Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Property Insurance Probable Maximum Loss and Tail Capital

Abstract. Probable maximum loss (PML) is among the most consequential and frequently underspecified quantities in property insurance. It is invoked in site engineering, catastrophe underwriting, reinsurance placement, enterprise risk appetite, solvency assessments, and communication with boards and investors. Its meaning changes with the underlying loss measure, the annual or event horizon, the insured exposure population, the model, the return-period convention, and the point in the contractual loss-transfer chain at which the loss is measured. This monograph reconstructs the mathematical, engineering, contractual, and regulatory foundations of PML and connects them to annual loss distributions, extreme-value statistics, capital adequacy, and the practical governance of model uncertainty. The research distinguishes occurrence exceedance probability (OEP) from annual aggregate exceedance probability (AEP); physical, ground-up, insured, gross insurer, and net retained losses; quantiles from tail means; scenario stress from probabilistic calibration; and a risk metric from the financial resources required to absorb a loss. It compares the United States catastrophe-modeling literature and actuarial standards with the 2026 Canadian earthquake capital framework, the International Association of Insurance Supervisors' Insurance Capital Standard, and the European Union's Solvency II architecture. Mathematical examples are explicitly hypothetical and reproducible; no proprietary portfolio dataset or empirical backtest is represented as having been conducted. The central finding is that PML becomes decision-useful through a fully specified loss definition, exceedance distribution, time horizon, contract transformation, and validation record. Tail capital then requires additional judgments concerning expected loss, dependence, reinsurance recoverability, nonmodeled exposures, parameter uncertainty, liquidity, and the institution's full balance sheet.

Keywords: Property insurance; probable maximum loss; PML; occurrence exceedance probability; aggregate exceedance probability; catastrophe models; annual aggregate loss; value at risk; expected shortfall; extreme value theory; reinsurance; tail capital; solvency; model risk; earthquake; climate risk; insurance regulation.

Introduction

Property insurance converts damage to buildings, machinery, inventories, contents, and related insured interests into contingent financial obligations. An insurer must understand both the expected cost of ordinary claims and the potential for unusual events to damage many properties simultaneously. A hurricane crossing a densely insured coast, a large earthquake affecting industrial and residential structures, or a sequence of severe convective storms can produce losses whose size and concentration differ radically from ordinary independent claims. The NAIC account of catastrophe models describes the sequence of simulated hazard events, engineering vulnerability, exposure characteristics, and policy terms through which a physical event becomes a modeled insured loss. That sequence supplies the starting point for an exact analysis of PML.

The term probable maximum loss has persisted partly because it translates a difficult probability distribution into an apparently straightforward financial amount. The label is used for engineering judgments about a single industrial location as well as quantified portfolio catastrophe losses. In contemporary portfolio modeling, a PML often denotes a point on a specified exceedance-probability curve. In a property survey, however, an engineering loss scenario may reflect assumed failure of fire barriers or suppression equipment, with probability estimated only qualitatively. Treating those outputs as interchangeable produces an avoidable category error: the first is a distributional quantile under specified stochastic assumptions; the second is a conditional engineering estimate that may lack a calibrated annual frequency.

The research object is the loss distribution confronting a property insurer, measured at different stages of the chain connecting hazard, physical damage, covered claims, reinsurance, and available capital. The analytical unit varies: a single property, insurance policy, peril occurrence, regional portfolio, annual portfolio, or consolidated insurer. The same vocabulary can be employed at all these levels, but each inference must specify its unit. This monograph concentrates on catastrophe-driven property losses. It connects closely to large individual property claims without replacing the separate study of threshold estimation for noncatastrophe severity.

Three questions organize the investigation. First, under what definitions is PML a statistically interpretable property of an insured-loss distribution? Second, how do event sets, annual aggregation, extreme-value tails, spatial dependence, and insurance contracts determine the modeled amount? Third, when and how should that output influence solvency capital, reinsurance decisions, and public or supervisory communication?

The method combines documentary analysis of official supervisory and professional standards, reconstruction of standard probability identities, comparative interpretation of capital rules, and deliberately hypothetical arithmetic demonstrations. The published catastrophe-modeling actuarial standard establishes a professionally grounded frame for the selection, use, review, and evaluation of catastrophe models. This research adds original analytical synthesis concerning metric definition, consistency across the physical-to-financial loss chain, and the distinction between contractual risk transfer and genuine absorption of residual uncertainty. The synthesis applies established mathematics and regulatory texts; it does not establish a new proprietary theory or doctrine.

Definitions, Historical Context, and Primary Source Foundations

The different meanings of maximum loss

A maximum possible physical loss is a statement about the largest consequence compatible with a specified set of physical possibilities. A maximum foreseeable or estimated loss is an engineering judgment using a chosen scenario and assumed operation or failure of protective systems. An insurer's probable maximum loss can instead mean a loss level assigned a particular exceedance probability by a stochastic model. None of these meanings follows from the word 'probable' alone. The operational definition must disclose the hazard universe, exposure set, insurance terms, loss denominator, aggregation convention, and probability statement.

At a single industrial plant, engineers might estimate the damage caused by an uncontrolled fire spreading across connected production buildings. Their result depends on construction, separation, fire load, detection, sprinklers, emergency response, and interruption of production. A portfolio catastrophe PML, in contrast, combines many geographically distributed assets under hazard event footprints and statistical event frequencies. The former may dominate an individual property limit; the latter informs concentration management, treaty design, and capital. Both may be useful when properly labeled. A numerical comparison between them requires an explicit common loss variable and probability basis.

The NAIC model description distinguishes hazard, vulnerability, exposure, and financial calculation. This layered interpretation matters because the catastrophe hazard catalog may simulate a physical earthquake intensity or wind field, while policy conditions determine which consequences create a legal claim. Between those stages stand occupancy, building class, mitigation, valuation, repair-cost inflation, and secondary effects. A model of strong physical shaking may be scientifically credible while an insured PML is biased by wrong earthquake deductibles or missing business interruption limits. Conversely, impeccable contract coding cannot repair a hazard catalog that omits relevant earthquakes.

The loss chain and five financial quantities

For a property at location i exposed to event e, let H(e,i) represent local hazard intensity and D(e,i) the random physical damage conditional on hazard and asset vulnerability. Let C(e,i) be the economic consequence, including qualified interruption, and P(e,i) the amount payable after the actual property-policy coverage transformation. Portfolio gross insured event loss G(e) is the sum of P(e,i) over covered properties, together with the relevant adjustment-expense components specified by the reporting convention. Reinsurance recoveries R(e) produce retained event loss N(e)=G(e)-R(e), subject to collectability and any further contract-specific cash flows.

