RESEARCH DOSSIER
Property Insurance Model Uncertainty and Validation Across Perils
Abstract. This monograph examines physical, statistical, economic and financial model uncertainty in property catastrophe loss simulation across earthquake, windstorm, flood, fire and other insured perils. It synthesizes documented primary evidence from NAIC, IAIS, OpenQuake, Oasis, FEMA and EIOPA with original research on uncertainty, vulnerability dependence and catastrophe model data quality. Six scientific chapters distinguish model-conditioned event variability, epistemic alternatives, measurement error, stochastic simulation precision, out-of-event and out-of-region validation, insurance contract nonlinearity and international model governance. Original examples illustrate variance decomposition, simulation and input sensitivity, compound hazard loss overlap, peril-specific observed-to-expected ratios, annual expected loss and dependence-sensitive portfolio tails. The methodology insists on explicit model, information, population, asset and financial loss definitions, and on independent evidence when a model is applied to an insurer's actual portfolio. Findings constitute an original analytic synthesis of established catastrophe and actuarial science; no private insurer loss data or original empirical calibration is claimed.
Keywords: catastrophe modeling; property insurance; model validation; aleatory risk; epistemic uncertainty; seismic hazards; flood risk; windstorm; wildfire; annual average loss; exceedance probability; exposure data; reinsurance; insurance internal models; Solvency II; simulation convergence.
Introduction
Property insurance catastrophe models calculate distributions of financial loss by combining stochastic hazard events, building and asset exposures, physical damage mechanisms and policy terms. An insurer may use these results for underwriting, reinsurance, claims forecasting, reserves, catastrophe protection and capital. Validation must be specific to each decision because the accuracy of an estimate of annual average loss does not establish the reliability of a distant tail quantile or a particular event's insured loss. An apparently precise probability can still rest on uncertain construction classifications, changing hazard inputs and incomplete claims records. The scientific object of this monograph is the measurement and attribution of those uncertainties across fire, flood, wind, earthquake and other insured perils.
The NAIC catastrophe-model overview describes the simulation of natural events, engineering damage and contractual financial transformation as distinct operations. The Oasis modeling documentation similarly separates hazard, exposure, vulnerability and an event loss table before summarizing financial exceedance. These models depend on inputs that may be independently valid or defective. Two erroneous components can cancel in an aggregate loss amount and create an apparently good historical fit. The validation problem is therefore to test individual modules and the assembled system against appropriate evidence, distinguishing measured event outcomes from hypothetical losses and official supervisory requirements.
An event catalog describes the stochastic peril events and their frequency representation. A hazard model translates each event to intensity at a specified asset location. Exposure data describe the assets actually at risk, including building characteristics and values. Vulnerability relates peril intensity to damage. Financial processing applies coverage and insurer obligations. At every stage, both numerical variability and uncertainty about the scientific adequacy of assumptions may arise. A model's output becomes meaningful only after the unit, region, peril, population, financial perspective and forecast time have been declared.
This international synthesis uses original official and peer-reviewed sources and original reproducible mathematical examples. It examines the scientific consequences of changing data, model versions and event sets; the challenges of validation in new geographic and climatic conditions; the need for tail-risk convergence tests; and insurance model governance under established supervisory frameworks. No new proprietary catastrophe theory or independently fitted insurer loss model is claimed.
Scientific Foundations and Primary Sources
Source architecture and physical causation
The OpenQuake model inputs guide distinguishes source and rupture models, ground-motion models, site information, epistemic logic trees, exposure, vulnerability, consequences and reinsurance. Their separation enables a precise scientific question for every change in an insurer's output. A ground-motion model can be wrong because its tectonic-zone parameters are inadequate, while a building's loss can be wrong because a structural class was assigned incorrectly. If those two errors compensate each other, validating only final financial loss fails to identify the internal deficiencies.
The GEM vulnerability method relates conditional probabilities of damage states to economic loss consequences. A mean ground-up vulnerability curve reflects the expected damage cost for specified intensity and asset type; a conditional distribution also requires dispersion. The consequence for insurers is material: deductibles, policy limits and aggregate reinsurance are nonlinear in physical loss. Two vulnerability models with equal mean damage may have different probabilities of extreme partial or complete destruction, generating different insurer financial tails.
OpenQuake logic trees allow alternative plausible seismic source and ground-motion models to be represented as formally weighted branches. These branches describe epistemic model alternatives and their consequences. They do not resolve uncertainty by themselves, because the chosen set of branches, weights and underlying evidence also require scientific judgment. A similarly explicit framework can be used for storm frequency assumptions or flood-defense scenarios, but physical meanings and parameter constraints differ between perils.
Historical evolution and original empirical findings
Property catastrophe modeling expanded as computational simulation, insurance exposure databases and meteorological and seismic information improved. Born (2023) examines differences in modeled average loss attributable to variations in input building locations, construction, height and event catalogs in a specific earthquake-exposed campus. The study demonstrates the importance of source data quality and correlated input assumptions. Its results are about a defined model and sample; they do not supply one percentage correction applicable to all countries or insurers.
