RESEARCH DOSSIER
Property Insurance Large-Loss Threshold and Tail Modeling
Abstract. Large property-insurance losses occupy an intersection of physical damage, contract-defined indemnity, administrative observation, and statistical extrapolation. The threshold that causes an insurer to report a claim, the deductible at which indemnity begins, the level above which a peaks-over-threshold approximation becomes credible, and the layer at which reinsurance attaches are distinct quantities. This monograph establishes their respective mathematical roles and investigates the consequences of confusing them. It examines the observational construction of a property-loss population; left truncation, interval and right censoring; threshold selection; generalized Pareto and competing severity distributions; event dependence and exposure normalization; tail-index uncertainty; capital estimation; and the limits of extrapolating a short loss record into very rare events. Classical extreme-value results are read alongside actuarial large-claim research, public supervisory standards, and the practical division between hazard, economic damage, insured loss, and final payment. An explicitly hypothetical numerical study demonstrates how increasing the tail shape from 0.15 to 0.35 changes the 99.9th claim-severity quantile while leaving the exceedance threshold and scale fixed. The main analytical conclusion is that threshold selection is part of defining the estimand and the observable insurance portfolio: even a mathematically well-fitted tail can provide a misleading capital measure when claim inclusion, policy limits, loss development, event clustering, and changing exposure are handled inconsistently. The proposed research workflow is an application and synthesis of established methods, not the announcement of a new theory.
Keywords: Property insurance; large losses; extreme value theory; peaks over threshold; generalized Pareto distribution; left truncation; censoring; deductibles; policy limits; tail index; exposure normalization; catastrophe dependence; parameter uncertainty; expected shortfall; solvency capital; model validation.
Introduction
Property-insurance loss severity is a sequence of transformations, rather than a single amount found ready-made in an accounting ledger. A peril affects an asset; an engineer or adjuster estimates physical damage and associated economic consequences; the insured interest, valuation clause, exclusions, sublimits, deductible, and policy maximum determine a contractual obligation; claims administration records paid and outstanding amounts at particular evaluation dates. The same occurrence may create multiple claims and multiple payments, while one claim may be developed, reopened, subrogated, or reported late. A statistical model that begins at the paid-claim column and calls the result the distribution of physical property damage has changed the scientific object without declaring the change.
A second transformation is statistical. Ordinary insurance samples contain relatively many small claims and comparatively few extreme losses. Their upper tails require extrapolation beyond observed quantiles. Classical extremal-event theory for insurance establishes asymptotic approximations under specific distributional and dependence conditions; the applied question is whether a particular property portfolio meets those conditions at a credible threshold. Work on threshold selection and uncertainty shows that raising a threshold generally trades approximation bias for estimation variance. It does not turn a finite catalogue of very large claims into directly observed hundred-year experience.
The research object of this dossier is the distribution of large property-insurance claim severity and the choice of inferential thresholds used to estimate its tail. It covers insured residential, commercial, industrial, and catastrophe-affected property, including claim-level and event-level aggregation. Its focus remains distinct from the full economics of insurance, property contract doctrine, general catastrophe engineering, portfolio reserving, and stand-alone reinsurance design, which supply necessary boundaries but each have broader scientific objects.
Three linked questions organize the analysis. How do contractual and administrative thresholds alter what enters a loss database? Under what assumptions can generalized Pareto or alternative distributions represent the observable upper tail? How should threshold, dependence, model, and parameter uncertainty enter risk-capital decisions across different property portfolios and jurisdictions? The evidence base combines foundational statistical results, a documented actuarial large-claim application, primary supervisory sources, and transparent mathematical derivations. Public references establish methods and institutional context; no proprietary insurer loss file has been analyzed for this monograph.
Definitions, Historical Context, and Primary Source Foundations
The loss object and its successive monetary measures
Let H denote the realized hazard event, A the affected physical assets, V their replacement or insured values, and D_phys the physical damage state. An engineering loss function converts these into repair or replacement cost C_phys, potentially augmented by costs such as debris removal. Economic damage E can include interruptions and uninsured consequences. The contractual covered loss X depends on the policy's insuring agreement, causation provisions, insured interest, exclusions and valuation basis. Gross insurer liability before reinsurance is Y=g(X;P), with P the complete policy wording and applicable interpretation. Payment at valuation date t, denoted Y_paid(t), can differ from ultimate liability because of settlement, salvage, subrogation, reinstatement, or late development. These measures have related but unequal distributions.
The difference matters to public comparisons. The EIOPA protection-gap dashboard separately draws on economic losses, insured losses, and coverage data across European countries. Its historical panel is an account of reported catastrophes, while its current view incorporates risk-model estimates. An insured-claims tail calculated from one company's policy database therefore cannot be assumed to represent the tail of total economic catastrophe damage. The EIOPA catastrophe data hub likewise distinguishes insured building exposure from incurred losses and explicitly says that its participating-company sample is not the entire market.
A property claim must also have a unit of analysis. A building-level loss, policy-level claim, insurance group event, occurrence-related set of claims, and annual aggregate each answer a different question. Ten buildings damaged by one storm may generate ten claim severities but only one common-cause event. Treating the ten as independent observations in a portfolio return-period calculation can make rare-event uncertainty look much smaller than it is.