These are related quantities, yet their magnitudes can diverge. Replacement costs may exceed insured limits, and policy exclusions may remove otherwise severe physical damage. Substantial ground-up damage can therefore coexist with a comparatively small covered property claim. A nominally large reinsurance recovery may reduce accounting net loss while leaving credit risk and claims-payment timing exposure. For this reason, a reported 'gross PML' needs an explicit definition: gross of which reinsurance, after which policy deductibles, and inclusive of which expenses? The Canadian regulator supplies a concrete illustration: under the 2026 Minimum Capital Test, earthquake gross PML is after policyholder deductibles and before catastrophe and other reinsurance protection, with stated adjustments for model and data limitations.

An analytical convention used here is to reserve 'ground-up' for physical or economic losses before policy insurance conditions, 'insured gross' for the insurer's contractual claim exposure before ceded reinsurance, and 'net retained' for the claim exposure after modeled treaty recoveries. Actual collections and liquidity are evaluated separately. Each reported PML must further identify whether the loss amount includes allocated or unallocated adjustment expenses, loss amplification, demand surge, taxes, reinstatement premiums, and postevent contractual changes.

A primary-source hierarchy

The scientific meaning of PML derives from probability theory and catastrophe-model construction. Its use in professional actuarial practice is informed by standards, and its capital treatment can become binding under jurisdiction-specific law and supervision. These are different evidentiary categories. The ASOP No. 38 catastrophe standard addresses models as professional actuarial instruments in the United States; it does not create a universal statutory capital formula. The OSFI 2026 capital guideline applies within Canada's specified supervisory population and requirements. The IAIS Insurance Capital Standard sets an international group-capital framework with its own coverage and implementation context. The Solvency II capital provision establishes a one-year, 99.5% calibration of basic own-funds risk within the European Union's legal framework.

An evidence-based comparison must therefore retain both the date and legal perimeter of a rule. A 500-year Canadian earthquake preparedness requirement cannot be relabeled a universal worldwide solvency requirement. A 99.5% annual aggregate catastrophe charge cannot simply be treated as a 99.5% quantile of one occurrence. A financial stress test developed by a rating agency must not be represented as a binding insurance statute. The scientific content may overlap while the institutional function differs.

The evolution from observed catastrophes to probabilistic portfolios

Traditional property risk evaluation emphasized site inspections, probable maximum fire loss, insurance schedules, and aggregate insured values. Portfolio catastrophe models expanded the field by simulating large sets of geographically explicit events and calculating the damage to every exposure affected by each event. The NAIC history and technical overview locates modern catastrophe-model development in the technological and data improvements beginning in the 1980s. Event catalogs created a way to quantify rare yet plausible losses beyond the particular catastrophes seen in a short claims record.

This development did not eliminate judgment. Historical records are short relative to long return periods, building stock changes through time, and some current perils lack mature observation or vulnerability models. The statistical result remains conditional on assumed hazard rates, exposure geography, repair costs, deductibles, and dependencies. Scientific progress makes these assumptions explicit and testable rather than eliminating their influence.

PML Definitions and Return Periods

A random variable before a number

Define a probability model for the annual number of relevant catastrophe occurrences K and their insurer losses L1,...,LK, with zero when no occurrence takes place. The annual maximum event loss is

M=max⁡(0,L1,…,LK),M=\max(0,L_1,\ldots,L_K),

and the annual aggregate loss is

S=∑j=1KLj.S=\sum_{j=1}^{K}L_j.

Here L is measured at a clearly specified financial stage; gross insured losses and net retained losses define distinct variables. The annual occurrence exceedance probability at threshold x is OEP(x)=P(M>x). Annual aggregate exceedance probability is AEP(x)=P(S>x). For nonnegative losses, S is at least M in every simulated year. Therefore AEP(x) is at least OEP(x) at any common threshold x. Their corresponding upper quantiles are likewise ordered when calculated from the same consistent loss definition and event-year model. The ordering may cease to be directly interpretable if one curve is gross of reinsurance and the other is net, or if the modeled loss components differ.

The NAIC explanation of exceedance curves explicitly distinguishes occurrence and aggregate ranking and connects return periods to model output. The result is an important warning for capital users: a $100 million 100-year occurrence PML and a $100 million 100-year aggregate PML are distinct claims about event accumulation. Multiple medium-size catastrophes can make the annual aggregate tail larger than the maximum-event tail, even when no individual catastrophe looks extraordinary.

For probability p and annual loss variable X, define the upper loss quantile as Q(1-p)=inf{x:P(X<=x)>=1-p}. Under a sufficiently continuous distribution, P(X>Q(1-p))=p. Under discrete distributions or atoms at zero, equality can fail, so a model report should state its quantile and exceedance conventions. This detail matters for low-frequency portfolios, simulation resolution, and contractual layers with large probability masses at an attachment point.

The return-period convention and its limits

In a stationary model with a well-defined annual exceedance probability p, the exceedance return period is T=1/p years. A return period of 100 years therefore means a 1% annual exceedance probability for the particular loss threshold and model, not that the insurer waits a century between events. The probability of at least one annual exceedance in n independent model years is 1-(1-p)^n. For a 100-year threshold and 30 independent, identically distributed years, this yields approximately 26.03%. For a 200-year threshold the corresponding one-year exceedance probability is 0.5%; for a 500-year threshold it is 0.2%.

Independence across years is required for this particular n-year formula. Serially dependent climate patterns, changing exposure, or nonstationary hazard rates require an event-time representation or a product of conditional survival probabilities, not a mechanical extension of the stationary expression. Even a stationary Poisson event model can generate several damaging events in one calendar year, and annual aggregation must account for them. The return period is thus a communication of a model probability conditional on time and exposure, not a physical calendar prediction.

A reproducible Poisson-event calculation

Assume, solely for mathematical illustration, that K has a Poisson distribution with mean λ events per year and event losses L are independent identically distributed nonnegative marks with cumulative distribution F(x), independent of K. The number of events whose losses exceed x is Poisson with mean λ[1-F(x)]. Consequently,

P(M>x)=1−exp⁡{−λ[1−F(x)]}.P(M>x)=1-\exp\{-\lambda[1-F(x)]\}.

For an occurrence threshold with return period T, the corresponding single-event exceedance fraction satisfies 1-F(x)=-ln(1-1/T)/λ. This event-mark probability is not in general numerically equal to the annual exceedance probability 1/T. Confusing the two probabilities miscalibrates risk, particularly when event frequency is material.