The Florida Public Hurricane Loss Model data study by Pinelli and colleagues shows how county tax appraisals, wind insurance exposure and other original sources enter hurricane vulnerability model construction and validation. Resolving gaps and conflicts among those sources can reduce epistemic uncertainty. In contrast, adding more simulated events to a model does not repair a systematically wrong building-construction label, insured value or hazard relationship. The distinction between data error and simulation uncertainty is central to the interpretation of capital models.
The original flood model benchmarking review by Gerl and colleagues documents diversity in flood-damage function families, valuation denominators, geographical scope and validation practices. A regional flood-depth curve may be appropriate to its original building inventory but transfer poorly to another area's construction types and elevation conventions. Similarly, observed fire loss may be sensitive to local mitigation and building systems, and earthquake damage to code vintage and structural design. Comparisons across perils should therefore retain the full taxonomy and the source of the intended outcome, rather than mechanically ranking loss curves by an aggregate fit measure.
Aleatory Variability, Epistemic Uncertainty, and Data Error
Mathematical decomposition
A catastrophe loss model contains random variability from event occurrence and severity, parameter uncertainty within estimated event and damage functions, and uncertainty from structural assumptions or missing data. Let L denote annual loss and M a model branch. Under a defined model mixture, the total variance identity is Var(L) = E[Var(L|M)] + Var(E[L|M]). The first component describes event and damage variability conditional on a specified branch; the second describes differences among branch means. The mixture must specify a valid distribution of M and finite moments. This formal identity does not prove that the model branches or weights represent all scientifically plausible outcomes.
Consider two equally weighted hypothetical branches, one predicting $8 million annual average insurer loss and the other $14 million. The mixture mean is $11 million. The between-model standard deviation of mean losses is $3 million. Even if millions of stochastic simulation years make the numerical estimate inside each branch extremely precise, their modeled expected losses still differ by $6 million. This difference is epistemic within the selected branch system. It does not become a confidence interval for observed future insurer experience without further assumptions about branch completeness and plausibility.
The Kaczmarska, Jewson and Bellone simulation study investigates how nested stochastic event models create different convergence behavior for annual average loss and tail risk statistics. Extreme quantiles can be especially unstable relative to their mean. A serious scientific review therefore measures convergence for the exact risk measure to be used. It records random seeds, simulation periods, event-sampling methods, losses before and after financial terms, and the repeatability of tail estimates across independent runs.
Exposure and measurement uncertainty
Actual properties carry incomplete descriptions. An address may geocode to the wrong parcel, a height or number-of-stories field may be absent, construction material may be inaccurate, and replacement value may refer to market valuation rather than reconstruction. These inputs affect model calculations differently by peril. Flood loss is sensitive to elevation and below-grade exposure, wind damage to roof structures and connections, earthquake to structural type and local shaking, wildfire to building materials and surroundings. A uniform ten-percent exposure uncertainty correction cannot reliably represent these heterogeneous mechanisms.
The Born empirical model assessment and Pinelli vulnerability data study demonstrate that data quality and model specification can affect average loss and portfolio decisions. Repairing an exposure error can be more valuable for a particular insurer than improving an already stable stochastic simulation algorithm. A scientific validation plan ranks data needs according to the insurer's actual sensitivity and the uncertainty of the decision, not according to data collection ease.
Event and peril dependence
Natural disasters can trigger overlapping hazards. Hurricanes cause wind, surge and inland flood; earthquakes can induce fire, landslides and other consequences. Loss aggregation requires physical component accounting and insurer occurrence rules. If a first hazard destroys 40% of a building's value and a second hazard damages 25% of the remaining undamaged portion, the combined loss ratio is 0.40+(1−0.40)×0.25=0.55. A naive sum gives 0.65, overstating damage by ten percentage points under this hypothetical sequential model. The arithmetic demonstrates the requirement to declare overlap and conditional loss bases. Actual compounding must be supported by engineering evidence.
Cross-Peril Validation and Empirical Design
Hazard, vulnerability, and contract separately
Validation should begin with the observable for each component. Hazard science can be tested against meteorological records, earthquake ground-motion measurements, mapped flood depths and fire perimeters. Vulnerability science can be assessed against engineering damage and reconstruction costs for matching construction types and physical intensity. Exposure data can be verified through insured schedules, building inspections, cadastral records and valuation. Insurance financial transformations can be compared with actual policy terms and an independently developed claim ledger. An event loss prediction is only as credible as the definitions of those inputs and their completeness.