Historical and disciplinary foundations
Asymptotic probability provided a mathematical language for maxima and excesses before it was adapted to insurance. Balkema and de Haan (1974) developed limit results for residual life at high levels. Pickands (1975) studied inference from extreme order statistics, and Hill (1975) provided a high-order-statistic tail estimator for power-law-type upper tails. Davison and Smith (1990) made exceedance modeling and diagnostics practical through likelihood, regression and dependence analysis. The results concern classes of stochastic distributions; property-insurance applications additionally require observed contracts, exposures, peril mechanisms, and a correct accounting of which losses are visible.
Two large syntheses clarify the breadth of the available mathematics. Embrechts, Klüppelberg, and Mikosch (1997) connect maxima, sums, upper order statistics, stochastic processes, and insurance loss applications in Modelling Extremal Events. McNeil, Frey, and Embrechts (2015) integrate risk measures, dependence and aggregation in Quantitative Risk Management. In a distinct empirical setting, Cebrián, Denuit, and Lambert (2003) apply generalized Pareto methods to large medical insurance claims. The latter study demonstrates an actuarial use case and does not itself provide evidence that a specific property peril has the same tail parameters.
A central distinction is between an asymptotic theorem, the adequacy of a finite-sample approximation, and the inferential decision based on that approximation. A theorem says what a suitably rescaled distribution approaches under its assumptions. It does not choose a threshold for a real insurer. An estimation procedure can deliver a converged parameter and conventional confidence interval even when the record was truncated at an undocumented reporting level. A capital decision adds obligations, risk appetite, legal accounting and the interaction of several stochastic hazards.
Definitions that remain distinct throughout the dossier
A reporting threshold r is a monetary or administrative level below which a record may be absent. A contractual deductible d modifies the insurer's payment. A statistical threshold u is a level beyond which an upper-tail approximation is fitted. A reinsurance attachment a determines where a specified reinsurance layer begins to respond. A risk tolerance or supervisory threshold k is a criterion for a capital or management action. These values can coincide numerically, but numerical coincidence does not give them identical meanings.
The upper-tail survival function of a nonnegative gross severity X is S_X(x)=Pr(X>x). A heavy right tail can mean several things in different disciplines; in strict regular-variation language S_X(x)=x^{-alpha} L(x), where L varies slowly and alpha>0. A long-tailed or subexponential law has other precise mathematical definitions. Lognormal severity can be subexponential without having a regularly varying Pareto tail. This distinction matters when a sample supports extreme skewness but not a confidently estimated positive generalized Pareto shape.
The target estimand must be written before fitting: examples are the gross ultimate claim quantile for a given property class, the net-of-contract claim quantile, the annual occurrence exceedance probability, the annual aggregate loss percentile, or the conditional mean of loss beyond a high quantile. Each target has its own observation law and potential model validation.
Source hierarchy and jurisdiction
The foundational journal articles and research monographs support theoretical and methodological statements. Published datasets or authority pages document actual coverage of their own statistical samples. The IAIS Insurance Core Principles are internationally used supervisory standards, while implementation and binding obligations depend on national law. In the European Union, Solvency II Directive 2009/138/EC, Article 101, specifies a one-year 99.5% value-at-risk calibration for the Solvency Capital Requirement in terms of basic own funds. That legal calibration is not equivalent to a 99.5% quantile of a single property's ground-up loss. Comparisons to other regulatory systems must make explicit the relevant capital object, aggregation period and jurisdiction.
Threshold and Censoring Selection
Physical, administrative, and statistical selection processes
Consider a property portfolio in which every insured location has a physical loss X, potentially zero. It may be possible to identify every policy-year exposure, yet only claims above a reporting condition r appear in the severity file. If losses below r never enter the dataset, observations are left-truncated: the observed density is f(x | X>r)=f(x)/S_X(r) for x>r, conditional on the reporting process being a deterministic and consistently applied monetary rule. The recorded count then represents exceedances, not all physical-loss events. Fitting the unrestricted density f(x) to this truncated set without the factor S_X(r) makes the severity parameter describe a different data-generating process.
When every incident is logged but the exact loss is known only to lie below a recording threshold, the observation is left-censored rather than truncated. Its likelihood contribution can be F(r), conditional on the relevant observation scheme. When the record states only that a loss exceeds a terminal amount c, perhaps because the insurer's maximum payment has been reached, it supplies a right-censored contribution S(c). An interval estimate [l,h] supplies F(h)-F(l). It is the information retained about the underlying quantity, not simply the magnitude of the threshold, that determines the proper likelihood.
Property systems frequently mix all four situations. A small claim may be absent entirely; another may be logged with zero paid because of the deductible; an open large loss may have an estimated range; a policy-limit loss may reveal only that the ground-up severity exceeded the cap. The loss file therefore needs a data dictionary recording exact, interval, censored, and truncated observations separately. The principles of likelihood-based exceedance modeling can accommodate incomplete observations when their mechanism is correctly specified, but an analyst cannot reconstruct an unobserved exposure denominator from severity records alone.
Contract transformations and their implications
For a simplified occurrence deductible d and payment cap l, define the insurer payment as Y=min{max(X-d,0),l}. This is a mathematical illustration, not a universal interpretation of property-policy wording. Percentage deductibles, franchise clauses, replacement-cost conditions, sublimits, valuation rules, aggregate deductibles and reinstatement provisions produce other mappings. Even this simple mapping has a mass at zero, a continuous section between zero and l, and a mass at l when the ground-up loss may exceed d+l. A continuous generalized Pareto likelihood fitted to Y as though it had no atoms is misspecified.