Under the same assumptions and a finite mean event loss, the expected annual aggregate loss is E[S]=λE[L]. It follows from conditioning on the event count. Expected annual loss is a central pricing or provisioning input, whereas a high quantile of M or S expresses tail intensity. A risk portfolio can have a moderate mean and an extreme upper quantile. A capital requirement may be based on an unexpected-loss concept, in which case comparing the full quantile to available capital without explaining the already-funded expected component can double-count part of the economic requirement.

Event clustering, zero-loss years, and the meaning of an exposure period

A stochastic catastrophe model may generate a full year of correlated events rather than independently drawing isolated footprints. This matters because tropical cyclones, aftershocks, flood sequences, or weather regimes can create clusters. AEP reflects the total severity and number of covered events within the specified year. OEP reflects only the largest modeled event in that year, though clustering can still alter its distribution through the frequency process. A claim adjustment timeline or catastrophe occurrence clause may group physical disturbances differently from a catastrophe model's event identifier.

If an insurer reports a PML for an annualized portfolio, the exposure base should reflect the relevant inception dates, anticipated new writings, and planned exits. A static single-day list of insured buildings can be informative for a current accumulation snapshot but is not automatically a full-year stochastic exposure. Currency, replacement-cost indices, and the chosen exposure valuation date are part of the loss definition. A return period quoted without an effective exposure date and horizon can become economically obsolete before the underlying hazard model has changed.

Operational specification required for every PML disclosure

A complete PML disclosure identifies (1) the loss variable and currency, (2) the insured interests and exposure valuation date, (3) the perils and territories, (4) an event-maximum or annual-aggregate basis, (5) the gross or net financial stage and contract version, (6) the annual probability or quantile and its convention, (7) the simulation horizon and key dependency assumptions, and (8) model version and treatment of nonmodeled exposures. These elements are logically necessary to compare two numbers. An apparently precise figure without them is an incomplete measurement statement.

The same discipline applies when a site engineer reports a maximum-loss scenario without a probabilistic distribution. The report should state the assumed initiating event, protection state, physical spread, repair and interruption costs, applicable coverage, and whether a frequency or return period has actually been assessed. That engineering scenario can be a useful stress input while remaining distinct from a statistically calibrated PML.

Scenario Versus Stochastic Tail

The engineering scenario and the simulated year

A deterministic catastrophe scenario begins with selected premises: an earthquake rupture, a cyclone track, a fire spreading from a specified ignition point, or a flood exceeding a defense. Hazard intensity is propagated through vulnerability functions and insurance conditions. The result can reveal concentrations, fragility, and critical financial consequences that a point estimate of annual probability conceals. The scenario's informational value does not depend on assigning it a spurious return period. Its validity depends on whether the assumed hazard and physical responses are credible, whether insurance calculations are consistent, and whether consequences are documented.

A stochastic catastrophe model adds a weighted population of possible scenarios and a rule for their occurrence. It can simulate event frequencies, locations, footprints, damage ratios, and annual event sequences. It supplies an annual loss distribution from which OEP, AEP, expected annual loss, quantiles, and conditional tail means may be calculated. The distinction is methodological: a very severe single-event scenario can be part of a distribution, but its severity alone establishes neither its annual frequency nor its contribution to an aggregate annual quantile. The NAIC catastrophe-model architecture expressly separates event occurrence, vulnerability, insurance terms, and resulting loss metrics.

Scenario and stochastic analyses should communicate with each other. A credible extreme scenario tests whether the stochastic event catalog omits a material physical possibility. Conversely, the event catalog can show whether a familiar historical disaster is typical of the modeled tail or unusually concentrated on today's exposure. A targeted deterministic stress may challenge a critical reinsurer, a flood defense, or a construction-type vulnerability assumption without pretending that the stress probability has been objectively measured.

Construction of the modeled catastrophe-loss distribution

The event catalog assigns candidate events spatial footprints, intensity measures, and occurrence rates. For a windstorm, the model may use local wind speed and duration; for an earthquake, ground-motion intensity, geological conditions, and secondary processes; for flooding, depth, duration, and topography. The exposure database locates insured assets and records their construction, use, replacement value, and policy details. Vulnerability functions map event intensity and asset attributes into distributions of damage. Financial modules apply deductibles, coinsurance, limits, coverage grants, and special provisions to obtain insured losses.

Every layer can distort the tail. Mislocated insured assets may be assigned an incorrect floodplain or wind footprint. Occupancy or construction coding may be particularly important for warehouses, large commercial complexes, or industrial roofs. Updating building values without updating business-interruption exposures can understate total loss. A model calibrated to ordinary repair costs may miss demand surge after a widespread event. A model of physical damage may fail to represent coverage for ordinance-and-law upgrades or debris removal. Jointly, such errors can change not only expected loss but the ranking of portfolio segments in the far tail.

Secondary effects demand special care. Fire following earthquake, storm surge after tropical cyclone, or supply-chain interruptions caused by multiple damaged sites can magnify the same primary event. The IAIS catastrophe specifications require associated secondary perils and loss amplification to be considered within the relevant catastrophe treatment. Whether a peril is classed as primary or secondary for modeling convenience should not determine whether its legally covered loss is included in a complete insurer risk inventory.

The simulated year is then formed by sampling an event process and applying corresponding financial transformations. When policy contracts have event-wise deductibles and aggregate annual limits, the order of operations is material: a deductible applied to each event is not equivalent to a deductible applied once to annual total loss. If the model simulates event loss distributions but later reinsurance calculations assume fixed recovery shares, treaty attachment and exhaustion may be materially misstated.

Extreme-value theory and its domain of application

Historical catastrophe records typically contain few observations of the very largest portfolio losses. Extreme-value theory supplies asymptotic tools for studying tails while requiring careful decisions about thresholds, stationarity, and dependence. Coles's treatment of extreme-value modeling covers block maxima, threshold exceedances, dependent sequences, nonstationary extremes, and multivariate problems. The general theory is relevant, but its assumptions must be assessed for the insurance-loss variable actually modeled.

In a peaks-over-threshold analysis, if a sufficiently high threshold u is appropriate and Y=X-u conditional on X>u, a generalized Pareto approximation for the conditional exceedance survivor is

P(Y>y∣X>u)≈(1+ξyβ)−1/ξ,P(Y>y\mid X>u)\approx\left(1+\frac{\xi y}{\beta}\right)^{-1/\xi},

where β>0, the support requires 1+ξy/β>0, and the limit ξ=0 corresponds to exp(-y/β). The shape ξ governs heaviness of the tail: for ξ>0 the upper tail is unbounded and heavy; for ξ<0 the generalized Pareto upper endpoint is finite. In this fitted family, a mean exceedance exists when ξ<1, subject to the assumptions. A threshold is therefore part of the scientific model and must be justified with diagnostics, uncertainty estimates, and sensitivity analyses.