A model whose financial prediction happens to match aggregate past loss may still be materially wrong in construction-specific components. In a synthetic example, imagine a model overpredicts $25 million of industrial property loss and underpredicts $25 million of residential loss. The net total error is zero. An insurer concentrated in either segment would experience an important systematic error even though its comparison to an overall market aggregate looked excellent. Validation must therefore include subgroup residuals, geography, perils, construction, occupancy and value bands.
The FEMA Hazus 7.2 primary announcement documents the June 4, 2026 release and integration of USGS ShakeMap features, following retirement of the prior ArcGIS Desktop-supported Hazus generation. Model versions can change the information and processing that enter scenario losses. Scientific reproducibility requires both original input datasets and the exact engine version. Applying a new model retroactively to an older event creates a useful new analysis but must not be described as the model's historical real-time performance.
The Oasis model validation initiative recognizes the importance of average annual loss and loss exceedance as distinct outputs. A rigorous validation archive should publish target population, event and peril catalog, source extraction time, loss perspective, mapping to insurance contracts and the statistical uncertainty of each comparison. Engineering inspection, satellite data and claims may all be needed, but they answer separate evidentiary questions.
Out-of-event and out-of-region testing
Catastrophe model calibration from historical events risks overfitting the peculiarities of the few disasters with detailed claims data. An event-based test holds out whole catastrophes from the estimation sample; geographically blocked tests hold out countries or regions; construction-specific tests hold out classes poorly represented during training. These structures reflect the dependence of observations that share one storm, event valuation or local claim practice. Random splits of individual buildings within a single hurricane can leak event-specific features and overstate generalization to an unseen hurricane.
The EIOPA catastrophe data hub presents insurer exposure and incurred loss data drawn from a defined reporting sample. Its statistical population differs from all buildings in Europe and from the entire insurer market. A model's error measured in that sample can be informative, but external validity requires comparing insured values, coverages, reporting completeness and event mix. The EIOPA protection-gap dashboard separately tracks modeled current and historical economic versus insured losses. Those perspectives must not be combined into one assumed insured-claim ground truth.
Calibration by peril
Suppose hypothetical observed wind loss is $120 million and its model prediction is $80 million, while flood observed loss is $80 million and prediction is $120 million. The pooled observed-to-expected ratio is (120+80)/(80+120)=1. The wind ratio is 1.5 and flood ratio is 2/3. A pooled score of one conceals offsetting error across perils and might induce unsuitable underwriting or reinsurance decisions. Calibration reports should disclose both pooled and peril-specific diagnostics, with confidence intervals reflecting the size and dependence of the underlying event samples.
Portfolio Tail Risk, Insurance Finance, and Governance
Expected annual loss and exceedance
Annual average loss is an expectation. It is additive across peril components under correctly defined and nonoverlapping losses. Loss quantiles and exceedance thresholds are nonlinear properties of the joint portfolio distribution. The Oasis modeling methodology explains that combined return-period exceedance curves must be recalculated, even when peril-region annual average losses can be added. Dependence between hazards, shared insured buildings, event clustering and contract aggregation affect the tail.
A synthetic example isolates dependence. Two losses each equal $1 million with 10% annual probability and otherwise zero. Under independence, the chance of at least one loss is 1−0.9²=0.19 and chance of two losses is 0.01. With perfect positive dependence, the chance of at least one loss is 0.10 and of two losses is 0.10. Both models imply the same $200,000 expected annual aggregate loss. Their aggregate financial tails, and the value of a reinsurance layer attaching beyond $1 million, are materially different. Mean agreement cannot validate tail dependence.
The Silva seismic vulnerability correlation study demonstrates the importance of conditional damage variability and correlation for aggregate loss. Correlation from shared hazard intensity should be distinguished from residual correlation in damage conditional on the hazard field. Modeling both twice can misstate the portfolio distribution. Extreme event stress tests should examine both dependence assumptions and the economic consequences of building and insured contract concentrations.
Insurance contract nonlinearity
The Oasis financial module separates ground-up loss, direct insured loss and net insurer loss after reinsurance. A simplified primary-policy payment is min(max(L−d,0),u), where L denotes otherwise covered physical loss, d the deductible and u the insurer's payment limit. If L is random, E[g(L)] generally differs from g(E[L]). A full validation should test the entire loss distribution against policy layers rather than applying financial terms only to an average damage ratio.
Suppose a building experiences no covered loss with probability 0.8 and a $500,000 covered loss with probability 0.2. Expected ground-up loss is $100,000. With a $25,000 deductible and $200,000 insurer-payment cap, the expected insured payment is 0.2×$200,000=$40,000. Applying the payment formula to the $100,000 mean would yield $75,000. The $35,000 difference is entirely due to the modeled conditional distribution crossing a nonlinearity. The example is hypothetical; exact rules for actual policies must follow their operative wording.
The Oasis Open Results Data specification defines event, period, average and exceedance loss tables under a consistent schema. Standardized outputs allow models to be compared on the same financial perspective when exposure and peril definitions also agree. Standardization of the file alone cannot establish that a claimed result reflects the appropriate contract or that a model has passed actuarial validation.
