions.

Sensitivity of distant extrapolations

The numerical example shows why a remote quantile should be reported as a conditional scenario, not a naked point estimate. Holding pu=0.025p_u=0.025, u=1u=1, and σu=0.8\sigma_u=0.8 million fixed, compare the severity 99.99th percentile under ξ=0.15\xi=0.15, 0.250.25, and 0.350.35. The formula uses pu/(1−p)=250p_u/(1-p)=250. The three tail assumptions yield sharply different remote quantiles, even though they agree on the exact exceedance probability at uu and have the same local scale there. This is a structural shape sensitivity exercise, not a confidence interval and not an empirical estimate.

The local differential makes the effect transparent. Put A=log⁡[pu/(1−p)]A=\log[p_u/(1-p)]. Then qp=u+σu[exp⁡(ξA)−1]/ξq_p=u+\sigma_u[\exp(\xi A)-1]/\xi, and for ξ≠0\xi\neq0,

∂qp∂ξ=σuξ2[ξAexp⁡(ξA)−exp⁡(ξA)+1].\frac{\partial q_p}{\partial\xi}=\frac{\sigma_u}{\xi^2}\left[\xi A\exp(\xi A)-\exp(\xi A)+1\right].

For high quantiles, AA is large and positive, so shape uncertainty becomes especially consequential. Tail-shape sensitivity alone understates the full risk because the estimated threshold exceedance rate, changing exposure, scale, dependence, and model class also have uncertainty. A mathematically computed 500-year level may be useful for scenario comparison while remaining poorly identified by a short historical data set.

Threshold, Sample, and Model Uncertainty

Bias versus variance is the governing tension

Extreme-value approximations are asymptotic, and threshold choice mediates a fundamental conflict. A lower threshold provides more excesses and usually reduces statistical variance while increasing the possibility that the GPD approximation is biased by the body of the distribution. A higher threshold improves the intended asymptotic regime in favorable conditions but reduces the number of observations and makes parameter estimates more volatile. This trade-off is a central conclusion of Scarrott and MacDonald's (2012) threshold-estimation review. It also explains why choosing a threshold solely because one graph looks straight is scientifically weak.

The relevant question is which thresholds yield inferences that are both stable enough and supported by a credible sampling regime. A rigorous analysis can predefine a family of plausible thresholds, report fits at each, and carry threshold variation into its final interval or model-uncertainty assessment. Automated threshold choice procedures may help discipline judgment, but they rely on assumptions about convergence and data dependence. In particular, selecting the threshold after viewing an exceptional loss can bias inference unless the selection procedure is represented in the analysis.

Likelihood, uncertainty intervals, and extrapolation

Parameter standard errors obtained from an asymptotic Hessian provide one approximation to uncertainty. For rare-tail quantities, profile-likelihood intervals are often more informative because the likelihood in shape and return-level parameters can be skewed, especially near support constraints. Bootstrap intervals are another possibility when the resampling unit respects the original dependence and observation process. Resampling individual payments independently from a storm's claims is not a valid proxy for resampling independently occurring storms when the event is the actual risk unit.

Bayesian methods may formally combine prior information and sparse tail data, and they can yield posterior predictive distributions for remote levels. A narrow credible interval, however, requires scrutiny of prior strength and model completeness. Coles and Tawn's (1996a) Bayesian extreme-rainfall analysis demonstrates the possible influence of expert prior information; its documented rainfall inference example concerns environmental extremes, not an automatically portable insurance-capital calibration. Whenever the data are weak, a prior can be substantively decisive and should be shown through sensitivity tests.

Two uncertainty targets should be distinguished. A confidence or credible interval for the unknown TT-year return level expresses uncertainty about a distributional parameter-derived functional under a model. A prediction interval for the maximum loss in a future period additionally incorporates future randomness. The intervals answer different questions and should not be relabeled as one another. A third quantity, a model-class or scenario range, expresses differences among plausible structures; it has no probability coverage unless a justified weighting or confidence procedure is supplied.

