RESEARCH DOSSIER
Actuarial Catastrophe Loss Modeling
Abstract. Actuarial catastrophe loss modeling estimates the probability distribution of insurance losses arising from low-frequency, high-severity events and from the accumulation of more frequent damaging events. It transforms physical event occurrence into insurance cash flows by combining stochastic event catalogs, spatial hazard fields, exposure inventories, vulnerability or fragility relationships, insurance contract terms, and risk-transfer structures. This monograph reconstructs that sequence as a transparent statistical and financial calculation. It specifies the random objects and conditioning assumptions at each stage; distinguishes event loss, annual aggregate loss, expected annual loss, occurrence exceedance probability, aggregate exceedance probability, and reinsurance recovery; and identifies the data and model errors capable of changing each output. The study integrates catastrophe-engine documentation, actuarial standards, observed hazard and claims datasets, engineering vulnerability methods, climate assessments, and international insurance supervision. A fully disclosed hypothetical portfolio demonstrates event-by-event indemnity, expected annual insured loss, an occurrence reinsurance layer, and analytically distinct occurrence and aggregate exceedance probabilities. Sensitivity and validation procedures show how a model should be challenged before its outputs influence pricing, capital, reinsurance, or public policy. The principal analytical conclusion is that the actuarial reliability of catastrophe losses depends on the integrity of the entire hazard-to-contract mapping, especially the preservation of spatial dependence and contractual aggregation, rather than the apparent precision of an isolated tail estimate. This conclusion is an original synthesis of established methods, not a claim of a new theory. International differences in hazard data, claims practices, contract law, regulatory calibration, and climate change require explicitly stated application boundaries.
Keywords: Catastrophe modeling; actuarial science; catastrophe event sets; hazard footprint; insurance exposure; vulnerability function; fragility; insured loss; financial module; expected annual loss; occurrence exceedance probability; aggregate exceedance probability; probable maximum loss; reinsurance; model validation; climate nonstationarity; capital adequacy; stochastic simulation; tail risk.
Introduction
A catastrophe loss model is an actuarial representation of a particular chain of uncertainty. An earthquake, tropical cyclone, flood, severe convective storm, wildfire, or another catastrophe has physical causes and a spatial footprint. Buildings, infrastructure, contents, people, and economic activities occupy that footprint. Exposure characteristics determine how the physical event becomes damage. Insurance contracts determine which portion of damage becomes the insurer's obligation. Reinsurance and other risk transfers determine how that obligation is divided. The final object is a distribution of insured financial consequences over a defined period and under a defined portfolio and contract configuration.
The NAIC description of property catastrophe models explicitly identifies modeled catastrophic events, engineering-derived damage, and insurance policy terms as successive elements of this process. The classical edited volume by Grossi and Kunreuther (2005), Catastrophe Modeling: A New Approach to Managing Risk, established a multidisciplinary context for catastrophe-model use in risk assessment and financial risk management. Contemporary open implementations make critical pieces of the calculation auditable: the Oasis modeling methodology describes footprints, vulnerability, loss uncertainty, and financial calculation, while the OpenQuake event-based risk methodology documents stochastic seismic events, ground-motion fields, damage ratios, and loss exceedance outputs.
Actuarial catastrophe work must explicitly define its output perspective. Economic loss describes damage and disruption from an event, some of which is uninsured. Ground-up loss represents the cost of damaged insured interests before financial terms. Insured loss reflects deductibles, limits, coverage, exclusions, and the policy structure. A direct insurer's gross loss, its ceded reinsurance loss, and its retained net loss are different cash-flow quantities. Expense and adjustment cost treatment, tax, reinstatement premium, salvage, subrogation, and time-value conventions create further distinctions. The Oasis financial module describes these perspectives and the ordering of coverage, location, site, policy, and layer calculations.
Observed catastrophe costs illustrate why definition is indispensable. Swiss Re Institute (2026) reported USD 107 billion of insured losses in 2025 in its March 2026 analysis. That retrospective estimate describes an observed global insured-loss aggregate, with estimation and reporting conventions of its source. It does not supply a probability distribution for an individual insurer, demonstrate the adequacy of a catastrophe model, or identify the financial loss under a particular reinsurance program. Macroeconomic totals, company-level model estimates, and premium or capital requirements operate at different analytical levels.
The objective here is to establish a fully traceable actuarial account of catastrophe loss modeling. The method is an integrative analysis of technical standards, original institutional documentation, risk mathematics, observed source catalogs, and an explicitly hypothetical numerical example. Each important empirical or institutional attribution is linked at the proposition it supports. The study develops interpretation and analytical conclusions within existing actuarial and statistical science. All amounts in the worked example are synthetic, and no example result is represented as a market estimate.
Research Object, Definitions, and Inferential Architecture
Let a portfolio at valuation time be represented by an exposure set , and let the annual catastrophe process be the random marked point process . Here is an event time, a peril or event family, and a vector containing location, magnitude, path, extent, and other event characteristics. The event produces a hazard field at spatial location . Asset has location , insured-interest value , construction and occupancy taxonomy , and a contingent physical damage ratio . A vulnerability function or distribution connects to , , and relevant mitigation characteristics.
The financial mapping then transforms damage-cost vectors into insured event loss:
A further reinsurance mapping converts the event sequence of insured losses into ceded and retained losses under treaty wording. A separate annual aggregation operator supplies . Time, currency, valuation date, and interest basis must all be fixed in interpreting and . For inflation-sensitive insured values, an event occurring two years after the exposure snapshot generally requires an explicit prospective value adjustment rather than implicit use of stale nominal figures.
The model's object is not a single number. It may estimate the joint distribution of event count, event intensity, location-specific loss, annual insured aggregate, net catastrophe cash flow, and payment timing, conditional on exposure , contract wording , model assumptions , and the valuation horizon . A notation such as preserves these dependencies. An apparent “one-in-100-year loss” without the portfolio, annual horizon, exceedance definition, and model edition is underdetermined.
The ASOP No. 38 revised standard addresses U.S. actuarial services involving the selection, use, review, or evaluation of catastrophe models. The more general ASOP No. 56 on modeling extends to model design, development, selection, modification, use, review, and evaluation within its stated scope. Their application depends on the actuary's professional assignment and governing actuarial standards; neither is a globally binding insurance statute. This methodological distinction matters for the governance of a model built by engineers, statisticians, commercial vendors, open-source communities, and actuaries working together.
A complete inferential architecture identifies: the stochastic source of events; the hazard transformation; the valuation and geolocation of affected interests; the damage mechanism; the contract-to-claim mapping; the event and annual loss aggregation; and the decision for which the model is used. Every stage should produce an inspectable intermediate object. When only the final loss curve is retained, a user cannot isolate whether an anomalous result originates in hazard generation, location geocoding, vulnerability selection, deductible ordering, or loss aggregation.
Catastrophe Event Sets and Stochastic Occurrence
An event catalog is a set of plausible catastrophe occurrences together with occurrence information. A historical catalog supplies observed records; a stochastic catalog supplies synthetic occurrences generated under explicit probabilistic assumptions. Each is useful in a different role. Historical data anchor physical plausibility and validation, while synthetic event sets explore rare combinations and tail outcomes that a short instrumental or insured-claims record cannot adequately sample. Catalog construction must account for changes in detection, station coverage, coding, landfall recording, and the treatment of events with no insured consequences.