This approximation is not a universal replacement for a spatially explicit catastrophe model. An individual building's repair bill, a single-event aggregate over thousands of properties, and a full year's insured aggregate are different random variables. A fit to individual claim severities cannot be transferred automatically to portfolio event PML, because common event footprints induce correlated claims. Fitting a tail distribution to synthetic annual catastrophe loss output can be useful for smoothing or extrapolation, but then its empirical basis is the catastrophe simulation and all of that model's assumptions, not an independent field observation.

A common methodological error is to extrapolate a return period much longer than the information content of both historical data and the model event generator. Mathematical possibility of fitting a parametric tail does not establish empirical identification of the farthest quantiles. Threshold selection, claim-limit censoring, event dependence, parameter uncertainty, and secular trends all require explicit evaluation. Model risk should be analyzed across plausible structural specifications rather than hidden inside a single nominal confidence level.

Why annual aggregation changes the capital problem

Consider a hypothetical insured property portfolio experiencing two independent contract-defined catastrophe events during one year, with gross insured event losses of $90 million and $70 million. Its occurrence maximum is $90 million and its annual aggregate gross loss is $160 million. If a simple occurrence excess-of-loss treaty pays min(max(G-50,0),50) in millions for each qualifying event, the first event recovers $40 million and the second $20 million. Net retained event losses are $50 million each; annual retained loss is $100 million. This assumes no aggregate treaty cap, reinstatement cost, exhaustion from earlier events, counterparty failure, or other policy conditions. It is an arithmetic example, not a claim about an actual insurer.

Now imagine the same $160 million gross amount arises from one qualifying occurrence. The same single-occurrence layer recovers only $50 million, leaving $110 million retained. Equal gross annual totals thus generate different net results when event grouping and treaty exhaustion differ. If the treaty has a single annual aggregate cap of $50 million, the two-event year's recovery also changes from $60 million to $50 million. Reinsurance results must consequently be calculated on the relevant event-year path, with actual occurrence definitions and aggregate provisions, before a net AEP is derived.

The exercise establishes a general point: the appropriate annual net loss variable is a function of the complete path of events and contracts, not ordinarily a constant fraction of gross annual losses. Quantile algebra cannot replace scenario-by-scenario application of a nonlinear treaty. In particular, Q(G-R(G)) generally differs from Q(G)-Q(R(G)); additivity of quantiles is a special-case property, not a general identity. Recoveries at the nominal 200-year gross event are likewise not automatically the recoveries associated with the 200-year net annual loss.

Tail quantiles and expected shortfall

For annual loss S, value at risk at confidence q is the quantile VaR_q(S)=inf{x:P(S<=x)>=q}. Expected shortfall, defined in the generalized quantile-integral form, is

ESq(S)=11−q∫q1VaRv(S) dv.ES_q(S)=\frac{1}{1-q}\int_q^1 VaR_v(S)\,dv.

For a continuous distribution it equals the mean loss beyond the appropriate q-quantile, provided the tail mean exists. For discrete distributions the integral definition handles mass at the quantile more carefully than naive conditioning on a strict exceedance. The coherent-risk framework of Artzner, Delbaen, Eber, and Heath (1999) explains why subadditivity matters when aggregating risk, and Acerbi and Tasche (2002) examine the coherent interpretation of expected shortfall.

A VaR point answers how far a specified percentile reaches; expected shortfall reports the average of the worst 1-q probability mass. For a hypothetical annual loss distribution with loss $0 in 90% of years, $100 million in 9% and $1 billion in 1%, VaR at 99% under the stated infimum convention is $100 million, while ES at 99% is $1 billion. The difference does not imply that expected shortfall is automatically superior for every regulatory purpose; rather, it demonstrates that a single percentile can conceal loss severity beyond the selected cutoff.

Tail estimates still depend on model completeness. If two competing hazard models have similar VaR at 99.5% but radically different losses beyond that percentile, expected shortfall may expose their divergence, although both measures inherit the models' uncertainty. A complete capital conversation can report quantiles, expected shortfall, stress scenarios, and model ranges without conflating them into a single universally valid risk number.

Numerical confidence is not scientific confidence

Suppose an insurer samples 10,000 independent simulated annual catastrophe outcomes and estimates the 99.5th percentile from their empirical distribution. On average only 50 draws lie beyond the true 99.5th percentile. The binomial standard deviation of this exceedance count is sqrt[10,000×0.005×0.995], about 7.05. This computation is a count diagnostic, not a direct confidence interval for the estimated loss quantile; that requires information about the loss density and sampling dependence near the threshold. The distribution may also be generated by an event-weighting algorithm rather than independent identical annual draws.

More simulations can reduce Monte Carlo sampling error conditional on the model. They do not establish that hazard frequency, vulnerability, exposure values, or insurance wording are correct. A million resampled synthetic years from a misspecified model can produce a highly stable yet misleading 200-year PML. Statistical precision, parameter uncertainty, and structural validity must therefore be reported separately.

Capital Adequacy and Communication

The move from a catastrophe loss to a capital requirement

Capital is an available financial resource that supports an insurer's capacity to honor obligations across adverse states. An estimated property catastrophe PML is a modeled loss amount, not an independently sufficient statement of required regulatory capital. The relationship depends on expected-loss funding, accounting liabilities, diversification with non-property risks, contractual risk mitigation, counterparty exposures, and the eligible composition of capital. It also depends on the rules governing the institution.

A simple analytical expression is capital need ≈ selected adverse net loss minus the part of expected loss already funded and legitimately available for absorption, plus relevant additional risks and margins. That schematic is a diagnostic, not a replacement for binding solvency formulas or insurer balance-sheet valuation. It makes explicit that one cannot compare a ground-up 200-year physical damage number directly with reported equity and infer solvency. The insurer may have insurance exclusions, large deductibles, significant reinsurance, asset risks, and claims provisions whose interaction changes both the numerator and the denominator.

The regulatory examples demonstrate substantive diversity. The Solvency II Article 101 criterion is the one-year 99.5% value at risk of basic own funds for the undertaking's quantifiable risks, with future business and risk-mitigation effects treated within the law. This is an enterprise balance-sheet calibration, rather than an instruction that property gross event PML must equal the total Solvency Capital Requirement. The risk-module design in Article 104 separately aggregates underwriting, market and counterparty modules under specified rules.