Sampling design, validation, and adversarial tests

Insurance data are frequently available only as a portfolio of paid or incurred claims after administrative transformations. Large claims may settle over longer periods, creating right censoring in snapshots. Industry data may mix insurer portfolios with different retention, reporting, and exposure rules. Catastrophe catalogues may revise historical event estimates as new exposure and damage information appears. A scientific tail study should create an auditable data dictionary including currency basis, gross/net status, losses per claim or occurrence, valuation dates, thresholds, exposure measures, missing events, and policy-limit treatment before fitting any model.

Validation has multiple layers. Distributional checks compare tail empirical behavior and fitted excesses. Temporal checks compare predictive behavior across genuinely separated periods without contaminating training with future information. Spatial or portfolio checks examine transportability to distinct insured books. Actuarial consistency checks reconcile the modeled tail with known aggregate loss, claims counts, policy limits, and retained layers. Decision checks ask whether plausible changes in tail parameters alter the recommendation materially. Gilleland and Katz's (2016) extRemes software paper documents a reproducible computational environment for univariate extreme-value analysis, but a software pass is distinct from validation of insurance data construction.

Useful adversarial tests include excluding the single largest observation, shifting the threshold over a disclosed range, changing the block definition, reestimating after inflation and exposure normalization, modeling calendar and accident years separately, and comparing occurrence maxima with annual aggregate loss. These tests help reveal how much of a remote estimate is supported by recurrent tail behavior rather than by one disputed or changing observation. They are particularly valuable when a fitted point estimate is likely to influence retention, solvency, or insurance-linked security pricing.

Structural change, tail ambiguity, and epistemic boundaries

Historical stability is a substantive assumption. Climate change can alter the distribution of hazards; building practices and land use can change vulnerability; inflation and accumulation can transform monetary consequences; underwriting can change the covered population; and changes in legislation or litigation can affect liability severity. Observing a time trend in insured losses does not uniquely identify any one of those causes. A causal attribution exercise needs physical and institutional evidence beyond fitting a trend to past payouts.

The Intergovernmental Panel on Climate Change's AR6 assessment of weather extremes documents observed changes and evidence of human influence for multiple classes of physical extremes, with confidence differing by peril and region. The 2023 Synthesis Report findings connect human-driven climate changes to widespread risks and losses. These assessments justify testing nonstationary hazard assumptions in relevant applications; they do not provide a universal multiplier for every insurer's loss tail. The scientific chain from climate driver to peril, exposure, damage function, policy attachment, and booked claim must be reconstructed at each step.

Several distinct epistemic uncertainties persist even after mathematical fitting: uncertainty over whether the observed tail has reached its asymptotic domain; uncertainty over whether unobserved catastrophes are physically possible; uncertainty over the future portfolio and its coverage; and uncertainty over whether rare loss mechanisms can be represented by past data. Increasing numeric precision does not settle these questions. A credible deliverable records model disagreement and provides decision sensitivities at a scale commensurate with available evidence.

Catastrophe Insurance, Reinsurance, and Capital Applications

Translating physical extremes into insured loss

A catastrophe loss is produced through a chain of mechanisms. A physical event has a location, footprint, intensity, and duration. Exposed properties and people have values and vulnerability functions. Insurance contracts specify which consequences are covered and how indemnities are calculated. Claims administration determines the recorded monetary amounts and the timing of their revisions. Therefore, an EVT model of extreme rainfall, maximum wind speed, or peak flood height is situated upstream of an EVT model of insured losses. The tail of one is not mathematically identical to the tail of the other.

Consider a schematic insured event loss Le=∑j∈Eeg(He,j,Vj,Cj)L_e=\sum_{j\in\mathcal E_e}g(H_{e,j},V_{j},C_j), where He,jH_{e,j} denotes local hazard intensity affecting insured risk jj during event ee, VjV_j denotes property value and vulnerability characteristics, and CjC_j denotes policy contract terms. The payout function gg maps those characteristics to insured loss after the relevant clauses. This expression is an accounting and analytical decomposition, not a newly introduced probabilistic theory. Its purpose is to expose the causal and contractual links that must be audited before any fitted event-loss tail can be interpreted. A changeset in the location portfolio Ee\mathcal E_e, the hazard function, or coverage terms can change LeL_e even if the distribution of a single hazard metric remains stable.