For a stationary independent event family, a simple starting model is a Poisson process with annual rate . The number of events in an interval of length then satisfies . If each event carries an independent loss mark , the annual loss is a compound Poisson sum , with when the mean exists. This is an analytical benchmark, not an assumption that all real perils behave independently, occur homogeneously through time, or have unchanged intensity distributions. Seasonal tropical cyclones, earthquake aftershocks, clustered convective storms, cascading floods, and antecedent wildfire conditions can violate simple homogeneous specifications.
Empirical event catalogs require precise provenance. NOAA's International Best Track Archive for Climate Stewardship, known as IBTrACS, integrates tropical cyclone observations from multiple agencies, with documented interagency and early-record differences. The U.S. National Hurricane Center's tropical cyclone and best-track archive provides basin-specific post-analysis track and intensity records and reports. For earthquakes, the U.S. Geological Survey's ANSS Comprehensive Earthquake Catalog provides contributed earthquake source parameters and associated products. These sources describe observed hazards and their reconstructed event parameters; they are not ready-made insured-loss catalogs.
Frequency estimation separates physical occurrence probability from exposure geography. A tropical cyclone can occur in a basin without entering an insurer's insured territory. A severe earthquake can occur near an area with little covered exposure. An annual count of hazard occurrences cannot be substituted for a claim-producing event count without a hazard and exposure mapping. Equally, a count of insured catastrophic claims depends on insurance penetration, policy thresholds, claim reporting, and event definitions. A historical claims catalog has a strong selection mechanism: events that caused no covered claim are often missing from it.
Alternative stochastic constructions include nonhomogeneous Poisson intensities , renewal or clustered processes, mixtures conditioned on climate states, and physically simulated event catalogs. Such sophistication is justified when source evidence and validation can identify the required parameters. An event rate estimated from an incomplete historical series must carry a period-of-record statement, missing-event analysis, and uncertainty interval. A tail sample spanning a handful of extreme years cannot silently establish a precise 1,000-year rate.
The relation between rare-event extrapolation and observation is examined rigorously in Embrechts, Klüppelberg, and Mikosch (1997), Modelling Extremal Events for Insurance and Finance. Extreme-value techniques can support tail inference under defensible limit and sampling assumptions, yet they cannot recover absent physical mechanisms, a changed climate regime, or missing contractual exposure. The data-generating process must always be specified alongside the fitted tail.
Hazard Footprints and Physical Event Characterization
A catastrophe event is converted into an asset-relevant hazard through a footprint: a spatially and sometimes temporally varying field of damaging intensity. Earthquake intensity may be expressed by spectral acceleration, peak ground acceleration, or other ground-motion measures. Tropical cyclone intensity for wind loss may require site-specific gust speed and duration, adjusted for local terrain and surface roughness. Flood modeling may require inundation depth, velocity, contamination, and duration; wildfire may require fire intensity and ember exposure. Each intensity measure must match the vulnerability function's definition and unit system.
The USGS ShakeMap product produces maps of ground motion and shaking intensity for significant earthquakes, including archives and scenario tools. ShakeMap supplies a hazard characterization; it does not by itself encode an insurance coverage decision. It is also important to separate a near-real-time instrumental footprint from a later revised reconstruction. Data revision can materially alter retrospective catastrophe loss attribution, and different ShakeMap versions should not be combined without a common observation and metadata policy.
Oasis's documented footprint methodology explains how local topography, surface roughness, soil, and geological conditions can affect surface hazard. The scientific problem is spatial interpolation and physically justified transformation under incomplete observations. At coarse resolution, a hazard cell can contain insured buildings with very different elevation, construction, and shielding. A physically credible regional footprint may still be poorly aligned with the insurer's exposure resolution.
Where flood damage is investigated, Huizinga, de Moel, and Szewczyk (2017) offer a global flood depth-damage database designed for comparative applications. Its available curves and maximum damage values supply a documented starting point; they require adaptation to local building stock, construction economics, and the particular insured interest. Extrapolating a continental-average loss ratio directly to an individual commercial property creates a mismatch of spatial and economic scales.
The spatial joint distribution matters more than a collection of marginal intensity estimates. For two nearby insured assets, the same underlying storm track affects both losses. If model simulation samples their physical intensity independently, it can understate or distort portfolio-level tail dependence. If a footprint already contains common physical uncertainty and the user independently adds another correlated shock, it can double-count that dependence. Hazard uncertainty therefore should be decomposed into event-parameter uncertainty, local field variability, and observation or interpolation error, with explicit information about which terms remain dependent across locations.
Multiple hazard components from the same event require coherent event identities. A tropical cyclone can simultaneously produce wind, storm surge, river flood, and rain-driven urban flooding. Summing stand-alone peril losses may double-count the same damaged building or violate coverage causation and hours clauses. A compound-peril model should either jointly simulate physical mechanisms and damage states or carefully reconcile combined damage. The scientific requirement is consistency of shared event conditions and insured interests across all peril modules.
Exposure Inventories and Spatial Accumulation
An exposure inventory determines where insured interests exist, what they are, how much they are worth, and how their characteristics influence damage. Typical fields include geocoded location, coverage value, occupancy, structural system, building age, elevation, number of stories, roof type, foundation, contents, business-interruption values, policy and account IDs, and the effective dates of insurance and mitigation. The appropriate granularity is a function of the peril: elevation can dominate local flood damage; roof construction can dominate wind vulnerability; soil and construction vintage can determine earthquake response.
The Oasis Open Exposure Data standard is an interoperability-oriented response to the problem of moving exposure and policy data between catastrophe model providers and analytical tools. Standardization makes fields interpretable and facilitates reproducible exchanges; it does not guarantee the underlying insured locations or values are accurate. A coordinate derived from a postal centroid and a verified building address can share a technically valid schema while having radically different flood or wildfire accuracy.
The insured value should be linked to the insured interest. Building replacement cost, contents limit, business-interruption declared value, and total insured value are different objects. A declared policy limit is not always a reliable replacement-cost estimate. The risk model should distinguish an economic valuation dataset used to estimate physical damage from the contractual limit that caps a claim. An insured property reconstructed for validation at current replacement prices also requires explicit deflation or inflation adjustments when compared with a paid historical claim.
Exposure quality audits should examine missing fields, duplicate locations, incorrect geocodes, multi-building accounts collapsed into one location, implausible values per unit area, and unlocated policies. Missingness is frequently nonrandom: older rural housing, informal buildings, or small insured businesses may be underrepresented precisely where vulnerability is uncertain. Straightforward imputation with a single portfolio mean can conceal geographic concentrations. Better practice is to preserve missing-data flags, use defensible imputation classes, and quantify how alternative classifications affect insured aggregate and tail losses.
Portfolio accumulation is driven by shared hazard footprints. A insurer with ten similar buildings within one floodplain may face more catastrophe concentration than one with the same total sum insured distributed across uncorrelated flood basins. The difference cannot be inferred from total insured value alone. Exposure growth and urban development can also increase insured loss without a change in annual physical event frequency. Swiss Re Institute's 2025 natural catastrophe study explicitly discusses expanding asset exposure and adaptation as determinants of observed loss levels.
An exposure snapshot must carry a transaction and valuation date. For annual aggregate risk, one may either model a fixed current portfolio over a one-year period or simulate scheduled changes in business volume and coverage. Mixing a current-day total insured value with loss observations from a materially different portfolio population yields a distorted claims frequency and severity relationship. Territorial exclusions, policy inception and expiration, co-insurance, and covered-interest percentages also belong in the inventory-to-financial mapping.