For exceedance analysis the safest starting point is a specified gross loss X, with the contract map applied separately. When the available data consist solely of capped insurer payments, ground-up upper-tail identification may be impossible without external uncapped losses, policy metadata, or credible parameter assumptions. Two different gross distributions can generate identical observed capped payments while giving very different probabilities for a hypothetical higher layer. That is a genuine identification problem, not merely a problem solved by a larger optimizer.
Adjustments must preserve valuation date and currency. A USD 1 million loss in one construction-cost environment and a USD 1 million loss several years later are not automatically comparable events. Appropriate exposure normalization may involve insured value, replacement-cost indices, coverage share, insured unit count, peril class, and the geographical distribution of insured assets. Normalizing X by insured value can reveal damage ratios, but ratios are naturally constrained by the nature of the asset and contract, whereas absolute reconstruction costs can change with inflation and supply shocks. Analysts should examine both a severity scale and the economic interpretation of the resulting target.
Selection of the tail-model threshold
Let u>=r be a candidate statistical threshold where individual exceedances X-u are available. A lower u produces more observations but may include a body distribution inconsistent with the generalized Pareto approximation. A higher u leaves fewer observations and makes extrapolation unstable. Scarrott and MacDonald (2012) review this bias–variance threshold tradeoff. Wadsworth and Tawn (2012) develop threshold diagnostics and uncertainty, emphasizing that treating a subjectively chosen threshold as fixed can understate inferential uncertainty.
A defensible selection process begins with policy and data-audit thresholds, then considers a grid of feasible statistical values. At each candidate u the analyst records the number of independent events or claims retained, the threshold exceedance rate, fitted tail shape and scale, confidence intervals, diagnostics, and out-of-sample performance. A mean residual life plot can be informative when moments exist; a stable-looking plot is not proof of an asymptotic regime. Parameter-stability plots require the correct change of scale: for an exact generalized Pareto tail with shape xi, raising u to v changes its scale according to beta(v)=beta(u)+xi(v-u), where the support condition is met. Comparing beta values without this adjustment creates spurious evidence of instability.
The threshold should not be chosen solely because it produces a desired solvency number. It should also not be picked by maximizing in-sample goodness of fit over a large unchecked grid and then treating the winning estimate as if no search occurred. Threshold sensitivity belongs in the final tail capital range. Time-blocked or event-blocked validation is preferable when claims are clustered; randomly splitting claims from the same flood into fitting and test sets leaks event information across samples.
Frequency calibration and reporting thresholds
For an observation window with exposure e, let N_u denote the number of losses above u and lambda_u=E[N_u]/e their average exceedance rate under a stationary count model. In a basic Poisson implementation Pr(N_u=n)=exp(-e lambda_u)(e lambda_u)^n/n!. Frequency calibration requires policy exposure records or audited aggregation; the number of rows in a claims spreadsheet is not a risk exposure denominator.
If a claim-severity sample comprises n fully observed claims and n_u exceed u, the empirical exceedance fraction n_u/n may approximate Pr(X>u) for the specified population of reported claims. It cannot automatically represent the probability that an insured building suffers a large loss in a policy year. The latter combines an occurrence process, exposure, zero losses and an ultimate-loss definition. A different claim submission practice can move n without any change in physical peril. These two denominators must never be silently interchanged.
Reporting can be stochastic and severity-dependent. Very large events may cause delayed notification, consolidated claims, litigation, or incomplete post-disaster inspection. Then Pr(recorded | X=x, z) depends on loss size and covariates z. A fixed threshold correction cannot restore representativeness; the data mechanism itself requires a model, auxiliary sample, or a credible sensitivity bound.
Threshold selection as an estimand decision
The most consequential editorial observation is that thresholds determine the population to which results refer. Statistical u defines conditional excesses; reporting r defines the observed records; deductible d defines a contractual transformation; event definition defines which losses are aggregated. A claim model estimated after all these operations can be internally consistent and still answer the wrong financial question. Accurate documentation of those operations is therefore a precondition of scientific replication.
Heavy-Tailed Loss Distributions
Generalized Pareto excess law
For sufficiently high u under an appropriate extreme-value domain of attraction, the conditional excess Z=X-u | X>u may be approximated by a generalized Pareto distribution (GPD). With scale beta>0 and shape xi, its survival function is Pr(Z>z)=[1+xi z/beta]^(-1/xi) on 1+xi z/beta>0, and its xi=0 limit is exp(-z/beta). A positive xi represents a Pareto-type heavy upper tail; xi=0 has exponential-type decay in the excess approximation; negative xi yields a finite upper endpoint in the fitted model. The foundational justification derives from Balkema and de Haan (1974) and Pickands (1975), with likelihood-based applied development in Davison and Smith (1990).
If p_u=Pr(X>u), the unconditional survival above u is approximated by S_X(x)=p_u[1+xi(x-u)/beta]^(-1/xi). Consequently the severity quantile at probability a, for 1-a<p_u, is q_a=u+(beta/xi){[p_u/(1-a)]^xi-1}, with logarithmic limit q_a=u+beta log[p_u/(1-a)] when xi=0. This expression depends on both the conditional excess tail and a defensible estimate of the unconditional exceedance probability. It does not work when p_u has been misestimated from a truncated claim sample or where a heterogeneous collection has no stable observation base.