Supervisory and internal model standards
The IAIS Insurance Core Principles address internal model governance, capital calibration and use in insurer risk management. National regulators implement those requirements under applicable domestic rules. The Solvency II internal-model provisions include use tests and ongoing appropriateness of internal models within the EU regime. The linked consolidated legal version is date-specific; any current legal application needs the controlling updated text and jurisdiction.
Scientific adequacy and regulatory permission are separate evaluative objects. An institution may have a technically useful catastrophe model for underwriting that has not been approved for regulatory capital. Conversely, supervisory approval does not eliminate the need to update scientific assumptions when hazard, exposure or claims practices change. A complete model inventory should identify model owner, latest source versions, external vendor dependencies, significant revisions, testing authority and independently documented findings. Model users must know how the output changes under key uncertainties and which decisions it can responsibly inform.
International Synthesis and Conclusion of Scientific Findings
Interpreting model output changes
Changes in catastrophe loss estimates can arise from hazard science, asset inventory, vulnerability, repair costs, insurance terms or numerical simulation settings. When several components change together, an attribution study reruns old and new versions under consistent input sets to isolate important sources. Because the complete loss mapping is nonlinear, attribution can depend on update order. A disciplined report states the chosen path or averages across paths instead of calling every increase in modeled loss a corresponding increase in real catastrophe hazard.
The original Born input-uncertainty comparison, Pinelli hurricane data management study and Kaczmarska numerical uncertainty study support a common scientific interpretation: correct inputs, scientifically adequate functions and stable simulation are separate sources of model credibility. Progress in one domain can be valuable while leaving other sources unresolved. This integrated source-to-claim synthesis does not establish a new doctrine; it specifies how established findings can be compared in an insurance model governance system.
The international context requires precise output definitions. Regulatory capital, insurer reserving, catastrophe rate adequacy, risk prevention and public protection gaps each select different target populations and financial perspectives. The same flood event can imply a very large regional economic damage amount and a smaller private insured loss because penetration and contract terms differ. A seismic hazard model calibrated for one structural stock may transfer poorly to another. Windstorm damage is sensitive to construction and geographic climate; wildfire depends on distinct fire propagation and building details. Cross-peril validation should preserve those differences while comparing common actuarial estimands where appropriate.
Replicable research priorities
A validation study should retain the exact event catalog, model software version, exposure and contract snapshot, code and risk-class mappings, hazard observation vintage, valuation currency and inflation basis, and actual claims development period. It should report output-specific convergence, out-of-event calibration, geographic transfer, unusual event exclusions, and sensitivity to correlation and epistemic alternatives. Where private data access is restricted, synthetic calculations and public simulations can validate technical identities and source structures, but empirical claim accuracy requires independently available target observations.
A further priority is the study of compound catastrophe damage. Correlated wind, surge and rainfall may interact, while separate policies assign different occurrences and perils. An event-loss database must prevent double counting repairs to the same components and must preserve real coverage differences. Improved multi-peril models can be tested against appropriately grouped historical events, engineering field surveys and claims files that disclose the timing and development of loss amounts.
The scientific conclusion is that model validation must be matched to the intended insurance decision. Component realism, insured population comparability, monetary calibration, occurrence and annual aggregate tail modeling, and regulatory governance all contribute to the credibility of a catastrophe estimate. A high degree of numerical precision within one simulation is insufficient to establish empirical certainty. International property insurance science can improve by making each observation, model assumption and financial transformation traceable and by testing claims in precisely defined populations.
The statistical meaning of return periods and decision limits
In a modeled annual loss distribution, a one-in-100-year aggregate loss threshold is conventionally the loss exceeded with annual probability 1%. This is an annual exceedance interpretation under the specified model, not a guarantee that such a catastrophe occurs only once during a hundred-year calendar span. The probability of at least one exceedance during 20 independent years with the same fixed 1% annual probability is 1−(0.99)^20, approximately 18.2%. The stationary independence assumption is important: changing hazards, concentrations of insured values and dependence among years can alter the multi-year probability.
The Oasis public model-validation standardization discussion emphasizes that AAL and loss exceedance represent different dimensions of risk. Tail output is conditional on the simulated event catalog and the empirical or theoretical frequency model. An occurrence exceedance curve describes the probability that the largest event in a year exceeds a loss threshold; an annual aggregate curve describes the probability that summed losses in the year exceed that threshold. These quantities can diverge in regions where several moderate storms occur in the same season. Reinsurance policies that cover single occurrences and those with annual aggregate attachment operate on different stochastic quantities.