Vulnerability Functions and Damage Uncertainty
A vulnerability function maps physical intensity and asset characteristics into a distribution of damage or loss. Write for the fraction of an insured interest's replacement value physically damaged in a given event, and let be a peril-appropriate intensity metric. A model may specify a conditional mean , a variance, an entire probability mass function, or a distribution of engineering damage states followed by a separate consequence function. These representations are related but not interchangeable. A fragility curve describes the probability of reaching or exceeding a damage state; a vulnerability curve estimates a fractional damage or economic loss response, possibly derived from fragility states and consequence values.
The OpenQuake risk calculator couples ground-motion intensity with a vulnerability relationship and samples loss ratios with configurable correlation assumptions. The Oasis sampling specification represents both intensity uncertainty and damage conditional on intensity through discrete distributions, then samples an effective damage distribution. These implementations illustrate a central identification issue: the mean vulnerability response, hazard uncertainty, and secondary damage variation must be separately described before they are combined.
For example, two nominally identical warehouses could have the same modeled mean wind-damage ratio but different realized losses because of installation quality, internal contents, maintenance, door failure, and unobserved protective measures. Assigning each warehouse exactly its conditional mean damage creates artificial determinism. Independently sampling each warehouse's damage can create the opposite distortion if a shared construction defect, identical roofing material, or common flood depth causes a correlated failure. Damage correlation therefore can be driven by both observed common hazard and residual common factors after conditioning on hazard.
Vulnerability calibration commonly combines damage surveys, engineering experiments, post-event reconnaissance, claim-derived loss ratios, and expert judgment. Each source has a different observation process. Claims represent covered insured interests and the policy adjustment process. Engineering damage surveys may describe physical damage across both insured and uninsured structures. Laboratory experiments can isolate a physical component under controlled conditions but cannot reproduce all field installation practices. A calibration dataset should record the loss denominator, hazard intensity scale, construction taxonomy, and treatment of destroyed or unrepaired assets.
The JRC depth-damage methodology exemplifies the need to distinguish fractional damage functions from maximum damage values expressed in monetary units. If a depth-damage curve predicts 30% damage to a building class, the gross physical loss estimate still depends on what cost base is multiplied by 0.30. National construction-cost conversion, contents assumptions, and business interruption may substantially alter the resulting total. Damage ratios should be bounded and tested against the actual insured interest: an engineering percentage of structural repair cost cannot automatically be applied to the full building-plus-inventory sum insured.
Vulnerability functions introduce epistemic uncertainty when fitted outside their supporting hazard range, region, building code, or construction class. Parametric smoothing and monotonicity constraints can provide plausible behavior, but no mathematical constraint can replace physical evidence for new construction materials or emerging peril mechanisms. A sound study therefore reports both fitted response and sensitivity to plausible competing curves. Where calibration data are thin, loss distribution outputs should be reported with a model-range or scenario qualifier rather than as a uniquely identified empirical forecast.
Insurance Contracts and Financial Terms
The passage from physical damage to indemnity is an actuarial and legal transformation. Let be gross physical damage cost for asset in event . For a deliberately simple policy with a monetary deductible and a maximum insurer payment , the indemnity is
This equation is only a stylized policy form: it treats as a post-deductible payment cap. If the contract instead defines a total amount of covered loss before deductible, or applies a percentage deductible, occurrence aggregate, coinsurance, franchise, sublimit, waiting period, special catastrophe deductible, or an annual aggregate, the exact formula changes. Every financial computation must be traced to the relevant policy wording and the ordering of terms.
The Oasis financial-module documentation enumerates coverage, site, policy, condition, and layer processing and distinguishes ground-up, insured, and net views. These levels explain why a portfolio-level deductible cannot simply be subtracted from each asset independently. Applying a coverage limit at the wrong level may materially change both expected annual loss and the severity of extreme events. Conversely, assuming that all reported economic damage is insured can overstate liabilities in areas where floods, earthquakes, or business interruption are excluded or subject to specific sublimits.
A financial module must determine coverage causation as well as arithmetic. A cyclone may create wind and water damage whose contract treatment differs depending on cause-of-loss provisions, flood exclusions, anti-concurrent-causation wording, endorsements, and jurisdiction. A catastrophe “event” in an engineering catalog may not be the same event under a policy's hours or occurrence definition. The appropriate claims unit also depends on whether deductibles are per location, per claim, per peril, or per event. These questions are inputs to the stochastic financial model because they change the distribution of paid claims.
Business interruption introduces distinct timing and economic relationships. Property damage may trigger business income coverage, but the measured loss can depend on waiting periods, period of restoration, continuing expenses, extra expense, supply-chain dependencies, and policy definitions. Demand surge after a widespread catastrophe can elevate reconstruction prices; claims development and litigation can change ultimate loss values after an initial event estimate. Model applications that require liquidity or unpaid-claims projections should use a settlement and payment-timing module rather than equating ultimate insured event loss with immediately payable cash.
Financial terms also determine the relation between gross and net loss. Direct insurance recovery is computed under insurance wording before it is assigned to reinsurance treaties, unless the specific risk-transfer architecture dictates another recognized basis. Treaty recoverability may depend on attachment, coverage scope, exclusions, occurrence definition, annual aggregate restrictions, reinstatements, credit risk, or collateral. A simple event-wise subtraction of a constant “reinsurance percentage” is rarely sufficient for complex towers or stop-loss arrangements.
Event Loss Tables and Portfolio Aggregation
An event loss table records event identifiers and modeled portfolio financial loss measures, often with their occurrence rates, weights, periods, and optional loss distributions. Its primary role is to retain the connection between each underlying catastrophe and its insured financial consequence. For each event , write if policies are individually separable and no portfolio-level contractual aggregation applies. Otherwise use and maintain the original insured-interest identifiers so higher-level financial terms can be applied correctly.
The OpenQuake event-based calculation workflow can produce event loss tables and loss exceedance curves and supports distinct annual occurrence and aggregate outputs. It is therefore useful to keep the event-level loss view and annual-aggregate view separate in data storage and QA. A catastrophe model that preserves only the sorted event loss distribution loses important information about co-occurrence, annual clusters, and treaty exhaustion unless the necessary dependence structure is stored elsewhere.
There are at least four aggregation choices. Asset aggregation sums damage across insured interests for the same physical event. Policy aggregation applies contractual provisions to the relevant covered set. Event aggregation combines losses from different locations and covered causes under the legally relevant occurrence. Annual aggregation combines all insured events assigned to the calendar or contract year. A fifth operation, reinsurance aggregation, applies treaty terms across covered events and annual periods. The order is substantive: arithmetic operators such as a deductible and summation generally do not commute.
For instance, for two asset losses of 0.8 each and an event-wide deductible of 1.0, the event indemnity is . Subtracting 1.0 separately from each asset gives zero. The latter computation violates the specified contract, although both results use superficially plausible “deductible” arithmetic. This small example demonstrates why model validation must compare policy-level and event-level transactions, not merely portfolio totals.
An event table may store one mean loss per event, many simulated conditional losses for the same event, or a conditional distribution. These are distinct statistical objects. Replacing conditional event loss distributions by their means generally suppresses tail variation. A simulation approach must preserve joint conditional event losses across exposures; an event-wise conditional mean table may remain useful for expected annual loss but can be insufficient for extreme-loss quantiles and reinsurance layers with nonlinear recovery functions.