The Insurance Capital Standard's natural catastrophe definition

The IAIS Insurance Capital Standard articulates a particularly instructive definition in Level 2 text L2-184 to L2-186. Stochastic catastrophe models may be used, and loss amounts must consider affected lines of business, nonmodeled exposure, expected new business, and material unmodeled perils and regions. The natural catastrophe risk charge is the difference between the 99.5th percentile and the mean of total annual aggregate losses net of protections. The annual aggregate combines regions and perils. This specification identifies the loss period, financial stage, scope, tail probability, and expected-loss subtraction explicitly.

If S_net denotes the annual aggregate covered catastrophe losses net of eligible protections within the ICS scope, its stated calculation can be represented as VaR_0.995(S_net)-E[S_net], subject to the standard's full definitions and valuation treatment. It is accordingly improper to describe that component simply as the 200-year gross occurrence PML. Nor should this catastrophe component be described as the entirety of an insurance group's capital requirement. It is one part of a larger regulatory method and operates within the IAIS implementation perimeter.

The mathematical relationship between OEP and AEP demonstrates why the distinction matters. A portfolio with moderate-sized, repeated windstorms may have an annual aggregate 200-year tail significantly larger than its single-occurrence 200-year tail. The difference can be amplified by inflation, territorial concentration, and treaty exhaustion. Across a combined book, peril correlations and nonmodeled perils also influence capital. A treaty designed solely around the largest single event could leave annual aggregate capital exposed to repeated events.

The Canadian earthquake-capital example

Canada's 2026 Minimum Capital Test, effective January 1, 2026, offers a jurisdictionally specific contrast. In section 4.5.1, OSFI assesses financial preparedness for a country-wide earthquake event at a 500-year return period. The eastern and western Canada 500-year PMLs represent 99.8th-percentile exceedance-curve levels, with adjustments for data quality, model uncertainty, and nonmodeled exposures. For the modeled approach, the guideline specifies a country-wide aggregation involving the sum of the eastern and western PMLs raised to the 1.5 power and the result raised to the reciprocal 1.5 power.

The 500-year preparedness benchmark is more remote than the 200-year (99.5%) annual quantile often encountered in other capital frameworks. The comparison does not show that Canadian law universally demands a larger whole-company solvency buffer, because the object of measurement, use of reinsurance, aggregation and available resources differ. OSFI's test is tied to earthquake exposure and its designated financial-preparedness mechanism. The guideline identifies an earthquake reserve component and the relevant financial resources including eligible capital and surplus, reinsurance, specified earthquake premium reserves, and approved capital-market financing, with limits on recognition.

The choice of Canadian gross PML, after policyholder deductibles and before reinsurance, is especially informative. It isolates the initial property catastrophe claim exposure before applying a separate test of funding and protection. When evaluating such a standard, the researcher must preserve the legal definitions of qualifying resources and aggregation formulas instead of replacing them with a generic net-value-at-risk heuristic. The source document, publication date and effective date are also necessary to prevent obsolete regulatory comparisons.

Solvency II's whole-balance-sheet orientation

Under Article 101 of the Solvency II Directive, a company's Solvency Capital Requirement is calibrated to all quantifiable risks through a one-year 99.5% value-at-risk measure of basic own funds. The loss of basic own funds reflects both assets and liabilities. Market value changes, counterparty default, insurance provisions, catastrophe losses, and risk mitigation can therefore interact. A property-catastrophe model contributes evidence to a module or internal-model calculation, but the organization's whole-balance-sheet capital is its own analytic object.

European internal-model rules reinforce this separation. Article 112 on model approval defines supervisory approval of full or partial internal models; Article 122 on calibration allows alternative periods or risk measures in internal modeling only when the required equivalent policyholder protection can be demonstrated. The Article 124 validation requirement calls for ongoing model validation, testing against experience, sensitivity to key assumptions, and data appropriateness. A high reported PML confidence level alone satisfies none of these wider validation obligations.

The distinction also prevents a common use-of-model error. An insurer may calculate a gross windstorm PML for underwriting concentration, a net occurrence PML for treaty placement, an annual net aggregate VaR for catastrophe capital, and a consolidated own-funds risk metric for solvency. All four outputs can be correct and materially different. They must be connected through reconciliation of definitions, contract transformations, and balance-sheet effects, not forced into numerical identity.

Reinsurance, securities, and residual capital

Reinsurance transfers part of the insurer's contractual loss exposure to another balance sheet, subject to policy wording, event definitions, attachment, limits, reinstatement, exclusions, and recoverability. Catastrophe bonds and other alternative risk-transfer arrangements may also provide contingent funding under precisely defined triggers. Their risk-reduction value depends on whether contract triggers match the insured losses, on collateral and timing, and on whether adverse events create multiple financial demands. A strong modeled net PML is credible only when these mechanisms are represented at the correct event or annual level.

An indemnity excess-of-loss treaty with low attachment can reduce moderately severe property loss yet exhaust in the farthest tail. An aggregate protection can reduce repeated moderate events but may provide less coverage for a single extraordinary event if its cap is reached. A parametric contract may settle quickly but create basis risk between the measured trigger and the actual covered property claims. The capital effect cannot be inferred from nominal treaty limit alone. A defaulted reinsurer may restore some catastrophe exposure at the same time as the insurer must fund claims, and an expensive reinstatement may erode postevent financial resources.

In addition, the paid-loss timeline is separate from ultimate insured loss. Property reconstruction may take years, with adjustment, litigation, temporary accommodation, inflation, and repair-capacity constraints affecting when payments become due. A firm could be adequately capitalized in an ultimate-loss sense yet struggle with immediate claims liquidity. This monograph therefore treats modeled loss, legally recoverable reinsurance, collectible cash flow, eligible capital, and liquid resources as distinct dimensions of catastrophe preparedness.

Governance and communication of capital metrics

Board reports should identify the question answered by each measure: concentration exposure, catastrophe probable maximum, survival under a chosen stress, treaty sufficiency, or compliance with a supervisory standard. The presentation should state at least the peril, territory, exposure date, gross/net convention, annual or occurrence basis, percentile, model and version, uncertainty treatment, and major exclusions from scope. A comparison table should preserve the same definition across years or explicitly restate prior values when assumptions change.

Professional obligations support this discipline. The ASOP No. 56 modeling guidance concerns model purpose, appropriateness, validation and communication, while the ASOP No. 38 catastrophe guidance directly addresses reliance on sophisticated catastrophe models. In the European internal-model context, Article 125 documentation requires documentation of theory, assumptions, mathematical and empirical bases, and circumstances in which a model may not operate effectively.