For xi<1 the conditional mean excess is beta/(1-xi), with the support and moment conditions observed. For xi>=1 the theoretical GPD mean is infinite; for xi>=1/2 its variance is infinite. These are statements about a fitted unbounded model, not an assertion that actual insured property payments can exceed physical limits without bound. Real portfolios contain finite exposure and contract limits, though those ceilings can be uncertain and the fitted model may approximate behavior over a practically relevant finite range. A model shape that implies infinite mean deserves diagnostic attention, especially in finite-exposure contexts.
Hill estimation, tail index and selection of k
For iid positive observations with a regularly varying upper tail, let X_(1)<=...<=X_(n) be the order statistics. Hill's estimator of gamma=1/alpha, based on the largest k observations, is H_k=(1/k) sum from j=1 to k of log[X_(n-j+1)/X_(n-k)]. It directly uses ratios of large order statistics. Hill (1975) develops tail inference from top observations. Its appropriateness requires a Pareto-type tail and a suitable k, rather than merely a visibly skewed sample.
A plot of H_k over k can reveal sensitivity, but a flat-looking region need not identify the true asymptotic range. Changes in exposure class, currency normalization, records at policy caps, or clustered catastrophe losses can each generate visual regimes with different interpretations. For an insurance analyst, k is effectively another threshold choice. Hill and GPD estimators should not be treated as statistically independent confirmations when both use the same largest observations.
Alternative severity structures
A lognormal, gamma, Weibull, Burr, log-logistic, spliced body–tail model, generalized beta, or mixtures of peril-specific distributions can fit portions of a property sample. They imply materially different extrapolation outside the observed range. The choice between them should arise from physical and underwriting heterogeneity, support constraints, external validation and quantile stability, rather than the lowest in-sample information criterion alone. Standard statistical analysis of extremes provides a framework for comparing parametric assumptions with extreme-value approximations.
A spliced model uses a body distribution below u and a tail above u, with probability mass matched at the junction. Continuity of the distribution function and, where required by an application, density continuity should be checked rather than assumed. If the body was observed only through reported claims, splicing must preserve that reporting conditionality. An apparent excess probability computed from all policy-years cannot simply be joined to a body fitted to only paid claims.
A single unbounded GPD is often a poor representation of capped indemnity. One approach models ground-up X and applies payment clauses in simulation. Another directly models Y, including its point masses and layer support. The first is generally more informative for changing contract design; the second may be sufficient for forecasting an unchanged book if its payout definition, historical limits, and exposure stay stable. The difference reflects the purpose of the model.
What makes property losses appear heavy-tailed
The observed severity tail can be shaped by sums of building characteristics, geographical accumulation, industrial concentration, partial or total destruction, and co-occurring insured interests. A mixture of ordinary perils and rare catastrophe perils can be heavier-tailed than each finite-range component over a particular interval. That empirical appearance does not establish regular variation at asymptotically infinite monetary values. The physical processes and asset stock impose changing bounds, and rebuilding costs may be time-varying.
Catastrophe-related dependence creates a second distinction. A very large event loss can arise from many moderate component claims. The distribution of the event sum depends on the number of claims, their common hazard intensity and dependence structure. It is possible for individual claim severities to be moderately tailed while the event aggregate is extremely skewed because of correlated claim counts and broad-area damage. The catastrophe data resources of EIOPA's building-exposure hub reinforce that exposures and incurred losses are measured at different observational levels.
Covariates may improve interpretation. Construction type, occupancy, property value, flood depth, wind speed, wildfire interface, local labor prices, geographic hazard and contractual breadth can enter severity and event-frequency models. Chavez-Demoulin, Embrechts, and Hofert (2016) demonstrate covariate-dependent frequency and severity extremes in operational-risk data. Their method illustrates a general modeling possibility, while its empirical coefficients cannot be transferred to property insurance.
A fully hypothetical severity calculation
Assume for illustration that X is a claim-level gross ultimate property-loss severity measured in millions of constant-value U.S. dollars, with statistical threshold u=1.0, exceedance probability p_u=0.020, generalized Pareto scale beta=0.60 and shape xi=0.25. These are invented parameters for mathematical demonstration; they are not fitted insurer experience. The fitted 99th severity percentile is q_0.99=1.4541 million, the 99.5th is 1.9941 million, the 99.9th is 3.6754 million, and the 99.99th is 7.6254 million. The conditional mean of X given X>1.0 is 1.8000 million. The expected shortfall beyond q_0.999, calculated as q_0.999+[beta+xi(q_0.999-u)]/(1-xi), is approximately 5.3672 million.
Sensitivity to tail shape is substantial even with p_u, beta and u held fixed. For the same 99.9th severity percentile, xi=0 (the exponential excess limit) yields 2.7974 million; xi=0.15 yields 3.2692 million; xi=0.25 yields 3.6754 million; xi=0.35 yields 4.1772 million; and xi=0.50 yields 5.1666 million. These values are direct substitutions into the quantile identity, not model validation results or forecasts. They demonstrate the cost of over-interpreting a point estimate for the tail index.
The example does not state a policy-year probability. Even if the source sample contained 200 exceedances among 10,000 independently and fully observed claims, the ratio 0.02 would condition on observed claims. Translating it to annual insurer losses requires independent incidence and exposure calibration. If the insurer's policy limit is 3.0 million for each claim, the example's ground-up 99.9th percentile cannot be read as the 99.9th percentile of its capped payment distribution.