Exceedance Probability, Return Periods, and Tail Metrics
An exceedance curve expresses the probability that a selected loss metric exceeds a threshold. Occurrence exceedance probability (OEP) is the probability that the largest qualifying loss in a year exceeds , whereas aggregate exceedance probability (AEP) is the probability that the sum of qualifying losses in the year exceeds . Let , with if no event occurs, and . Then
For nonnegative event losses on an identical annual and financial basis, pathwise, so . This ordering can fail as a claimed comparison if one curve is net of annual aggregate reinsurance while the other is gross, or if the curves use different exposure snapshots or contractual periods. Any published comparison must state the common basis.
For independent Poisson events of type , with frequency and loss mark , the annual occurrence probability is
This derivation uses Poisson thinning: the number of events exceeding is Poisson with mean . The simple expression breaks down for certain clustered and non-Poisson event processes; the compound annual simulation remains the general operational alternative. The OpenQuake distinctions among EP, OEP, and AEP provide a corresponding software interpretation.
For a strictly defined annual exceedance probability , the reciprocal is the conventional annual exceedance return period. A return period is a statistical description of the modeled exceedance frequency under the assumed process, not a schedule on which losses must occur. A 1% annual exceedance probability does not mean that a particular event arrives once every century or that an insured portfolio faces exactly one such loss during a hundred-year horizon. Under independent stationary years, the chance of at least one annual exceedance in 100 years is , approximately 63.4%. Dependence or nonstationarity across years alters this calculation.
Expected annual loss (EAL or AAL) is for the stipulated annual loss measure. For a nonnegative annual aggregate , the layer-cake identity is
provided the expectation exists. This mathematical relationship is important for numerical QA: an AEP curve estimated with a sufficiently resolved threshold grid should integrate to an EAL consistent with direct sample averaging. It does not authorize integrating OEP as if it were annual aggregate expected loss.
“Probable maximum loss” (PML) is used inconsistently unless a return period, loss basis, time horizon, and definition are supplied. An asset or portfolio PML at a stated annual exceedance probability is a loss quantile derived from a particular curve, not a physical upper bound. The OpenQuake loss-curve guidance makes explicit that finite simulated event durations restrict the supported range of return periods and that inadequate simulation may require an undefined output rather than extrapolation. Tail quantities are especially sensitive to catalog size, event weighting, correlation, and parameter uncertainty.
Tail estimation can be supplemented by generalized extreme value or peaks-over-threshold methods where their convergence, threshold choice, independence or declustering, and data sufficiency assumptions are defensible. The reference treatment by Embrechts, Klüppelberg, and Mikosch (1997), statistical methods for extremal events, supplies the underlying probability context. Extrapolation methods should be tested against the physical event model and the actual insured-loss sampling design, especially when modeling a changed portfolio or a changing hazard regime.
Simulation Design and Statistical Accuracy
Simulation converts stochastic catalog, hazard, vulnerability, contract, and treaty assumptions into a distribution of annual insured and retained losses. A practical simulation generates independent or appropriately dependent annual histories; draws events and marks within each history; computes hazard at every exposed asset; samples conditional damage with coherent spatial dependence; applies policy terms; computes reinsurance recoveries; and stores annual outcomes. It is essential to carry the same hazard realization through all exposed assets within an event. Resampling the event independently at each location destroys the accumulation structure.
The Oasis effective-damage and Monte Carlo specification shows how hazard and vulnerability uncertainty can enter through probability distributions and sampling. The OpenQuake event-risk guidance describes simulation periods, asset correlation controls, and the computation of output curves. The technical freedom to change sample count, correlation, random seed, and event weighting implies an editorial obligation to document these choices so that independent analysts can reproduce the result.
For independent simulated years with annual loss outcomes , the sample mean estimates EAL. Its familiar standard error estimate is , but heavy tails, clustered years, importance sampling, common parameter uncertainty, and rare-event conditioning require care. If observations were generated with deliberately unequal importance weights, direct unweighted averaging is invalid. If all simulated years share an estimated model parameter, increasing annual simulations reduces conditional Monte Carlo error while leaving epistemic parameter uncertainty unchanged.
For a threshold , an empirical annual AEP estimator based on independent years is . If only ten simulated years exceed an extreme threshold, even a very large nominal annual run may yield a relatively unstable estimate. The relevant evidence for tail precision includes the number of effective exceedances, the effective sampling horizon, intersample dependence, and sensitivity to event parameters. Confidence intervals should reflect the actual sample design and distinguish stochastic error from fitted-parameter uncertainty.
Model execution should record the random seed, model versions, input hashes, event-set duration, simulated years, peril definitions, currency basis, deductible ordering, coverage scope, damage sampling, correlation parameters, and result calculation. It should also include deterministic unit tests for each financial calculation. A change in a vulnerability table or an exposure column must be traceable to altered event losses and ultimately to changed EAL and exceedance outputs. Such provenance is the basis of defensible actuarial reproducibility.
Reinsurance Recoveries and Net Loss
Catastrophe reinsurance changes the distribution of losses carried by a direct insurer. A stylized occurrence excess-of-loss layer with attachment and monetary layer limit pays
The retained event amount is before any further recoveries or costs. The precise insured loss to which the layer attaches must be specified: subject ultimate net loss, gross paid claim, ultimate incurred loss, or another contract-defined measure. An insurer may have quota share beneath occurrence excess layers, several ordered excess layers, annual aggregate protections, event caps, exclusions, and aggregate deductibles. Recoveries cannot generally be computed from one portfolio-wide annual gross percentage.
The Oasis reinsurance financial-module implementation explicitly separates underlying insured loss and reinsurance calculations, including attachment and participation. Complex treaty structures require event identity, covered exposure, financial ordering, and annual exhaustion logic to remain available. If an occurrence cover is exhausted by an event and can be reinstated only against payment of premium, the annual loss and expense distribution depends on subsequent events and reinstatement conditions. A sample-by-sample event table without contract-year chronology may therefore understate the insurer's risk.
Reinsurance can reduce an insurer's net volatility while creating counterparty-credit exposure and operational uncertainty in collection. Catastrophe modeling used for solvency planning should link expected recoveries to counterparty and liquidity stress, rather than treating every projected cession as contemporaneous cash. A treaty limit measured in a different currency also needs an explicit exchange-rate convention and currency-risk treatment. Catastrophe loss models can supply underwriting event distributions, but total company solvency requires integration with investment, liquidity, operational, reserving, and counterparty risk.
Event definitions are especially important for catastrophe reinsurance. A single broad storm system could create several physically distinct loss episodes; contractual hours clauses may allow a cedant to choose certain time windows or impose separate event definitions for different perils. Such choice may require optimization under contract rules and should never be guessed from the engineering peril label. If policy losses and treaty terms are incompatible in coverage or time, modeled recoveries may be legally unavailable even when a simple mathematical attachment formula shows payment.
Empirical Calibration and Model Validation
Validation evaluates whether a model is credible for a clearly specified decision and data domain. It is not equivalent to checking that code runs, that a historical event is recognized, or that modeled aggregate loss equals one reported industry total. Catastrophe validation needs multiple forms of independent comparison, preferably with preserved uncertainty and an explicit test design. Relevant targets include event occurrence, spatial hazard, vulnerability damage fractions, conditional loss distributions, reported insured claims, contract financial outcomes, annual aggregate loss behavior, and event-to-event dependence.