A PML presentation should use dollar ranges or sensitivities when the evidence does not warrant apparent exactitude. For example, it may report the baseline 200-year net annual aggregate loss alongside uncertainty from location data, building values, severe-convective-storm vulnerability, and treaty exhaustion. It should also show whether the number has changed because exposure increased, the model was recalibrated, policy wording changed, or the selected risk measure changed. This attribution is necessary for decisions: a higher PML caused by updated geocoding has a different practical meaning from a higher PML caused by substantial new property accumulation.

Empirical Verification, Uncertainty, and Competing Explanations

Evidence appropriate to the question

Empirical validation begins with specifying what can actually be observed. Historic physical-hazard records include instrument measurements, flood elevations, wind fields, earthquake intensities, and event footprints. Engineering evidence records structural and nonstructural damage conditional on those intensities. Insurance claims datasets record payments or reserves after terms, limits, reporting processes, and coverage disputes. Market catastrophe-loss publications aggregate across changing insured portfolios and reporting institutions. These datasets answer different questions. A discrepancy between a damage map and claims paid may result from policy exclusions rather than failed hazard physics; a discrepancy between claims ultimately paid and initial reserves may reflect subsequent repair-cost inflation rather than a change in event intensity.

The ASOP No. 23 data-quality standard requires attention to the selection, review, and limitations of data used for actuarial work within its professional scope. A scientifically accountable catastrophe assessment should document the location and geocoding quality of exposures, building attributes, coverage terms, data timestamps, incompleteness, duplicated insured assets, and the treatment of large commercial values. Insurance claims must be reconciled with event identifiers and policy versions. External hazard datasets should be evaluated for their period of observation, spatial resolution, instrument completeness, and suitability for the exact asset population. The ASOP No. 39 catastrophe guidance describes why catastrophe ratemaking requires attention to frequency, severity and changes in exposure beyond a simple historical catastrophe-loss ratio. Its pricing application is related to, but institutionally distinct from, a capital-tail calculation.

A well-designed empirical program includes at least four distinct tests. First, does the hazard model reproduce relevant measured intensity and spatial dependence for events within its observable range? Second, do vulnerability functions reproduce independent damage evidence for matched building classes and hazard levels? Third, does the financial module reproduce actual policy payouts when supplied with verified event damage and policy terms? Fourth, does the combined annual loss model produce a defensible distribution of observed portfolio losses over compatible exposure periods? Each question can fail independently, and the fourth is difficult at remote return periods because the observed record is short.

Identification and the difficulty of rare-event backtesting

When the assessed level is a one-in-200-year annual exceedance, even a thirty-year stationary observation window provides very little direct information about its frequency. Under an idealized independent Bernoulli annual model, the probability of observing no exceedance in thirty years is (199/200)^30, approximately 86.04%. A period without a modeled 200-year loss is therefore consistent with the model, even before allowing for portfolio changes. Such a record cannot validate the 99.5th annual quantile by simple absence of exceedances. Positive evidence must combine independent hazard science, engineering validation, catalog plausibility, exposure audit, loss-function tests, and carefully reported model uncertainty.

Claims experience also faces selection. Only insured hazards and policyholder populations generate covered claims, and the portfolio may change rapidly because of underwriting, nonrenewal, premium increases, construction, and changes in insurance take-up. A disaster's total economic loss, insured industry loss, and one insurer's contractual retained loss have different denominators. A claim sample drawn from postevent settlements overrepresents damaged and insured properties, whereas a vulnerability function requires information about damaged and undamaged exposed assets. Failure to make this distinction can exaggerate conditional damage frequency.

The 2026 Swiss Re Institute catastrophe analysis estimates USD 107 billion in worldwide insured natural catastrophe losses for 2025 and emphasizes wildfire and severe convective storm experience. This is contextual industry aggregate evidence for the importance of nontraditional or secondary peril portfolios. It is not a direct observation of a particular insurer's PML, nor does one year's aggregate establish the distribution of a 200-year portfolio outcome. Without normalization for exposure, insured penetration, currency, and claims maturity, such global totals cannot serve as a standalone validation of a local portfolio tail.

A threshold-sensitivity calculation

The consequences of parametric uncertainty can be demonstrated without introducing an empirical dataset. Suppose an illustrative annual aggregate loss variable X has a threshold u=$100 million and probability P(X>u)=0.05. Assume a generalized Pareto exceedance approximation with scale β=$10 million. For a desired annual exceedance probability 1/T smaller than 0.05, the quantile extrapolation under a fixed shape ξ different from zero is

xT=u+βξ[(Tpu)ξ−1],pu=P(X>u).x_T=u+\frac{\beta}{\xi}\left[(T p_u)^\xi-1\right],\qquad p_u=P(X>u).

For ξ approaching zero, the increment tends to βln(Tp_u). For T=200 years, Tp_u=10. With ξ=0, the resulting hypothetical level is approximately $123.03 million; with ξ=0.2 it is $129.24 million; with ξ=0.4 it is $137.80 million. The same threshold, probability and scale assumptions thus generate materially different remote loss levels solely through the chosen tail shape. A sensitivity result is not itself evidence that any one shape parameter applies to property insurance.

This example is deliberately based on the annual loss variable and therefore relates to an AEP-style threshold. If the underlying data were individual building claims, catastrophe event losses, or annual maxima, the mapping to an annual aggregate loss level would require its corresponding frequency and dependence structure. The relevant scientific literature includes Coles (2001), whose threshold-modeling treatment also emphasizes the role of dependent and nonstationary sequences. Parameter stability, graphical threshold diagnostics, confidence intervals or likelihood-based uncertainty, exposure change, and alternative family comparisons should all accompany real applications.

Structural model uncertainty

The principal components of a catastrophe model can be plausible in isolation while their combination misstates the tail. A windstorm event generator may reproduce coastal landfalls but underrepresent inland decay. A damage function calibrated on newer roofs may underestimate aging industrial stock. A portfolio geocoder may assign an enterprise's entire insured value to the headquarters rather than its insured warehouses. A policy engine may interpret a per-building deductible as a per-event deductible. Each error affects a different stage of the physical-to-financial chain. Model disagreement should therefore be decomposed by stage rather than summarized as an unexplained percentage margin.

Some losses arise from sources not fully present in established catastrophe models. Examples include debris-removal costs, damage from a peril linked to a primary event, contingent business interruption, legal disputes about coverage, postcatastrophe reconstruction constraints, and model gaps for certain territories or secondary perils. The IAIS nonmodeled-loss allowance demonstrates that a capital framework may require accounting for omitted exposures and hazards rather than assigning them a tacit probability of zero. Canada's OSFI earthquake adjustment instructions similarly incorporate data quality and model uncertainty in the PML definition used for its earthquake risk assessment.