Finite support and tail extrapolation limits
A fitted negative xi indicates a finite endpoint u-beta/xi. A fitted positive xi does not compel an unbounded real-world loss. Property replacement costs can grow; a policy value can change; exposed structures are countable; a claim can have contingent costs. Whether a physical or contractual maximum is knowable depends on the defined loss measure and valuation date. Limits can make a net insurer-obligation distribution bounded even where a hypothetical uncapped gross damage model retains an unbounded approximation.
Extrapolation beyond the maximum observed value is a model-based estimate. The greater the distance, the more assumptions about tail stability, exposure, inflation, jurisdiction, hazard and contract design determine the answer. The 2012 threshold review and likelihood uncertainty analysis support reporting that inferential uncertainty as a central result rather than hiding it beneath a single large quantile.
Uncertainty and Portfolio Capital
From single claims to annual aggregate
An insurer cares about cash and capital over a contractual and regulatory horizon. Define annual loss S=sum from i=1 to N of Y_i where N is a claim count and Y_i policy-adjusted ultimate liability. A basic collective-risk model takes N as Poisson and, conditionally on N, assumes independent identically distributed claims; it is useful as a baseline mathematical construction. Property catastrophe portfolios generally require more structure. N can be clustered by occurrence, common weather, geographic footprint, macroeconomic construction inflation and policy exposure. Catastrophe subevents, aggregate deductibles, reinstatements and program-specific terms can cause event and annual loss distributions to diverge.
There are at least three distinct exceedance measures: claim severity exceedance (the chance a claim is above x), occurrence exceedance probability (OEP, the chance the largest event loss in a year exceeds x), and aggregate exceedance probability (AEP, the chance all annual covered losses together exceed x). The model's definition should specify whether OEP concerns ground-up event loss, gross insurer event liability, or net liability after reinsurance. A return period is conventionally the reciprocal of a stationary annual exceedance probability; a 100-year return level has 1% annual exceedance probability under that model. It does not mean a loss occurs exactly once per century, nor that events are separated by one hundred years.
Consider a separate hypothetical event portfolio with a Poisson mean lambda_u=1.5 events per year whose event loss exceeds u=1 million, and conditional GPD parameters beta=0.6 million, xi=0.25. With independence between exceedance events, Pr(M_year<=z)=exp[-lambda_u{1+xi(z-u)/beta}^(-1/xi)] for z>=u. Solving Pr(M_year>z)=0.01 gives the 100-year OEP level z=6.9886 million. This event example is separate from the preceding claim-level p_u=0.02 calculation; the two rates have different denominators. Changing event frequency or spatial dependence changes the return level even if the severity parameters remain the same.
Contract structure and reinsurance
A capital model must implement deductibles, limits and insurance-to-value provisions before estimating the relevant insurer-loss distribution. Quota share and excess-of-loss reinsurance add layers, limits, reinstatement conditions and counterparty exposure. For a simplified per-event excess-of-loss treaty with attachment a and layer limit L_R, the nominal recovery is R=min{max(G-a,0),L_R} on an eligible gross event loss G; actual recovery depends on treaty wording, aggregate limits, reinstatements, basis risk and recoverability. Net event loss is G-R under the simplified fully collectible treaty, not a mere constant fraction of ground-up damage.
Aggregation creates competing effects. Diversification across independent perils may reduce portfolio quantiles relative to simple summation of standalone capital. Conversely, shared exposure, property-value inflation, storm sequences or a large wildfire affecting the same region can increase dependency beyond that assumed in a simplistic independent model. Lower-frequency but geographically concentrated portfolios may have greater capital sensitivity to a particular tail shape than broadly diversified books. The arithmetic of risk aggregation and dependence supplies the general statistical context, but portfolio dependence parameters require their own evidence.
Regulatory capital is a broader object
The Solvency II Directive defines the European Solvency Capital Requirement in relation to the distribution of the insurer's basic own funds at 99.5% over one year, encompassing multiple risks. A model's 99.5th annual property-loss percentile is one possible risk input; it is not by definition the legally required company-wide capital. The IAIS risk-management framework calls for insurer-specific integrated identification and management of material risks and includes stress and scenario analysis in its enterprise-risk-management guidance. Jurisdictions differ in capital models, accounting treatment, admissible risk mitigation and supervisory processes.
Risk measures serve different purposes. For a loss L, VaR_a(L)=inf{x:Pr(L<=x)>=a}. Expected shortfall can be defined as the average loss across the worst 1-a fraction of the distribution with appropriate handling of atoms; for a continuous loss it equals E[L|L>VaR_a(L)]. A high VaR summarizes one threshold, while expected shortfall responds to the severity of outcomes beyond it. The distinction between a single loss threshold and the average of severe outcomes makes dependence and tail averaging economically salient, but neither risk measure can compensate for an inaccurate observational definition of insured loss.
Uncertainty: parameter, process, model and change
Process uncertainty is variability in future outcomes conditional on the selected model and parameters. Parameter uncertainty is uncertainty about fitted frequency, scale, shape, dependence and covariate coefficients. Model uncertainty concerns choices such as GPD versus a tempered tail, treatment of a policy cap, event-definition rules or the missing-record mechanism. Structural change concerns physical peril, exposure, policy design, claims practice and economic value changing across time. These sources cannot be pooled into a single confidence interval simply by using more simulation draws from one calibrated parameter vector.