For seismic loss, USGS historical ShakeMap products allow an observed or reconstructed intensity field to be compared with model hazard outputs. The USGS ComCat data architecture retains earthquake sources and associated product versions. For tropical cyclones, NOAA's IBTrACS documentation describes multiple agency contributions and caveats. These public records can validate portions of the event or hazard process without revealing any insurer's actual contract-adjusted claim amounts.
Claims validation requires an exposure denominator and an explicit payment definition. The National Flood Insurance Program's public data limitations and disclosure rules indicate, for example, that public policy records are incomplete for earlier periods, that released coordinates are intentionally coarse, and that the claims records do not generally supply payment transaction dates. An analyst reconstructing a particular historical flood cannot infer accurate building-level hazard from coordinates rounded for privacy or reconstruct payment timing from date-of-loss records. Such constraints are substantive scientific limitations, not minor clerical issues.
A validation program should compare ground-up modeled costs with independent engineering observations where available, then compare insured modeled outcomes with contract-consistent incurred or paid losses. This two-stage approach helps identify whether discrepancies arise in hazard/vulnerability or in policy adjustment and claims development. A storm may be modeled correctly in physical terms but incorrectly in insurance terms because the actual portfolio contained different deductibles, policy endorsements, reinstatements, or claims inflation than the calibration dataset assumed.
Historical replay should reconstruct the insured portfolio as of the event date, use the then-current policy wording, adjust cost values consistently, and distinguish modeled ultimate loss from the claims development stage observed. Repeating today's exposure inventory against a historical footprint estimates a hypothetical counterfactual present-day loss, which is useful for stress analysis but is not a backtest of the historical insurer's loss. Clear labels for actual event, historical-exposure reconstruction, and present-day exposure replay prevent a common validation ambiguity.
Other independent checks include geographic holdouts, peril-specific holdouts, period-separated validation, alternative vulnerability calibrations, simulated exceedance counts, negative controls for unaffected assets, and cross-model comparisons. ASOP No. 56's model validation guidance requires model users to consider fitness for purpose, output validation, appropriate use, and communication within its scope. ASOP No. 38's catastrophe-specific guidance further addresses the special challenges of catastrophe models used in actuarial services.
One critical validation distinction is between a calibration success and an independent confirmation. A vulnerability curve estimated from a set of claims and then evaluated against the same claims has demonstrated in-sample fit. It has not established how the model will perform in unseen regions, changed construction types, new policy conditions, or extreme events outside observed ranges. A model approval should record the blind or out-of-sample evidence available, sensitivity results, and the domains for which validation remains limited.
Dependence, Compound Perils, and Climate Nonstationarity
Dependence enters catastrophe losses at several distinct scales. Geographic dependence arises because a single event damages many nearby assets. Within-event dependence also arises through shared construction characteristics, mitigation failures, and service interruptions. Across-event dependence arises when events cluster seasonally, when antecedent weather conditions affect subsequent hazards, or when repair resources and financial limits are depleted by prior losses. Across-peril dependence is particularly important when one meteorological system creates wind, surge, fluvial flood, and business-interruption losses. Each dependence structure can affect portfolio accumulation and net retained loss even if single-asset expected damage remains unchanged.
The IPCC assessment of compound events describes multiple compound hazard processes, including interactions among storm surge, extreme rainfall, and river flow. This physical assessment supports the need to examine coupled hazards at appropriate regional resolution. It does not provide a universal insured-loss dependence coefficient. That coefficient depends additionally on portfolio location, construction, insurance policy wording, and how insured interests overlap across the relevant hazard modules.
One approach is to simulate the common generating event and its physical fields jointly, then compute correlated damage conditional on that common event. An alternative uses a statistical copula or other dependence structure where a joint physical simulator is unavailable. The latter approach requires evidence for tail dependence, not merely observed correlation under ordinary conditions. A correlation estimate from a short insurance claims series may be driven by changes in exposure and insurer participation rather than by a stable physical dependence mechanism. Stress scenarios should therefore vary both the marginal hazard intensity and the assumptions about simultaneous or sequential events.
Climate nonstationarity complicates the standard “annual return period” vocabulary. Physical hazard intensity and frequency can change with warming, land-use transformation, urban drainage, vegetation, and protective infrastructure. The IPCC AR6 assessment of extremes reports robust evidence of increases in heavy precipitation across many regions and distinguishes more regionally uncertain patterns in river flooding and tropical cyclone occurrence. An insurer may legitimately incorporate a changing hazard into its prospective annual analysis, but the method must identify the reference climate, the modeled calendar year, and the scientific basis for each peril-specific adjustment.
The logical distinction is between a stationary risk measure , evaluated under a fixed parameter vector , and a time-indexed probability , in which hazard, exposure, vulnerability, or contract terms evolve. A reported return period has a year-specific interpretation. Over multiple future years with independent but varying exceedance probabilities , the probability of at least one exceedance is . Independence itself may be questionable if climate states or catastrophe clusters persist over time. A fixed “100-year storm” associated with a historical climate should not be used without qualification to describe future insured financial loss.
Climate-related changes in insured loss must be decomposed into hazard, exposure, vulnerability, and financial factors. An observed increase in paid catastrophe claims can coincide with growth in insured values, changing coverage penetration, rebuilding costs, litigation practices, and hazard severity. Swiss Re Institute (2026) emphasizes the role of expanding exposure and adaptation in catastrophe-loss outcomes. This illustrates why attribution of a financial trend requires more than a simple regression of historical nominal insurance payments on year.
Secondary uncertainty, epistemic uncertainty, and changing risk are separate analytical categories. Secondary uncertainty describes event-to-event variation around modeled conditional damage, often sampled within a fixed model. Epistemic uncertainty describes incomplete knowledge of model structure, parameters, data quality, and hazard mechanisms. Nonstationarity describes a time-dependent underlying distribution. A larger Monte Carlo sample can reduce stochastic numerical error while leaving both epistemic uncertainty and structural changes essentially untouched. A mature model risk report labels these contributions separately, estimates what can be estimated, and explains which risks are represented only by qualitative sensitivity or stress scenarios.
Pricing, Reserving, Capital, and International Supervision
Catastrophe modeling supports several actuarial decisions, each with a different target estimand. For prospective insurance pricing, modeled expected catastrophe loss must be adjusted to the insurance product's covered risks, exposure period, claims-expense basis, and cost environment. The ASOP No. 39 catastrophe ratemaking guidance addresses catastrophe loss provisions in applicable U.S. property/casualty ratemaking work. A catastrophe provision cannot be inferred solely from a tail quantile because a premium's expected-loss component concerns the distribution's mean, while risk loading, expenses, cost of capital, and competitive or regulatory considerations require additional analysis.
For reserving, a catastrophe model may help evaluate latent future claims development and unpaid event losses, but an expected annual future catastrophe model is not itself an estimate of claims already incurred. A completed insured event can generate paid and unpaid claims whose ultimate severity, litigation, business interruption, and reporting development must be estimated separately. The reserving problem conditions on an occurred event and observed claims; the prospective pricing problem conditions on potential future events and current coverage. Using an unconditional future-loss distribution as a substitute for case or incurred-but-not-reported reserves misidentifies the liability.