A physically based catastrophe model can support rare-event simulation by combining hazard catalogs with geographically resolved exposure and vulnerability. Extreme-value methods may validate the simulated upper tail, estimate peril-specific thresholds, or extrapolate a residual quantity for which asymptotic conditions are defensible. The method does not guarantee that a catastrophe catalog includes every plausible event mechanism. A catastrophe portfolio with a large, concentrated coastal exposure will respond differently from one with the same total insured value dispersed inland, even if both use the same marginal wind-speed tail. This is why the spatial and temporal definition of the event process matters before the numerical return period.

The statistical problem is illustrated in Coles and Tawn's (1996b) areal rainfall extremes study. That paper concerns inference for spatially aggregated rainfall from local observations, making spatial dependence central to the target quantity. Its direct subject is rainfall, rather than insured money. The insurance application requires a further transition through exposure, vulnerability, and contract mappings. The IPCC's climate-extremes assessment similarly distinguishes physical-hazard evidence from what insurers must infer about protected assets and financial outcomes.

Event losses, clusters, and spatial footprints

The reporting unit in a catastrophe program is often an occurrence, not an individual payment. One hurricane may give rise to many claims over different policy classes and geographic districts. A simple claims-level POT fit to the same individual payments may seriously understate dependence among the largest annual portfolio losses. Conversely, aggregating by overly broad named event can merge physically distinct loss processes and obscure the frequency of a coverage-defined occurrence. The event taxonomy should be fixed from the contract, perils, and physical process before determining what constitutes an independent exceedance.

Clustering can also occur at the event level. A season may contain several related storm systems, aftershocks, wildfire ignitions during persistent drought, or floods produced by repeated precipitation within a short period. In a stationary time series meeting certain mixing conditions, an extremal index can alter the limiting behavior of maxima. Northrop (2015) provides one estimation approach; Leadbetter, Lindgren, and Rootzén (1983) develop the underlying dependent-sequence framework. Neither source licenses a universal numerical adjustment for any observed insurance time series. Dependence must be examined at the resolution at which an insurer's risk and treaty response are defined.

The distinction between the largest single event in a year and the aggregate event loss is essential. Let NN be the annual event count and L1,…,LNL_1,\ldots,L_N the event losses. The annual occurrence maximum is M=max⁡i≤NLiM=\max_{i\leq N}L_i, while the annual aggregate is S=∑i=1NLiS=\sum_{i=1}^NL_i. For a Poisson event process with independent, identically distributed event severities, Pr⁡(M≤x)=exp⁡{−λPr⁡(L>x)}\Pr(M\leq x)=\exp\{-\lambda\Pr(L>x)\} for thresholds above the relevant zero-loss point. The compound-sum distribution of SS requires convolution or other aggregation techniques; the distribution of MM alone generally does not determine it. The difference becomes material when a year experiences several medium-sized events with aggregate consequences exceeding those of its single largest catastrophe.

For a subexponential severity distribution under appropriate independence and light-tailed claim-count conditions, the extreme right tail of a compound sum is asymptotically dominated by one exceptionally large summand, yielding a relation of the form Pr⁡(S>x)∼E[N]Pr⁡(L>x)\Pr(S>x)\sim\mathbb E[N]\Pr(L>x) as x→∞x\to\infty. This is a recognized heavy-tail asymptotic described in the actuarial mathematics literature by Embrechts, Klüppelberg, and Mikosch (1997). It applies only in its stated regime. It is not a finite-threshold equality for an arbitrary annual insurance portfolio, and it is not a generalization to strongly dependent catastrophe events or common-shock losses. At practical capital levels, multiple events, cross-line dependence, and treaty structures can still control the aggregate tail.