Sensitivity analysis should vary model assumptions that can influence decisions: hazard recurrence and climate trends, building-value inflation, geocoding error, asset vulnerability, insurance take-up, business-interruption duration, deductibles and exclusions, and contract exhaustion. Some variables may be correlated. Repair-cost inflation following a regional disaster can increase both the cost of damage and the insurer's need for liquid cash; treating those effects as independent could understate the relevant joint stress. A single optimistic scenario should not determine treaty adequacy when materially different plausible hazard and contractual paths exist.

Climate change, exposure change, and nonstationarity

A long-run hazard estimate assumes some representation of the climate or geophysical process under which losses occur. For weather-related perils, conditions can evolve. Exposure also changes because construction moves, building values rise, land use shifts, and mitigation improves or deteriorates. A return period estimated under one reference climate and asset population can become inconsistent with the insurer's forthcoming annual risk. The problem is not resolved by labeling one scenario as a 200-year event while leaving its climate reference period implicit.

A time-dependent exceedance probability p_t can be expressed as P(X_t>x) conditional on year-specific hazard and exposure. If annual exceedances are conditionally independent across years with different probabilities, the probability of at least one exceedance over n years is 1-Π_t(1-p_t). If temporal dependence persists, this product no longer follows without further modeling. This is a standard probabilistic identity, offered here to show why a stationary period label must be revisited for changing risk.

The source-based discussion of climate should separate observed trends, mechanistic physical attribution, model projections, and insurer-specific financial effects. An increase in industry insured losses might reflect property value growth, insurance penetration, construction in hazard zones, weather change, or combinations of these factors. The Swiss Re natural catastrophe assessment identifies rising exposure as a substantial driver of industry losses; it does not permit direct attribution of a particular portfolio's PML movement to climate alone. Useful model comparison therefore holds the exposure constant when studying hazard change and holds hazard assumptions constant when studying exposure change, before exploring their combined evolution.

Professional validation, disclosure, and competing hypotheses

The ASOP No. 38 model-review requirements and ASOP No. 56 modeling requirements support a documented account of purpose, input appropriateness, reliance, model limitations, and validation. Within Solvency II, Article 124 calls for internal-model performance monitoring, comparison with experience, sensitivity assessment, and evaluation of data adequacy. Validation must be organized around the relevant model and supervisory perimeter; an internal assurance memorandum cannot be automatically represented as formal regulator approval.

Competing explanations for an observed increase in PML should be tested in a controlled decomposition. The increase may come from a new hazard catalog, improved location records, exposure growth, a new vulnerability function, broadened business-interruption cover, changed treaty terms, inflation, or a revised annual aggregation convention. Analytically, rerunning the same portfolio under successive model versions can isolate part of the model effect; rerunning the same model on successive exposure snapshots isolates part of the portfolio change. The remaining interaction requires explicit treatment because model, contracts, and exposure may have changed jointly. The appropriate scientific claim is a documented attribution within stated assumptions, not a unique causal identification based on one headline PML number.

For this monograph, documentary verification covers the official texts and scholarly metadata cited in References, and the mathematical calculations are transparent synthetic examples. No full property-level claims database, catastrophe vendor implementation, observed loss backtest, proprietary contract review, or subscription-only monograph full-text audit has been conducted. The limits define the scope of the conclusions and prevent documentary reasoning from being misrepresented as new observational evidence.

Integrative Comparative Scientific Synthesis

A common analytic language for different institutions

A coherent research comparison can place every PML claim on the same sequence of questions. What physical event process is being represented? Which assets and risk characteristics are exposed? Which losses are covered under current policy terms? Are losses measured by maximum occurrence or annual total? What is the exceedance probability and time basis? Which reinsurance transformations are applied, in what event order? What uncertainties and unmodeled components remain? Finally, which capital, liquidity, or risk-management decision uses the result? These questions translate engineering, actuarial, and supervisory vocabulary into compatible dimensions without imposing identical legal rules on distinct institutions.

The result explains why the 2026 Canadian earthquake 500-year gross PML and the IAIS 99.5% annual aggregate net catastrophe loss are not contradictory. They measure different objects and perform different functions. The former forms a defined input to Canadian earthquake preparedness, with a separate consideration of qualifying financial resources. The latter measures a component of unexpected annual catastrophe loss across territories and perils in an international group-capital framework. Solvency II's 99.5% basic own-funds VaR moves still further from a stand-alone property event loss because it concerns the firm's one-year capital position. Each can be appropriate within its specified objective.

The NAIC model literature emphasizes measurement from event to insured loss, while ASOPs require users to evaluate the suitability of models and data for actuarial decisions. These positions complement the regulatory approaches. A model that produces a technically correct OEP curve is still incomplete as a capital instrument if expected new business, material secondary perils, reinsurance exhaustion, and nonmodeled exposure are omitted from the financial calculation.

Risk concentration as the connective mechanism

The central mechanism linking property PML with tail capital is concentration. A catastrophe hazard creates simultaneous or closely spaced damages among assets that share a geographical footprint or structural vulnerability. Insurance policy terms transform these damages into correlated obligations. Event aggregation and repeated events amplify the upper annual loss distribution. Reinsurance removes specified contractual slices, leaving tail outcomes exposed to attachment, exhaustion, and counterparty solvency. The residual probability distribution interacts with other balance-sheet risks to determine capital adequacy. Every stage has a distinct evidence base, and errors may accumulate.

This mechanism explains why diversification by number of insured buildings can be insufficient. Thousands of insured properties can have strong event dependence if they occupy the same cyclone corridor or flood basin. Conversely, more dispersed exposures may reduce extreme concentration even when ordinary claim frequency remains unchanged. The capital-relevant effect is measured through the full multivariate annual-loss distribution rather than through a count of properties or the sum of stand-alone building PMLs. Dependence assumptions should be validated at the event and peril level, with particular care for compound events and restoration cost shocks.

Loss concentration also creates distributional consequences for insurance markets. A catastrophe may simultaneously reduce available reinsurance capacity, increase reconstruction costs, and raise renewal premiums in a region. The insurer's PML is an instrument for preparing financial resilience; it does not measure the total societal protection gap or determine how much risk a household can afford to retain. Translating portfolio tail numbers into public policy requires independent data on insurance take-up, hazard mitigation, affordability, and uninsured economic damage.