For economic capital or risk appetite, the relevant object is an annual distribution of net loss under a specified balance-sheet horizon and solvency metric. An occurrence PML can inform concentration and treaty attachment but does not fully represent aggregate catastrophe accumulation. The capital consequence also depends on non-catastrophe insurance risk, market losses, asset liquidation needs, counterparty defaults, and potential dependence among these effects. Thus a single 1-in-200 event loss cannot be equated automatically with required company regulatory capital.
International regulatory systems use different quantitative conventions. The European insurance framework's Solvency II natural catastrophe sub-module identifies windstorm, earthquake, flood, hail, and subsidence components within its prescribed standard-formula architecture. The EIOPA natural catastrophe reassessment, published in January 2025, documents review of standard-formula calibration against changed scientific evidence and catastrophe experience. The listed regulatory sub-modules and their aggregation rules are prescribed jurisdictional capital tools. They should not be presented as the universal mathematical decomposition of every insurer's underlying catastrophe losses, nor as direct substitutes for a company's detailed internal model.
In the United States, the NAIC's property catastrophe model overview discusses regulatory model review and catastrophe-risk risk-based-capital activities, including the role of hurricane, earthquake, and informational peril collections. Such reporting must be interpreted using the applicable reporting year and jurisdiction. The international supervisory perspective in the IAIS Insurance Core Principles, ICP 16 establishes enterprise risk management and own risk and solvency assessment expectations in supervisory frameworks. The IAIS material is a global supervisory standard intended for jurisdictional implementation; it is distinct from direct locally binding law and from the technical specifications of a commercial catastrophe model.
Reinsurance purchasing and insurance-linked securities introduce further decision criteria. Selecting an attachment point depends on the insurer's risk appetite, event and annual loss distribution, reinsurer credit and price, reinstatement provisions, and liquidity. A layer's expected ceded loss is one component of its risk transfer economics. It does not alone establish its market premium, collateral demand, expenses, capital benefit, or legal enforceability. Models that evaluate a catastrophe bond, indemnity treaty, or parametric protection should measure the basis risk between the instrument's trigger and the modeled insured liabilities.
A sound managerial interpretation preserves at least three perspectives in parallel: expected annual insured loss for cost and profitability; occurrence and aggregate tail distributions for accumulation and capital; and cash-flow timing or recoverability for resilience. The same model may produce all three if its temporal, contractual, and financial modules are consistent. Where only an event-loss table is available, it is necessary to state which of these estimands can legitimately be recovered and which require additional annual history or contract information.
Reproducible Analytical Illustration
Portfolio and assumptions
Consider a hypothetical portfolio containing three asset coverages, all values in millions of U.S. dollars. Asset A has replacement value 10, deductible 0.2, and maximum insurer payment 2.5 per event. Asset B has value 8, deductible 0.2, and payment cap 1.5. Asset C has value 6, deductible 0.2, and payment cap 1.8. We use a common event-specific intensity-to-damage table and the stylized payment formula defined earlier. All three coverages are assumed in force for the full year. There are no additional exclusions, coinsurance provisions, demand-surge loads, business-interruption payments, claims expenses, inflation adjustments, or disputed coverage interpretations.
Four independent event types arrive according to stationary Poisson processes, with annual occurrence rates , , , and . Each occurrence of a specified event type is assumed to have deterministic damage ratios in this initial demonstration. This simplified construction intentionally excludes secondary uncertainty so that every result can be checked by arithmetic; a production catastrophe model would ordinarily consider stochastic event intensity and conditional vulnerability variation. The rates are invented for teaching and have no calibration to a named territory or peril.
| Synthetic event type | Annual rate | A damage ratio | B damage ratio | C damage ratio |
|---|---|---|---|---|
| E1 | 0.20 | 0.30 | 0.20 | 0.05 |
| E2 | 0.10 | 0.05 | 0.25 | 0.40 |
| E3 | 0.45 | 0.01 | 0.05 | 0.08 |
| E4 | 0.05 | 0.12 | 0.10 | 0.20 |
The physical damage cost of event E1 is . Its insured event loss is . The other rows follow from identical financial terms and expose the role of both deductibles and limits.
| Event | Ground-up A | Ground-up B | Ground-up C | Ground-up total | Insured A | Insured B | Insured C | Insured event total |
|---|---|---|---|---|---|---|---|---|
| E1 | 3.00 | 1.60 | 0.30 | 4.90 | 2.50 | 1.40 | 0.10 | 4.00 |
| E2 | 0.50 | 2.00 | 2.40 | 4.90 | 0.30 | 1.50 | 1.80 | 3.60 |
| E3 | 0.10 | 0.40 | 0.48 | 0.98 | 0.00 | 0.20 | 0.28 | 0.48 |
| E4 | 1.20 | 0.80 | 1.20 | 3.20 | 1.00 | 0.60 | 1.00 | 2.60 |
Expected annual loss and occurrence reinsurance
With independent Poisson arrivals, expected annual ground-up loss equals the sum of event-rate-weighted ground-up values:
Expected annual insured loss, before reinsurance, is
Suppose an unlimited-reinstatement occurrence reinsurance layer pays the part of each insured event loss between 1 and 4 million: attachment , limit , no coverage mismatch or credit loss, and no reinstatement premium is included in the losses shown. The layer yields ceded event losses of 3.00, 2.60, 0.00, and 1.60 for E1–E4, respectively. Its expected annual ceded loss is
The expected annual retained loss is million. These are expected modeled claim amounts. They are not premium indications; the reinsurance price, operational costs, and cost of capital would be separate. If reinstatements were limited, event-wise ceded annual expectation would require treaty-year accounting, and 0.940 would no longer automatically be the ceded annual mean.
| Event | Insured loss | Ceded occurrence layer | Retained event loss |
|---|---|---|---|
| E1 | 4.00 | 3.00 | 1.00 |
| E2 | 3.60 | 2.60 | 1.00 |
| E3 | 0.48 | 0.00 | 0.48 |
| E4 | 2.60 | 1.60 | 1.00 |
OEP and AEP at a common threshold
For insured loss threshold million, only E1 and E2 individually exceed it. The sum of their annual Poisson rates is 0.30, so
The annual OEP is about 25.918%, with a reciprocal return period of approximately 3.858 years. To compute AEP precisely under this toy distribution, take independently. The two high-loss types E1 and E2 must both be absent for aggregate loss to stay at or below 3.00. Conditional on that absence, two occurrences of E4 exceed the threshold, one E4 plus one E3 exceeds it (2.60+0.48=3.08), and seven E3 occurrences exceed it (7×0.48=3.36). Thus the only annual nonexceedance configurations are or , together with :
Consequently
or approximately 27.285%, with reciprocal return period 3.665 years. The 1.3664 percentage-point gap between AEP and OEP results purely from the possibility of several insured events accumulating during one year. It would be invisible if one treated the distribution of individual event losses as the distribution of annual aggregate loss. A Monte Carlo reconstruction of the same independent Poisson example should converge to these analytical values within the sampling error predicted by the number of simulated years.
Sensitivity and nonlinear contract transmission
Now increase every physical damage ratio by 15%, holding values, event frequencies, and contractual caps constant. The insured event totals become E1 4.145, E2 3.675, E3 0.612, and E4 3.080 million. Expected annual insured loss becomes 1.6259 million, an increase of about 7.96% from the original 1.506 million. The increment is smaller than the 15% vulnerability shift because some damages are below deductibles while other recoveries reach caps. The hypothetical occurrence layer's expected annual cession becomes 0.9715 million and net expected annual loss becomes 0.6544 million. This illustrates that proportional physical-damage stress can create non-proportional insurance and reinsurance effects.