An insurer can assess whether additional event-set simulation reduces numerical uncertainty by running independent simulations at progressively larger sample sizes and comparing mean annual loss, occurrence tail, aggregate tail and confidence intervals at stated thresholds. If convergence is slow, the appropriate computational response may involve variance reduction, importance sampling or other scientifically justified event-sampling schemes. These require proof that estimator weights and the target probability measure are correct. A high number of simulated events alone does not prove stable tail estimates, because event catalogs may omit scientifically plausible catastrophe modes or contain biased frequency parameters.
The original Kaczmarska and colleagues analysis investigates simulation uncertainty under nested catastrophe loss models. Its principal implication for insurance practice is that sampling methodology should be selected for the statistic of interest rather than only for the ease of generating events. A model built to estimate a mean can have different computational requirements from one intended to estimate the upper 0.5% annual aggregate loss. Results should display effective sample size and sampling error where feasible, with separate ranges for uncertainty about hazard model and vulnerability alternatives.
Data-version control, scientific traceability, and information timing
A reproducible catastrophe study records the exposure date, policy term date, hazard record collection period, event-forecast origin, original and final intensity fields, vulnerability function identifier, consequence-cost price year, catastrophe software version and random seed. These data permit a later researcher to distinguish changes in risk from changes in the description of the same risk. A location record corrected after a disaster may improve the final loss estimate while making a retrospective reconstruction appear artificially accurate. When model performance is asserted prospectively, every input must be frozen to the information that was actually available at the forecast origin.
The 2026 FEMA Hazus release notice provides a concrete case of model-version dependence. New ShakeMap-linked functionality changed the route from a seismic observation to a scenario damage estimate, while retirement of an earlier GIS foundation affected support for a previous software generation. A historical Hazus result and a newly recomputed Hazus result should therefore be accompanied by their versions. Likewise, a flood loss forecast derived from one digital elevation model may change when a newer topographic dataset resolves a building's first-floor datum. Neither change proves that the climate or the building's underlying structural vulnerability changed.
A detailed provenance matrix has at least five columns: source and version, model component and variable, observation or calibration population, information date, and known validity limitations. An additional column can record the output quality gate and who reviewed it. This is especially useful when a proprietary third-party catastrophe model sits alongside an insurer's own claims development or actuarial aggregation code. The vendor may revise its hazard event catalogue, while the insurer simultaneously updates its geocoding and contract values. Unless these changes are separated, a reported reduction in AAL cannot be attributed to any specific scientific or commercial driver.
The Oasis Open Results Data standard helps preserve results in consistent event, period and average loss table structures. Standardized schemas support independent technical comparison across source systems, but they do not ensure that two model outputs share the same economic definitions or event catalog. An analyst should not compare a gross occurrence exceedance curve from one model with a net annual aggregate curve from another simply because the files use similarly named fields.
The validation sample and its observational selection
A catastrophe insurer may hold private property-level claim files that include detailed coverage and incurred or paid-loss fields. These records are sampled by insurance penetration, reporting incentives, policy coverage, deductible thresholds and claims development time. Economic damage to an uninsured house is absent from private claims, and small damage below a deductible may never be submitted. A national disaster recovery database can represent a different population, especially if eligibility for public assistance differs from insurance coverage. Validation against either source requires an explicit target-population statement and, where possible, adjustments for sample inclusion.
The EIOPA catastrophe data hub aggregates a sample of reporting insurers and undertakings. It can support comparisons under its stated exposure and loss definitions. The EIOPA protection-gap dashboard presents modeled economic risk and historical insured loss perspectives. The scientific research question changes depending on which of those population levels is chosen. A model predicting property-insured loss for a single carrier should not be judged solely against a whole-economy physical loss figure; a whole-economy damage model should not be calibrated as if a private insurer's reported claims were a census of all loss.
An observed-versus-expected ratio compares the sum of observed losses and the sum of expected model losses over identically defined events and exposure. When there is high claim development or litigation, an early incurred estimate may not be comparable with ultimate paid-loss prediction. The error can be mechanical rather than a failure in catastrophe science. A validation report should specify whether losses are reported, incurred, paid or ultimate; whether they are adjusted for price inflation; whether insurer deductibles and limits are included; and whether a claim is indexed by catastrophe occurrence, policy record or financial accounting calendar.
Coverage differences can be spatially correlated with risk. Flood insurance may be rarer in a region where insurers restrict flood coverage; homes at high wildfire hazard may face different limits and deductibles. This selection can bias a comparison between an insurer's insured portfolio and a national hazard exposure inventory. A statistically credible study should compare the covariate distribution of the observed claims sample with the target model population, and state where missing or unverifiable data prevent identification. More flexible machine learning cannot substitute for missing physical and contractual observations.
Component interaction and attribution of change
Model upgrades can affect outputs nonlinearly. Suppose a hypothetical previous catastrophe model used an exposure database A and vulnerability curve V, and a new model updates both to A' and V'. One can compute losses under A,V; A',V; A,V'; and A',V'. The marginal effect of revising the inventory depends on which vulnerability curve is held fixed. If the difference between A,V and A',V' is large, explaining it by a single "new hazard risk" number is inadequate. A fair model change attribution might report each component's effect under both orders, or average them using an explicitly defined decomposition.