A parametric bootstrap estimates the sampling distribution of an estimator under its assumed family. An event-block bootstrap preserves clustering better than a claim-row bootstrap when common catastrophes produce dependent claims. A profile likelihood can reveal asymmetry in tail-shape or quantile uncertainty. Bayesian analysis can integrate parameter uncertainty through a posterior predictive distribution, but results depend on prior choice and model specification, especially with sparse extreme observations. Wadsworth and Tawn (2012) emphasize the need to incorporate uncertainty in threshold choice rather than conditioning silently on a convenient u.
For a particular annual portfolio estimate, the preferred communication is a set of clearly labeled distributions or scenario ranges: baseline fitted loss; parameter uncertainty at a fixed model; threshold alternatives; contract and exposure sensitivity; plausible dependency structures; and explicitly motivated stress scenarios. This separation permits a decision-maker to identify which uncertainties can be reduced by data collection and which require conservative governance or scenario planning.
Capital, liquidity and time
The time between peril occurrence and settlement matters. A catastrophe may trigger liquidity needs before ultimate liability is established; repairs can continue while market prices change. Portfolio capital and immediately available liquid assets solve related but different problems. Reinsurance may lower ultimate net loss while introducing collection timing or counterparty risk. An engineering damage scenario can inform claims surge, fraud review and operational capacity without supplying a calibrated 1-in-200-year capital estimate.
A further distinction concerns nonstationarity. If the building stock, construction costs, insurance penetration or hazard climate changes, a probability calculated under yesterday's exposure base is not automatically a forecast for tomorrow. The EIOPA economic and insured loss views illustrate why exposure and coverage need to accompany reported historical losses. A stable generalized Pareto shape estimated from an old pooled dataset is only one assumption about the changing insurance system.
Empirical Verification, Uncertainty, and Competing Explanations
Evidence needed for a real property portfolio
A reproducible empirical study requires a defined insurance book, effective policy dates, written exposure, construction and occupancy data, territory and peril information, insured values, deductibles, limits, currency and valuation adjustments, claims opened and closed, inception and occurrence dates, payment and case-reserve histories, salvage and subrogation, reinsurance recoveries, and unique catastrophe-event identifiers. Loss records need to distinguish paid, incurred and ultimate projections. A claims estimate at six months of development is not interchangeable with a fully developed loss from a mature accident year.
For each row, analysts should retain the observation mechanism: whether small claims were logged, whether zero payment reflects no covered liability, whether a positive payment was capped, whether a limit is shared with other claims, and whether losses are aggregated at building, policy, location or event level. Missingness is part of the study design. A convenience sample of paid large claims from a broker or one reinsurer can be scientifically useful for a specified conditional distribution, but it cannot be assumed to represent nationwide property damage or the unconditional insured population.
When jurisdictional comparisons are attempted, required coverage and loss-reporting standards must be documented locally. The European EIOPA sample in the catastrophe data hub includes particular participating insurers and warns that it does not represent the full market. It follows that publicly available insured-loss data may be appropriate for illustrating coverage patterns or evaluating particular event statistics but insufficient for a universal claim-tail parameter or jurisdiction-wide insured-exposure denominator.
An executable research protocol
The first research stage fixes the unit of loss and target measure. It produces a signed data dictionary, policy-year exposures, a time window, known reporting thresholds, and an inventory of censored and truncated observations. The second stage examines basic distributions, class and peril mixtures, inflation-adjusted values, and event clustering. The third fits a body model where scientifically needed and an upper-tail model across a defensible threshold grid. The fourth evaluates tail shape and scale, candidate alternative families, policy transformation consistency and event dependence. The fifth performs predictive checks and sensitivity analysis. The sixth maps the selected probability outputs into clearly defined capital, reinsurance or underwriting questions.
The workflow follows existing statistical practice rather than establishing a new proprietary algorithm. It is strengthened by preserving a record of analytical choices, including unsuccessful models. A threshold rejected after a poor fit, a failed source-integrity test, or an unstable tail shape is part of the scientific history, because silently removing it prevents readers from judging selective reporting.
Diagnostics and temporal validation
Goodness-of-fit plots for transformed exceedances, return-level plots, probability plots and residual checks are useful diagnostics. Shape stability, likelihood profiling and sample-size curves are essential. Yet in-sample diagnostics assess compatibility with the observed tail, not predictive accuracy at a quantile beyond every recorded loss. To evaluate forecast claims within an observable range, use out-of-time or event-blocked folds, with the model trained only on earlier available information where the study is temporal.
An out-of-time design must also handle loss development: the latest year often has less complete claims. One option is to evaluate a fixed development maturity across accident years; another is to model ultimate losses explicitly with uncertainty. Comparing mature historical settlements to immature recent incurred estimates can create artificial trends in the upper tail. Data pipelines should record the evaluation date and transformations applied to each observation.
For dependent extremes, estimate cluster or event-level properties and use simulated year aggregates or stochastic event catalogs. An independent-claims model can be retained as a transparent baseline, provided the difference from event-dependent outcomes is reported. The inferential limits of dependent and seasonal exceedances appear in Davison and Smith's analysis.
Competing explanations for apparent tail change
An increase in large paid losses can result from more severe physical hazards, higher property values, higher repair-price inflation, broader insurance coverage, higher policy limits, different claims settlement behavior, more extensive exposure or combinations of these. A purely statistical trend in nominal paid amounts does not identify which mechanism produced it. Peril-level external hazard evidence and exposure-normalized property records are needed before a causal interpretation is justified.