Alternatively increase each annual event rate by 25%, preserving the original event losses. Expected annual insured loss rises proportionally to 1.8825 million, as follows from Poisson linearity. But the annual OEP at 3.00 becomes , a 31.271% probability rather than 25.918%. Exceedance probabilities respond nonlinearly to event rates even when EAL responds linearly. When multiple event types share a climate driver, scaling all frequencies independently might also misstate joint annual behavior; the present calculation makes the independence assumption explicit.
What the example proves and what it does not
The example establishes arithmetic consistency among asset damage, insurance indemnity, event losses, EAL, an occurrence reinsurance layer, and annual OEP/AEP. It provides exact benchmarks for a small implementation. It does not establish tail credibility for rare return periods; four deterministic event classes do not span the range of catastrophe severity, dependence, exposure error, or contract structures encountered in a real catastrophe portfolio. In particular, the expected loss, exceedance rates, and reinsurance recoveries are conditional on synthetic event frequencies and damage ratios. A scientific actuarial assessment of a real territory would have to calibrate and validate every material input before such figures could support a decision.
Data Governance, Model Risk, and Reproducibility
Model governance begins with an inventory of source data, transformations, and ownership. Every major input dataset needs a title, provider, access date or version, geographic and temporal coverage, relevant license, transformation history, and the claims it can support. For example, NOAA IBTrACS and NHC best tracks provide meteorological observations with specific post-analysis histories; OpenQuake and Oasis documentation define computational methods; actuarial standards define certain professional expectations; and EIOPA or NAIC publications describe distinct regulatory applications. Substituting one category for another is a serious evidentiary error.
A deterministic financial-module test suite should cover zero loss, deductible boundaries, exact limit exhaustion, shared-policy deductibles, proportional participation, multiple covered perils, annual aggregate triggers, and reinsurance order. A stochastic calculation suite should compare event frequencies and conditional loss moments to their input distributions; reproduce known expected-loss sums; test the OEP/AEP inequality under common financial bases; and compare small Monte Carlo experiments with analytic benchmark cases such as the preceding example. Independent code review is especially valuable around low-probability events, floating-point threshold comparisons, joins between event IDs, and changes in portfolio or contract effective dates.
An input file can pass a schema check while containing incorrect geography or economically inconsistent valuation. The Oasis open data standards initiative makes data interoperability feasible, but semantic QA still requires checks against known insured-interest identifiers, locations, totals, and policy coverage. In a multi-vendor comparison, each vendor's “gross” or “net” label must be reconciled to the same contractual loss perspective. Comparison of unaligned model outputs may measure different definitions rather than different scientific risk estimates.
The ASOP No. 38 documentation and disclosure framework is particularly relevant to model reliance and communication in applicable U.S. actuarial services. A well-supported report discloses the degree of reliance on external model providers, material limitations, purpose and intended use, and departures that affect interpretation. Where a proprietary hazard module cannot be inspected, analysts can still demand event-level diagnostics, input descriptions, benchmarking, stress tests, change logs, and scope limitations. Proprietary implementation by itself supplies neither scientific validation nor proof of invalidity.
A complete reproducibility package for an applied analysis should include anonymized or appropriately protected exposure tables, model and data versions, documented equations or software contracts, random seeds, simulation horizons, financial terms, independent test cases, numerical tolerance specifications, and exact result-table definitions. Sensitive policyholder details require access control and privacy-preserving aggregation. Published scholarly discussion can describe models and calculations without exposing customer data or internal commercial catastrophe model intellectual property.
Editorial source review is part of reproducibility. The assertion supported by a source should be narrower than or equal to that source's own contribution. Hazard observations must not be cited as proof of contract indemnity; standards must not be cited as empirical loss estimates; publisher summaries must not be treated as independently validated individual model outcomes. Author-year references provide historical attribution, contextual inline hyperlinks permit reader verification, and complete bibliographic entries preserve the source identity. This research edition accompanies its own machine-readable EVIDENCE.json, which records source-to-claim relationships without exposing internal source identifiers in the scientific chapters.
Integrated Interpretation and Research Boundaries
The evidence supports an integrated interpretation of catastrophe loss as a conditional financial consequence of physically uncertain events. A more detailed hazard model does not necessarily improve an insurer's annual loss distribution if the exposure inventory is mislocated or if policy limits are applied at the wrong level. Conversely, precise claims records cannot validate a hazard frequency process beyond the narrow selection of insured and reported events they represent. The required validation effort should follow the error mechanism most capable of changing the intended actuarial output.
The first original analytical conclusion concerns loss-measure integrity. The complete calculation must preserve a coherent relationship among physical damage, covered indemnity, gross loss, ceded loss, retained loss, and time aggregation. These quantities are connected by defined operations, but they need not share the same tail shape. A deductible or excess layer can transform a modest change in damage into a discontinuous change in payment probability near attachment. Model QA should therefore examine loss measures on both sides of major contract thresholds, not only their final annual mean.
The second conclusion concerns spatial and temporal dependence. Portfolio diversification and annual accumulation are properties of joint event histories. Assetwise marginal vulnerability curves do not determine a portfolio's catastrophe tail in isolation. A scientifically credible event table must retain common event identity and a justified dependence specification, while an AEP calculation must retain the possibility of multiple events within the accounting year. The worked example proves this distinction exactly under its stated Poisson assumptions; real portfolios require the corresponding empirical or physical dependence analysis.
The third conclusion concerns uncertainty ownership. Monte Carlo sampling error, hazard-model parameter uncertainty, missing exposure characteristics, vulnerability-model uncertainty, and legal interpretation of policy terms are different inferential problems. They demand different evidence and different remedies. Increasing the simulated number of years may improve sampling precision but cannot correct the wrong indemnity clause, an incorrect geocode, or an out-of-date hazard calibration. Decision reports should identify which uncertainty sources were quantified and which were handled through bounded scenarios or explicit limitations.
The fourth conclusion concerns dynamic risk. Claims history provides evidence about realized losses under historical exposure, coverage, and hazard environments. Prospective catastrophe risk must explain changes in those environments. Climate projections, construction adaptation, protection measures, property inflation, and insurance penetration need distinct causal or statistical treatment. The scientific legitimacy of a climate-conditioned insured-loss scenario depends on tracing each adjustment through hazard, exposure, vulnerability, or contract mechanisms, rather than adding an unexplained global “climate factor” to all catastrophe payments.
The fifth conclusion concerns institutional translation. The same event and financial model can support pricing, reinsurance, ORSA, and regulatory reporting only after its output definition is aligned to each decision. An economic catastrophe-loss estimate is not directly an insurer claim; a claim is not automatically an unpaid reserve; a catastrophe PML is not automatically required solvency capital; and a reinsurance recovery estimate is not always immediately available liquidity. Insurance supervision and actuarial practice create explicit responsibilities for method choice and communication, while local contract law and capital regimes determine actual implementation.
The central limitation of this monograph is the absence of confidential insurer exposure and claims microdata, a specified real-world territory, and licensed proprietary catastrophe engines. Its purpose is to synthesize and verify the actuarial modeling architecture and provide a reproducible mathematical example. The numerical illustration is fully internally testable but intentionally synthetic. Documented external technical and regulatory materials substantiate the methods and institutional facts within their own scope; the work does not assert independent replication of the underlying empirical catastrophe studies or a comprehensive survey of every commercial model. Historical and future model calibration in any jurisdiction remains an application-specific research task.