For a fully hypothetical two-factor example, let the old model's annual expected cost equal $100 million. Updating exposure alone raises it to $110 million. Updating vulnerability alone raises it to $120 million, and updating both raises it to $135 million. The interaction effect is $135−$110−$120+$100=$5 million. The total increase of $35 million is composed of a $10 million exposure-only change, a $20 million vulnerability-only change, and a $5 million interaction. The decomposition is an algebraic identity for these modeled outputs. It demonstrates why historical changes should be attributed to jointly modified assumptions rather than automatically interpreted as an observed increase in the frequency of catastrophe events.
The Born catastrophe input study highlights interactions between building attributes, locations and the event catalog. These are precisely the kinds of dependencies for which a one-factor-at-a-time analysis may be incomplete. In a flood model, a geocoding change may affect both water depth and building construction class. In a seismic model, an asset's assigned location determines its hazard intensity, and its construction class determines the vulnerability response. In a hurricane model, local wind exposure and structural roof characteristics may both depend on coastal location. An attribution analysis should therefore include interaction testing and refrain from presenting causal claims that the design cannot support.
Tail concentration, correlation, and reinsurance decisions
The distribution of insurer catastrophe loss depends on shared hazards and correlated vulnerability residuals. An earthquake ground-motion field produces spatial dependence, while neighboring buildings with similar age and structural systems may have correlated damage even conditional on a common intensity map. The Silva 2019 seismic vulnerability study investigates correlation in conditional loss ratios and the consequences for portfolio loss-exceedance behavior. Correlation assumptions can materially affect a high catastrophe layer despite leaving the individual building's expected loss unchanged.
A simple calculation distinguishes dependence from expected loss. Let two $1 million insured assets each face 10% annual probability of complete loss. The total expected loss equals $200,000 under independent outcomes and under perfect common outcomes. Under independence, the probability of both total losses is 1%, and under perfect dependence it is 10%. A reinsurance layer that attaches at aggregate losses above $1 million has a different expected recovery in the two distributions because only the double-loss state triggers the layer. Its scientific valuation must therefore use the joint distribution, not separate marginal mean curves.
The Oasis combined-peril guidance distinguishes additive expectations from nonadditive tail quantiles and discusses loss exceedance recomputation. In practice, insurers may have several catastrophe models whose event sets are built independently. Their joint aggregation should state whether annual peril losses are assumed independent, whether event clustering is represented, and whether combined hazards share insured assets. A portfolio in a coastal region can face compound wind, precipitation, inundation and fire effects; assigning them separate catastrophe tags does not automatically make the insured losses independent or additive.
Contract definitions matter further. A primary building loss may cross a per-risk deductible, and an annual occurrence treaty may combine multiple locations under an hours clause. If the model aggregates financial terms in an incorrect order, expected gross and net loss can both be wrong even if the physical loss distribution is valid. Oasis financial processing makes these different financial perspectives explicit. A validation program should test contractual arithmetic on independently known examples, including partial loss, full loss, multiple locations and events that trigger reinsurance exhaustion.
Independent review and governance in international regulation
The validity of a catastrophe model for an internal pricing or claims decision depends on its observed and documented performance, while authority to use an internal model for statutory capital depends on applicable supervisory requirements. The IAIS ICP framework articulates principles for internal model quality, validation, governance, calibration and use. The European Solvency II directive places obligations on the use and maintenance of internal models in EU insurance capital regulation. These sources identify regulatory principles and dated legal provisions; they do not replace a jurisdiction-specific compliance assessment of a particular insurer.
A scientific internal review should document independent assessment of model inputs, code validation, financial transformation, historical fit, sensitivity and the consequences of material assumptions. Reviewers should have access to the exact model version and frozen datasets used to produce the results. They should assess whether unexplained differences in loss estimates arise from changes in scientific inputs, portfolio composition or contract layers. Approval may be conditional on risk appetite, intended use and data limitations. When a vendor model has undocumented proprietary elements, the insurer should record the resulting verification boundary and use independent observable benchmarks and sensitivity ranges where feasible.
The model review record must also specify what it does not establish. A full-source repository build validates data transformations and reproducibility; it does not establish that a rare 200-year event has been empirically observed or that a climate-projected future distribution has been calibrated by the insurer. A scientific article's model-validation findings may be strong for a specified geography and asset population but require transfer tests elsewhere. Published model accuracy and internal supervisory approval should each be attributed to the actual organization, population and method responsible.
Research integration
Across flood, fire, wind and earthquake, the principal source of scientific difficulty changes with the hazard and the corresponding observation mechanism, yet a stable validation architecture remains possible. It begins with the desired actuarial estimand and its financial perspective. It identifies the scientific event process, the intended insured assets and their physical characteristics, the vulnerability and cost relationships, the insurance coverage transformation and the observation time. It then tests the model's components using the kind of primary evidence appropriate to each, and the assembled outputs under independent event and portfolio comparisons.