Retention, exhaustion, and annual aggregate risk

Reinsurance transforms a loss through attachments, limits, shares, event definitions, and annual aggregate provisions. For a nonproportional occurrence treaty with attachment aa and width ww, the ceded occurrence payment is min⁡{(L−a)+,w}\min\{(L-a)_+,w\} under its simplified terms. The treaty is activated when L>aL>a, and the per-event layer becomes fully exhausted when L≥a+wL\geq a+w. The model supplies both frequencies through the occurrence loss survival curve. However, annual ceded loss may be constrained by reinstatements, hours clauses, aggregate deductibles, aggregate limits, and the dependence among occurrences. Treaty loss must therefore be constructed from contract wording and event history before its tail is fitted.

A gross-loss EVT fit does not imply that the corresponding net retained loss is GEV or GPD at the same parameter values. A cap can truncate the right tail and concentrate probability at the cap. A quota share changes the loss scale, and a nonlinear excess layer can alter the shape of the observed payment distribution. If losses are reported only when a layer attaches, the resulting sample is selected through the contract. Treating it as an unbiased draw from original gross event losses is an inferential error.

There are also two forms of economic uncertainty in reinsurance analysis. The first is the uncertainty of the catastrophe event and the losses generated by it. The second is the ability and timing of the reinsurance counterparty to pay, including disputes about event definition or aggregate recovery. The latter is not resolved by estimating a physical catastrophe return level. Capital governance must distinguish risk-transfer promises from confirmed recoveries and operational cash flows.

The tail model can support a transparent decision comparison across retentions. For each candidate attachment, one can evaluate expected layer loss, annual probability of at least one attachment, annual probability of exhaustion, expected annual aggregate recoveries under a specified frequency model, capital response, and uncertainty bounds. The same external event distribution should be used when comparing treaty alternatives unless changes in underwriting or behavioral response are explicitly part of the scenario. This makes the decision traceable and isolates the cost of tail uncertainty from differences in contract structure.

Capital requirements and governance

Regulatory capital is a defined legal and supervisory object. In the European Union, the Solvency II Directive specifies in Article 101 a calibration of the Solvency Capital Requirement to a one-year 99.5% value-at-risk measure of basic own funds, subject to the directive's scope and mechanics. Its capital calibration provisions are different from the colloquial idea of a catastrophe occurring “once in 200 years.” The 99.5th-percentile one-year economic loss of an insurer is a balance-sheet and aggregation question, involving underwriting, market risk, credit and counterparty risk, and risk mitigation. A 200-year level of one peril's annual maximum event loss cannot be substituted without the missing aggregation and valuation transformations.

The directive describes a standard formula and internal-model conditions within its framework; the present study does not claim that Solvency II rules are globally binding or that the same calibration applies unchanged in every jurisdiction. International insurers operate under distinct capital, accounting, reinsurance, and claims-law regimes. EVT is a statistical method that may be used within capital modeling or independent validation. Its numerical output becomes a regulatory input only after the insurer's relevant loss object, valuation horizon, rules, diversification, and model governance have been established.

A model-governance record should document the data population; event, claim, and exposure definitions; inflation and normalization adjustments; threshold or block choice; estimation method; dependence assumptions; diagnostics; parameter and model uncertainty; validation results; external datasets and their revisions; scenario sensitivities; and the final management use. For large-loss stress testing, the record should include plausible scenarios beyond the fitted mathematical family, especially where historical coverage is sparse. The full range of uncertainty deserves the same attention as the reported return level, because an extreme point estimate can be highly sensitive to a small set of observations or decisions.

Integrated analytical synthesis and operational research priorities

The mathematical and insurance literatures together support six connected analytical judgments. First, the extreme-value target must correspond to the actual decision variable. A claim-level quantile is informative for a claim excess treaty; a catastrophe occurrence distribution is relevant to occurrence reinsurance; an annual aggregate distribution is necessary for annual aggregate capital and treaty decisions. Selecting the target after fitting a convenient model weakens interpretation. Second, the hazard, exposed assets, insurance contract, and measured loss must be separately specified. This clarifies whether the research concerns physical extremes, financial loss formation, or the tail of realized insurer payments.

Third, a fitted tail has two independent components: the exceedance frequency or base probability and the magnitude conditional on exceedance. POT modeling requires both, and translating “per claim” tail probabilities into “per year” return levels requires a frequency model. Fourth, distant tail extrapolation derives much of its sensitivity from shape, threshold, and stationarity assumptions. More numerical digits cannot resolve this sensitivity; cross-threshold tests, profile intervals, alternative model classes, and independently justified stress scenarios can make it visible.