Conclusion
Actuarial catastrophe loss modeling is the disciplined transformation of stochastic peril occurrence into contract-defined insurance loss distributions. Its primary scientific tasks are to identify the event process, reconstruct a spatially coherent hazard footprint, value and classify exposed interests, estimate conditional damage, apply insured financial terms at the correct level, and retain the joint event history necessary to calculate annual losses and risk-transfer recoveries. Reliable output depends on all these stages together. Mathematical precision in the final exceedance curve has limited meaning when exposure, vulnerability, dependence, or contract semantics are inconsistent.
The analysis demonstrates that event loss, annual aggregate loss, expected annual loss, OEP, AEP, and net retained loss must be treated as separately specified estimands. In the reproducible hypothetical portfolio, insured EAL is 1.506 million dollars, expected annual ceded occurrence loss is 0.940 million, and expected retained loss is 0.566 million. At a 3 million insured annual loss threshold, OEP is 25.918% and AEP is 27.285%, a material difference arising from multi-event accumulation. The example also shows how proportional physical vulnerability changes can produce nonlinear insured and reinsured loss responses through deductibles and payment limits.
The primary-source review establishes complementary roles for physical hazard observations, engineering loss functions, transparent catastrophe engine specifications, actuarial standards, and international supervisory frameworks. It also establishes boundaries: historical observations alone cannot resolve extreme prospective tails; sampled damage ratios do not determine joint spatial loss without dependence assumptions; and internationally influential standards do not automatically impose identical legal obligations in every insurance jurisdiction. Model transparency requires source provenance, method-specific calibration, contract-faithful financial modeling, independently interpretable validation, and reproducible outputs.
For insurance risk science, the durable methodological conclusion is that catastrophe-loss credibility resides in a verified sequence of conditional transformations from event to financial obligation. The practice standard that follows from this synthesis is one of evidentiary coherence: each modeled loss measure should retain its physical, statistical, contractual, geographic, temporal, and institutional meaning from source data to final actuarial decision. These are analytical conclusions from established science and verified institutional documents. No new proprietary theory, doctrine, or conceptual system is claimed.
References
Actuarial Standards Board. 2000. Actuarial Standard of Practice No. 39: Treatment of Catastrophe Losses in Property/Casualty Insurance Ratemaking. Effective December 15, 2000; public online text last revised May 2011. ASOP 39 text.
Actuarial Standards Board. 2019. Actuarial Standard of Practice No. 56: Modeling. Document No. 195. Adopted December 2019; effective October 1, 2020. ASOP 56 text.
Actuarial Standards Board. 2021. Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas), revised edition. Adopted July 2021; effective December 1, 2021. ASOP 38 text.
Embrechts, Paul, Claudia Klüppelberg, and Thomas Mikosch. 1997. Modelling Extremal Events for Insurance and Finance. Stochastic Modelling and Applied Probability, vol. 33. Berlin and Heidelberg: Springer. doi:10.1007/978-3-642-33483-2. Springer book record.
European Commission. 2015. Commission Delegated Regulation (EU) 2015/35, Article 120, “Natural Catastrophe Risk Sub-Module.” Consolidated in the EIOPA Solvency II Single Rulebook online edition, accessed November 14, 2026. Solvency II provision.
European Insurance and Occupational Pensions Authority (EIOPA). 2025. “Opinion on the 2023/2024 Reassessment of the Nat Cat Standard Formula.” Published January 30, 2025. EIOPA opinion.
Global Earthquake Model Foundation. n.d. “Event Based Damage and Risk: Outputs.” OpenQuake Engine 3.26.2 Documentation, accessed November 14, 2026. OpenQuake outputs.
Global Earthquake Model Foundation. n.d. “Event Based Risk.” OpenQuake Engine 3.23.5 Long-Term Support Manual, accessed November 14, 2026. OpenQuake configuration.
Global Earthquake Model Foundation. n.d. “Probabilistic Event-Based Risk Calculator.” OpenQuake Engine 3.26.2 Documentation, accessed November 14, 2026. OpenQuake calculator.
Grossi, Patricia, and Howard Kunreuther, eds. 2005. Catastrophe Modeling: A New Approach to Managing Risk. Huebner International Series on Risk, Insurance and Economic Security, vol. 25. New York: Springer. doi:10.1007/b100669. Springer publisher record.
Huizinga, Jan, Hans de Moel, and Wojciech Szewczyk. 2017. Global Flood Depth-Damage Functions: Methodology and the Database with Guidelines. EUR 28552 EN. European Commission Joint Research Centre. doi:10.2760/16510. JRC research report.
Intergovernmental Panel on Climate Change (IPCC). 2021. “Chapter 11: Weather and Climate Extreme Events in a Changing Climate.” In Climate Change 2021: The Physical Science Basis, Working Group I Contribution to the Sixth Assessment Report. Cambridge University Press. IPCC chapter.
International Association of Insurance Supervisors (IAIS). 2024, as presented in the current online tool. “Insurance Core Principles and Common Framework: ICP 16, Enterprise Risk Management for Solvency Purposes.” Accessed November 14, 2026. IAIS online standards.
National Association of Insurance Commissioners (NAIC). 2025. “Catastrophe Models Property.” Insurance Topics. Last updated June 2, 2025. NAIC. NAIC guidance.
National Flood Insurance Program. n.d. “Frequently Asked Questions about NFIP Policies and Claims Data.” NFIP Services / FloodSmart, accessed November 14, 2026. NFIP data guidance.
National Oceanic and Atmospheric Administration, National Centers for Environmental Information. 2024–2026. “International Best Track Archive for Climate Stewardship (IBTrACS), Version 4r01.” Data portal and technical documentation, accessed November 14, 2026. IBTrACS data portal.
National Oceanic and Atmospheric Administration, National Hurricane Center. 2026. “NHC Data Archive: Tropical Cyclone Reports and HURDAT2 Best Track Data.” Data portal, accessed November 14, 2026. HURDAT2 archive.
Oasis Loss Modelling Framework. n.d. “Financial Module.” Oasis LMF Documentation, version 2.5.9, accessed November 14, 2026. Oasis financial documentation.
Oasis Loss Modelling Framework. n.d. “Modelling Methodology.” Oasis LMF Documentation, version 2.5.9, accessed November 14, 2026. Oasis modeling methodology.
Oasis Loss Modelling Framework. n.d. “Open Data Standards: Open Exposure Data and Open Results Data.” Official standards overview, accessed November 14, 2026. Oasis data standards.
Oasis Loss Modelling Framework. n.d. “Sampling Methodology.” Oasis LMF Documentation, version 2.5.9, accessed November 14, 2026. Oasis sampling methodology.
Swiss Re Institute. 2026. “Natural Catastrophes in 2025: The Persistent Rise of Wildfire and Storm Risk.” sigma, no. 1/2026, March 16, 2026. Swiss Re Institute. Swiss Re sigma study.
U.S. Geological Survey. n.d. “ANSS Comprehensive Earthquake Catalog (ComCat).” Earthquake Hazards Program, accessed November 14, 2026. USGS ComCat.
U.S. Geological Survey. n.d. “ShakeMap.” Earthquake Hazards Program: Ground-Motion Maps, Atlas, and Scenarios, accessed November 14, 2026. USGS ShakeMap.