The research findings in Gerl's flood model survey, Pinelli's wind data study, Silva's seismic correlation analysis, and the documentation of Oasis loss simulation converge on this methodological requirement. The central achievement of a source-grounded international synthesis is to articulate which differences among risk estimates are scientific, observational, computational, contractual or institutional. Reliable comparison becomes possible through exact source, exposure, peril and output definitions and through explicit uncertainty, not through suppression of differences between models.
A practical statistical validation matrix
An operational comparison should define a matrix rather than a single model-quality score. The rows represent each covered peril and material territory; the columns represent hazard observations, exposure correctness, vulnerability loss fractions, primary insured contracts, annual average losses, occurrence exceedance and aggregate exceedance. Within each cell the analyst records the evaluation population, observation period, independently verified data source, validation statistic and unresolved measurement limitations. This makes it possible to identify a high-performing seismic hazard model coupled with a weak local building exposure map, or a well-calibrated wind model that still misprices flood-covered business interruption.
A uniform observed-to-expected ratio can be computed for selected matched events as O/E = Σ observed losses / Σ predicted losses. The ratio is meaningful only for a common set of covered assets, event definitions, inflation-adjusted dollar values and claim development stage. A ratio of one is a statement about aggregate equality, not about the shape of the residual distribution. If a model overstates small claims and understates catastrophic claims, an AAL may appear good while the layer that protects insurer solvency is misvalued. Validation must include tail-sensitive statistics and comparisons by damage severity, value and geography.
For probability-based scenario events, calibration can be evaluated through proper scoring functions. If a model predicts probability p for a specified damage event and the observed outcome is y equal to zero or one, the Brier score is (p−y)². A model that assigns probability 0.8 to a building's damage and observes no damage receives 0.64 for that observation; if the building is damaged, the score is 0.04. Summing or averaging across a representative event sample creates a performance statistic, but observations within a catastrophe are correlated. Confidence intervals for such scores should therefore account for event clusters, and a rating model's score should not be claimed to assess financial severity when its dependent variable is only binary physical damage.
Cross-peril validation also requires consistent technical units. Flood intensity measured in meters above finished floor cannot be combined numerically with earthquake peak acceleration or three-second wind gust speed. The appropriate scientific comparison might be prediction error after translating each intensity through its correctly calibrated vulnerability function and the same definition of property value. Even then, different hazard observation methods and legal coverage rules cause substantive differences. A useful validation report states those differences instead of forcing a common normalized intensity measure where it lacks physical meaning.
Fully reproduced mathematical checks
The scientific examples in this monograph are designed to be independently recomputable. The equal-weight model mixture with expected annual losses of $8 million and $14 million yields a mean of $11 million and between-model variance of (3²+3²)/2 = 9 in squared millions, so the standard deviation is $3 million. The two-peril observed-to-expected example has wind O/E = 120/80=1.5, flood O/E=80/120=2/3, and pooled O/E=(120+80)/(80+120)=1. The balanced pooled result cannot demonstrate calibration within either peril.
For the two-risk dependence example, expected loss is 0.10×$1 million for each risk, totaling $200,000. Independence gives joint destruction probability 0.01; perfect positive dependence gives 0.10. The at-least-one-loss probabilities are 0.19 and 0.10, respectively. In the deductible example, 0.20×$500,000=$100,000 is expected ground-up loss. At the $200,000 insurer payment cap, indemnity in the positive-loss state is $200,000, so expected payment equals $40,000. Contract transformation applied only to the mean yields $75,000, a $35,000 error.
The sequential peril example gives 0.40+(1−0.40)×0.25=0.55. The naive sum gives 0.65, differing by ten percentage points. In the version attribution example, the old/new model outputs of $100 million, $110 million, $120 million and $135 million yield exposure contribution $10 million, vulnerability contribution $20 million and interaction $5 million. Finally, under an independent annual exceedance probability of one percent, the probability of at least one such exceedance within 20 years is 1−0.99²⁰ ≈ 0.1821. Every calculation is explicitly hypothetical and is evidence of mathematical consistency rather than an observed insurer's outcome.
Interdisciplinary research needs
A scientifically adequate insurer catastrophe model is developed across several professions. Seismologists, hydrologists, meteorologists and wildfire researchers establish physical event plausibility and hazard data. Engineers characterize structures and component vulnerability, including consequences of material and code variations. Statisticians and actuaries evaluate event frequency, stochastic dependence, sample bias and financial aggregation. Legal and claims specialists establish coverage conditions and the insurer's actual obligations. Data engineers preserve source lineage, versioned transformations and the relationship between building locations and insured contracts. A fully documented validation decision identifies the scope of each expertise and the tests actually performed.