Fifth, temporal and spatial dependence must be treated at the unit of risk used by the insurer. One storm's several property claims are not independent storm occurrences. One physical catastrophe can involve several perils and contracts. The statistical units in the data and the contract units in the economic question have to be reconciled. Sixth, decisions under uncertainty should communicate both conditional estimates and the conditions that could invalidate their use. The strongest EVT deliverable is an auditable mapping from observation, assumption, mathematical result, and decision context to a bounded conclusion.

These judgments are an analytical synthesis of established research and the insurance-specific transformations developed in this monograph. They are not presented as a new theory, formal doctrine, or original named model. Their principal contribution is operational clarity: rare-loss estimates acquire financial meaning through a reproducible sequence of definitions and transformations. This statement follows from the classical limit theory, empirical threshold-inference literature, and direct insurance use-case analysis rather than from a claim of theoretical priority.

Several research priorities arise from the remaining evidence gaps. Comparative studies should evaluate how EVT tail estimates change when identical historical events are expressed as physical intensity, normalized gross damage, policy-indemnified event loss, and post-reinsurance net loss. Long-term records need careful reconciliation with exposure growth and changing claims valuation. Threshold uncertainty should be assessed jointly with the event-definition problem, rather than only through GPD parameter confidence intervals. Compound catastrophe models need reliable diagnostics for clusters, missing weak events, and repeated exposures. Climate-sensitive portfolios require scenarios that specify both the evolving hazard distribution and the future insured asset base. Finally, validation should assess decision stability: whether plausible choices of thresholds, models, and data vintages change treaty attachments, exhaustion probabilities, or available capital materially.

The appropriate result of these priorities is better-identified estimation and clearer audit evidence. They remain proposed research directions; no field experiment, new insurance database, industry calibration, or independently verified scientific novelty is claimed here.

Conclusion

Extreme value theory gives the mathematics of far-tail inference a coherent foundation. For independent identically distributed block observations, suitably normalized maxima can converge to a generalized extreme value law. For distributions in appropriate domains of attraction, exceedances over a sufficiently high threshold can be approximated by a generalized Pareto law. These results explain why large-loss extrapolation can be conducted with parsimonious tail parameters and why a modest change in the estimated shape can transform the largest inferred claims. The inference remains conditional upon approximation quality, observation design, and the stability of the process being modeled.

For insurance, the decisive methodological fact is the identity of the loss variable. Gross claim severity, reported and paid claim severity, catastrophe occurrence loss, annual maximum, annual aggregate, and net reinsurance retention are related through count processes, dependence, exposure, vulnerability, and contracts. They are different inferential objects. The synthetic numerical case demonstrates this distinction concretely: its individual-claim percentiles, expected shortfall, and 100-year annual maximum have valid meanings only under the separate assumptions used to derive each. Converting one of them into another requires a frequency model and a defined financial transformation.

The tail-shape parameter has high decision relevance because extreme quantiles and layer losses can be sensitive to it. Threshold choice, limited exceedance counts, data censoring, clustering, model class, and structural change consequently belong inside the final uncertainty assessment. A classical limiting theorem does not substitute for empirical diagnostics, and historical insurance payouts do not by themselves identify physical climate trends. Scientific adequacy grows from explicit source–claim matching, mathematical derivation, actuarial data construction, and independently reproducible tests.

In catastrophe and solvency applications, the strongest practice is to link each reported return level to the event clock, exposure base, contract terms, and valuation horizon for which it is meaningful. Expected loss, attachment and exhaustion probability, aggregate annual risk, and capital adequacy then become comparable decision measures rather than interchangeable uses of an impressive extreme number. The research concludes that EVT is most informative when it is embedded in a transparent insurance-risk measurement process whose assumptions and uncertainty remain inspectable. This is an analytical finding grounded in prior theory and research, with no assertion of a newly discovered EVT law or authorial conceptual system.

References