An apparently finite tail can reflect contractual caps; a lognormal-looking tail may be a mixture of several property classes; a dramatic increase at one reporting amount can reflect claims handling or data extraction. An excess cluster around a single hurricane may inflate confidence in a tail shape if building claims are treated as separate independent events. A stable shape after pooling currencies can be an artefact of conversion or changing insured values. These alternatives are testable with policy, exposure, event and valuation metadata, not simply a different optimizer.
Comparisons of tail fits should therefore report the estimand, exact monetary basis, threshold, number of exceedances and effective number of independent clusters, estimation procedure, diagnostics, uncertainty method, temporal holdout and all major contract transformations. Sensitivity analyses should vary physical exposure assumptions alongside statistical tail models.
Reproducible mathematical verification
For the hypothetical GPD in Chapter 3, the unconditional survival relation and quantile formula are algebraic rearrangements of the stated excess survival. The reported quantile and expected-shortfall numbers can be reproduced directly with the parameters listed. The 100-year annual maximum example follows from the Poisson probability of zero events exceeding z. Its conditional excess survival at the return level is -log(0.99)/1.5, approximately 0.0067002, producing z approximately 6.9886 million.
These checks establish that the arithmetic is consistent with its assumptions. They offer no empirical confirmation of the assumed p_u, Poisson rate, beta, xi or independence. The difference between verifying a calculation and validating its input model must be preserved in any research package. Error checking must also inspect support conditions: xi<1 for the stated mean, the threshold range for all quantiles, and the applicability of a stationary event process for the annual OEP formula.
Documentary limits of the current study
The reviewed primary statistical research supports the mathematics of excess distributions, threshold diagnostics and methods of accounting for uncertainty. The supervisory texts establish international guidance and a specific European capital calibration. Public EIOPA resources document the existence and scope of European catastrophe and protection-gap datasets. The empirical large-claim study of Cebrián and coauthors concerns medical insurance, not property damage. The current dossier does not possess a microdata archive of global property insurance losses or the policy-wording sample necessary for estimating universal property tail parameters. Consequently its numerical scenarios are disclosed as hypothetical and its conclusions about portfolio behavior are conditional and methodological.
That limitation is scientifically productive. It prevents a technical monograph from presenting fitted parameters as observations when their provenance is absent. A defensible empirical follow-up can specify exactly which insurer data and documentation would make each claim estimable or refutable.
Integrative Comparative Scientific Synthesis
Threshold consistency across the entire research chain
The central synthesis is a consistency requirement: every statistical statement must remain attached to the same loss object through observation, estimation, contractual transformation and risk decision. Start with an event and exposure population. Document which losses are physically realized and which are covered. Apply and record any selection that governs what becomes a claim row. Estimate the relevant conditional severity and frequency distributions. Reapply contractual payouts and aggregate dependent events at the correct policy and annual level. Express a capital or reinsurance result with its exact risk measure and horizon. An interruption anywhere in this chain can alter the meaning of a correctly computed tail quantile.
This is a comparative research conclusion, not a newly named or proprietary theory. It follows from probability conditioning and the documented differences among insurance data and obligations. For example, an exact GPD on reported claims above a payment cutoff can describe those reported claims. It cannot be interpreted as the ground-up loss law below that cutoff or beyond a contractual cap without additional identifying evidence.
The portability question
Tail parameters are conditionally portable only where the underlying severity scale, contract design, claim population, exposure, peril mechanisms, valuation and data system are sufficiently comparable. Shape stability within one historical sample does not establish portability between residential fire and industrial explosion, between single-asset and catastrophe aggregates, or between jurisdictions with different limits and claims practices.
This explains why a methodological large-claim actuarial application can support use of excess modeling without supporting a universal numerical tail. The portable contribution of the method is the conditional likelihood, diagnostic logic and uncertainty treatment. The estimates themselves are specific to a dataset and estimand.
What the statistical and supervisory literatures jointly imply
Classical extreme-value theory gives principled approximations for sample maxima and threshold excesses. Applied statistical research identifies threshold choice and sparse observations as important sources of uncertainty. Supervisory guidance requires an insurer to understand integrated risk rather than simply report a well-fitting distribution. The IAIS supervisory principles and European solvency framework place the tail model inside broader institutional responsibilities. Together, these sources imply that a capital decision should record the effect of distributional and operational assumptions and specify which risks fall outside the modeled claim severity.
In international comparison, a universally stated requirement for reliable inference is stronger than a universally imposed numerical threshold. An amount that counts as a large insured claim differs across location, asset size, policy limits and currency. A percentage-of-value threshold may improve comparability for a defined class, but can erase information about absolute concentration and correlated reconstruction demand. The appropriate design may require both scales, with a declared transformation between them.
The most consequential decision errors
The first error is fitting a model to truncated records without reconstructing the observation mechanism. The second is treating a deductible-capped payment as an uncensored ground-up loss. The third is pooling claim-level and event-level observations while retaining independent-event frequency assumptions. The fourth is estimating a far-tail quantile from one selected threshold without displaying alternative credible thresholds and shape uncertainty. The fifth is transporting an extreme-event estimate into a company-wide solvency claim without matching the capital definition, horizon and financial structure.
Each error can survive conventional goodness-of-fit checking. They are detected instead by reconciling sources, units, conditional probabilities and contract mappings. A more elaborate statistical model does not repair an unrecorded loss population; a more restrictive contract does not by itself supply the lost information about the physical tail.