This interdisciplinary structure is particularly important in periods of changing exposure or climate. One cannot infer that a more recent catastrophe model is scientifically superior solely because its expected loss estimate is higher or lower than a previous one. The appropriate test compares model outputs with independent evidence, assesses physical and contractual inputs and explains the source of changes. International research benefits from common output definitions and auditable calculation protocols, while retaining peril-specific hazard science and the jurisdictional meanings of coverage. The product is a validated statement about a defined risk and decision, supported by measurable evidence and a complete provenance trail.
Conclusion
Property catastrophe modeling connects physical hazard, exposure, engineering vulnerability, economic consequence and insurance contract obligations. Reliable prediction and capital use require distinct evidence for these links, with explicit stochastic and epistemic uncertainty, validated event and population definitions, sensitivity to data quality and policy terms, and quantitative validation of both average loss and extreme exceedance. The reviewed institutional standards and peer-reviewed research support component-level testing, prospective event-holdout evaluation, convergence checks and sustained international model governance across insured perils. A model's credibility is the documented fit of its scientific assumptions and outputs to the specific risk decision, with uncertainty, dependence and historical source revisions preserved in the research record.
References
National Association of Insurance Commissioners. (2025). Catastrophe Models: Property. https://content.naic.org/insurance-topics/catastrophe-models-property
International Association of Insurance Supervisors. (2024). Insurance Core Principles and ComFrame. https://www.iais.org/activities-topics/standard-setting/icps-and-comframe/
Oasis Loss Modelling Framework. (2026). Modelling methodology, version 2.5.9. https://oasislmf.github.io/2.5.9/oasislmf/explanation/modelling-methodology.html
Oasis Loss Modelling Framework. (2024). Standardising Cat Model Validation. https://oasislmf.org/news/standardising-cat-model-validation
Global Earthquake Model Foundation. (2026). OpenQuake Engine: Input Models. https://docs.openquake.org/oq-engine/manual/latest/user-guide/inputs/index.html
Global Earthquake Model Foundation. (2026). OpenQuake Engine: Logic Trees. https://docs.openquake.org/oq-engine/manual/latest/underlying-science/logic-trees.html
Kaczmarska, J., Jewson, S., & Bellone, E. (2018). Quantifying the sources of simulation uncertainty in natural catastrophe models. Stochastic Environmental Research and Risk Assessment, 32, 591–605. https://doi.org/10.1007/s00477-017-1393-0
Born, P., Dumm, R., & Johnson, M. E. (2023). Epistemic uncertainty in catastrophe models—A base level examination. Risk Management and Insurance Review, 26(2), 247–269. https://doi.org/10.1111/rmir.12246
Pinelli, J.-P., Da Cruz, J., Gurley, K., Paleo-Torres, A. S., Baradaranshoraka, M., Cocke, S., & Shin, D. (2020). Uncertainty reduction through data management in the development, validation, calibration, and operation of a hurricane vulnerability model. International Journal of Disaster Risk Science, 11, 790–806. https://doi.org/10.1007/s13753-020-00316-4
Federal Emergency Management Agency. (2026, June 9). FEMA Bulletin: Earthquake ShakeMap Model Added to FEMA's Hazus Tool. https://content.govdelivery.com/accounts/USDHSFEMA/bulletins/41a9b67
European Insurance and Occupational Pensions Authority. (2025). Dashboard on Insurance Protection Gap for Natural Catastrophes. https://www.eiopa.europa.eu/tools-and-data/dashboard-insurance-protection-gap-natural-catastrophes_en
European Insurance and Occupational Pensions Authority. (n.d.). Catastrophe Data Hub. https://www.eiopa.europa.eu/tools-and-data/catastrophe-data-hub_en
Global Earthquake Model Foundation. (2026). Global Seismic Vulnerability Model: Vulnerability Models. https://docs.openquake.org/vulnerability/vulnerability_models.html
Oasis Loss Modelling Framework. (2026). Financial module, version 2.5.9. https://oasislmf.github.io/2.5.9/oasislmf/explanation/financial-module.html
European Union. (2009, consolidated 2021). Directive 2009/138/EC on insurance and reinsurance (Solvency II), Articles 120–126. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A02009L0138-20210630
Gerl, T., Kreibich, H., Franco, G., Marechal, D., & Schröter, K. (2016). A review of flood loss models as basis for harmonization and benchmarking. PLOS ONE, 11(7), e0159791. https://doi.org/10.1371/journal.pone.0159791
Silva, V. (2019). Uncertainty and correlation in seismic vulnerability functions of building classes. Earthquake Spectra, 35(4), 1515–1539. https://doi.org/10.1193/013018EQS031M
Oasis Open Data Standards. (2026). Open Results Data (ORD), version 2.5.7. https://oasislmf.github.io/2.5.7/ord/index.html