Decisions supported by a properly specified PML

For underwriting, PML identifies geographic and structural accumulations whose extreme losses may exceed an insurer's stated risk appetite. For catastrophe reinsurance, occurrence curves inform event limits and attachment while aggregate curves reveal repeat-event exposure and exhaustion. For capital planning, the annual net distribution supplies a basis for selected risk measures and is then reconciled with expected losses, capital eligibility, other risks, and the jurisdiction's solvency rules. For claims preparedness, a severe-event scenario can inform vendor capacity, adjuster mobilization, payment operations, and liquidity even when its precise annual probability is disputed.

For research, a useful PML report provides a reproducible lineage from underlying hazard assumptions and asset data through policy transformations and return-period computation. A separate sensitivity record shows how the result changes under credible alternatives. Reinsurer collectability and treaty wording belong in the financial lineage; a scientific model statement should not conceal legal uncertainty under statistical precision. These requirements follow directly from the construction of the underlying random variables, the mathematics of tail estimation, and the institutional distinctions found in the primary sources.

Original analytical findings

The analysis supports five substantive conclusions. First, PML is a conditional quantile or scenario estimate only after its loss and probability conventions have been specified; a bare maximum-loss label has insufficient measurement content. Second, the occurrence-versus-aggregate distinction changes the risk variable and therefore the relevant capital inference. Third, reinsurance's effect on tail capital must be assessed through contract-respecting event-year simulations rather than simple subtraction of nominal treaty limits from a gross PML. Fourth, statistical stability of simulated quantiles does not demonstrate validity of a hazard, vulnerability, or coverage model; empirical validation and structural sensitivity remain independent tasks. Fifth, a regulatory catastrophe charge and an enterprise solvency requirement operate at different financial and legal levels, so cross-jurisdictional comparison must preserve the full scope of each cited instrument.

These findings constitute original organization, comparison, and reasoned interpretation of existing knowledge. Their mathematical content follows established probability and risk-measure theory, while institutional details follow the official instruments cited. No new named doctrine, conceptual system, or unverified claim of priority is advanced.

Remaining research questions

The strongest opportunities for further empirical work are the independent verification of secondary-peril vulnerability, identification of spatial dependence under changing building stocks, quantification of event-definition effects on catastrophe treaties, comparison of extreme-loss estimates across vendor event catalogs, and study of net-loss liquidity under simultaneous catastrophe and reinsurer stress. Such studies require explicit data provenance, contract permissions, scenario timestamps, insurer-specific portfolios, and a defensible backtesting framework. The present documentary monograph establishes a common analytic framework for evaluating their results without pretending that the necessary datasets have been observed here.

Conclusion

Probable maximum loss is a powerful representation of property insurance exposure when its underlying loss variable, hazard and contract boundaries, return-period interpretation, and model are visible. A PML for the maximum annual event and a PML for aggregate annual catastrophe losses answer different questions. The gross insured, net retained, and whole-balance-sheet financial stages are likewise distinct. The mathematical treatment of Poisson event occurrence, annual aggregation, quantiles, expected shortfall, and tail extrapolation shows why a return period alone cannot carry the meaning of the measure.

The comparative evidence reinforces this conclusion. The 2026 Canadian earthquake standard employs an explicitly defined 500-year PML and dedicated financial-preparedness test. The IAIS Insurance Capital Standard calculates a net annual aggregate natural-catastrophe unexpected-loss charge at the 99.5th percentile, subject to its institutional scope. Solvency II calibrates a one-year, 99.5% risk measure of an insurer's basic own funds and requires wider internal-model validation. These approaches use related quantitative tools for different regulatory objects. Their meaning and requirements should be preserved at source.

The scientific and practical task is to maintain an auditable connection from modeled physical hazards to contractual insurer losses and onward to capital, liquidity, and governance decisions. A responsible PML presentation identifies exposure, peril, event or annual aggregation, probability, gross or net treatment, model version, sensitivity, nonmodeled risks, and legal context. The resulting measure can guide underwriting concentration, reinsurance design, capital adequacy and disaster preparedness with clear limits. The monograph's contribution lies in a source-grounded, reproducible integration of these established elements and a precise interpretation of their connections.

References

  • Acerbi, C., & Tasche, D. (2002). Expected shortfall: A natural coherent alternative to value at risk. Economic Notes, 31(2), 379–388. https://doi.org/10.1111/1468-0300.00091.

  • Actuarial Standards Board. (2011 revision). Actuarial Standard of Practice No. 39: Treatment of Catastrophe Losses in Property/Casualty Insurance Ratemaking. Official ASOP 39.

  • Actuarial Standards Board. (2016 revision). Actuarial Standard of Practice No. 23: Data Quality. Official ASOP 23.

  • Actuarial Standards Board. (2019 adoption; effective 2020). Actuarial Standard of Practice No. 56: Modeling. Official ASOP 56.

  • Actuarial Standards Board. (2021 revision; effective December 1, 2021). Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas). Official ASOP 38.

  • Artzner, P., Delbaen, F., Eber, J.-M., & Heath, D. (1999). Coherent measures of risk. Mathematical Finance, 9(3), 203–228. https://doi.org/10.1111/1467-9965.00068.

  • Coles, S. (2001). An Introduction to Statistical Modeling of Extreme Values. Springer Series in Statistics. Springer. https://doi.org/10.1007/978-1-4471-3675-0.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 101: Calculation of the Solvency Capital Requirement. EIOPA Article 101.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 104: Design of the Basic Solvency Capital Requirement. EIOPA Article 104.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 112: General Provisions for Approval of Full and Partial Internal Models. EIOPA Article 112.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 122: Calibration Standards. EIOPA Article 122.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 124: Validation Standards. EIOPA Article 124.

  • European Parliament and Council. (2009; current EIOPA rulebook compilation). Directive 2009/138/EC (Solvency II), Article 125: Documentation Standards. EIOPA Article 125.

  • International Association of Insurance Supervisors. (2024 adoption; online tool accessed November 2026). Insurance Capital Standard, Level 2, Sections 5.2.4.2–5.2.4.3, especially L2-182–L2-186. IAIS ICS online tool.

  • National Association of Insurance Commissioners. (2025 update). Catastrophe Models Property. NAIC catastrophe model overview.

  • Office of the Superintendent of Financial Institutions Canada. (2025 issuance; effective January 1, 2026). Minimum Capital Test—Guideline (2026), Section 4.5.1: Earthquake Risk Exposure. OSFI 2026 guideline.

  • Swiss Re Institute. (2026). Natural catastrophes in 2025: The persistent rise of wildfire and storm risk. Sigma 1/2026. Swiss Re sigma 1/2026.