Priority directions for future empirical inquiry
A first high-value empirical study would pair full property exposure records with ground-up claims, paid and incurred development histories, deductible/limit metadata, and independently identified catastrophe clusters. It would compare fixed thresholds, threshold-uncertainty approaches and alternative tail laws using event-blocked temporal validation. A second would measure how capitalization changes when reporting thresholds and policy caps are modeled explicitly rather than silently ignored. A third would compare tail-shape stability across harmonized construction, peril and region strata while tracking replacement-cost inflation and insurance penetration. All three can report preregistered estimands and sampling designs.
For market-wide interpretation, public and private datasets require careful reconciliation. The EIOPA protection-gap dashboard and catastrophe data hub illustrate useful but different measurement purposes. A public economic-loss record, an insurer claim archive, an actuarial claim severity sample and a modeled hazard catalogue can be used together only after their definitions and inclusion rules are aligned.
Conclusion
Property-insurance large-loss tail modeling begins with a precisely defined monetary and observational object. Physical damage, economic loss, covered claim, insurer obligation, payment, event loss and annual aggregate form related but distinct variables. Administrative reporting conditions, contractual deductibles and limits, statistical tail thresholds, and reinsurance attachment points alter those variables in different ways.
Generalized Pareto peaks-over-threshold methods provide an established mathematical basis for analyzing large claim excesses under their domain-of-attraction assumptions. Their application requires documented selection mechanisms, sufficiently credible exceedance samples, threshold diagnostics, relevant exposure denominators and explicit uncertainty. Hill-type estimators, alternative severity families, spliced distributions, event clustering and portfolio simulation provide complementary views whose relevance depends on the target estimand.
The illustrative calculations show the practical sensitivity of high quantiles to the generalized Pareto shape and demonstrate the difference between claim-level percentiles and annual event return levels. Those calculations are reproducible hypothetical examples, and no observed property-insurance tail parameter is claimed. The comparison of statistical research and supervisory documents establishes why a claim-tail estimate is only an input to broader solvency, liquidity, catastrophe and reinsurance analysis.
The principal analytical finding is that preserving the identity of the loss variable across observation, statistical inference, contract terms and portfolio decisions is necessary for scientific interpretability. High-quality tail modeling consists of defensible assumptions, source-linked evidence, verifiable arithmetic, honest empirical scope and decision-relevant uncertainty. That conclusion is grounded in established extreme-value methods and the institutional structure of insurance; it introduces no new proprietary theoretical system.
References
Balkema, A. A., and L. de Haan. (1974). “Residual Life Time at Great Age.” The Annals of Probability 2(5): 792–804. Original journal article.
Cebrián, A. C., M. Denuit, and P. Lambert. (2003). “Generalized Pareto Fit to the Society of Actuaries' Large Claims Database.” North American Actuarial Journal 7(3): 18–36. Journal article.
Chavez-Demoulin, V., P. Embrechts, and M. Hofert. (2016). “An Extreme Value Approach for Modeling Operational Risk Losses Depending on Covariates.” Journal of Risk and Insurance 83(3): 735–776. Journal article.
Davison, A. C., and R. L. Smith. (1990). “Models for Exceedances Over High Thresholds.” Journal of the Royal Statistical Society: Series B (Methodological) 52(3): 393–425. Journal article.
Embrechts, P., C. Klüppelberg, and T. Mikosch. (1997). Modelling Extremal Events for Insurance and Finance. Stochastic Modelling and Applied Probability 33. Berlin and Heidelberg: Springer. Publisher edition.
European Insurance and Occupational Pensions Authority (EIOPA). (2025). “Dashboard on Insurance Protection Gap for Natural Catastrophes.” Updated December 5, 2025. Official dashboard.
European Insurance and Occupational Pensions Authority (EIOPA). (2023). “Catastrophe Data Hub.” Official data platform and scope note. Official data hub.
European Parliament and Council. (2009). Directive 2009/138/EC of 25 November 2009 on the Taking-Up and Pursuit of the Business of Insurance and Reinsurance (Solvency II), especially Articles 101 and 104. Official Journal of the European Union L 335, 17 December 2009. EUR-Lex legal text.
Hill, B. M. (1975). “A Simple General Approach to Inference About the Tail of a Distribution.” The Annals of Statistics 3(5): 1163–1174. Original journal article.
International Association of Insurance Supervisors (IAIS). (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups, revised December 5, 2024; see ICP 16 on Enterprise Risk Management for Solvency Purposes. Official online standards.
McNeil, A. J., R. Frey, and P. Embrechts. (2015). Quantitative Risk Management: Concepts, Techniques and Tools. Revised edition. Princeton and Oxford: Princeton University Press. Publisher book record.
Pickands, J., III. (1975). “Statistical Inference Using Extreme Order Statistics.” The Annals of Statistics 3(1): 119–131. Original journal article.
Reiss, R.-D., and M. Thomas. (2007). Statistical Analysis of Extreme Values: With Applications to Insurance, Finance, Hydrology and Other Fields. 3rd ed. Basel: Birkhäuser. Publisher edition.
Scarrott, C., and A. MacDonald. (2012). “A Review of Extreme Value Threshold Estimation and Uncertainty Quantification.” REVSTAT—Statistical Journal 10(1): 33–60. Open-access journal article.
Wadsworth, J. L., and J. A. Tawn. (2012). “Likelihood-Based Procedures for Threshold Diagnostics and Uncertainty in Extreme Value Modelling.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 74(3): 543–567. Journal article.