Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Property Insurance Catastrophe Model Vulnerability Curves

Abstract. Catastrophe vulnerability functions connect the intensity of natural hazards to the distribution of physical damage and economic loss for buildings. This monograph examines their foundations, construction, calibration, transferability, uncertainty, and eventual application to property insurance contracts. It distinguishes seismic fragility curves from expected loss-ratio vulnerability curves; engineering damage states from repair-cost outcomes; hazard-event uncertainty from vulnerability uncertainty; and ground-up losses from insured and net financial losses. International primary evidence is drawn from the Global Earthquake Model Foundation, OpenQuake, FEMA Hazus, the Oasis Loss Modelling Framework, USGS earthquake-impact tools, and EIOPA catastrophe-loss documentation. Illustrative calculations are constructed examples rather than observed loss estimates. The analysis identifies validation requirements for multi-peril portfolios, including exposure taxonomy, measurement provenance, loss-cost denominators, dependency, model versions, reporting selection, and uncertainty around extreme outcomes.

Keywords: catastrophe modeling; property insurance; vulnerability functions; fragility curves; expected damage ratio; flood depth-damage; seismic risk; windstorm losses; exposure taxonomy; replacement cost; stochastic event sets; basis risk; model uncertainty; insurance deductibles.

Scientific Foundations and Vulnerability Terminology

Catastrophe risk modeling estimates the consequences of hazardous events for a specified distribution of assets, population, and financial interests. For property insurance, a basic analytic path runs from a hazard occurrence and its location-specific intensity to physical damage, repair or replacement costs, policy-responsive losses, and ultimately to distributions of contractual financial obligations. Vulnerability functions occupy the middle of that path. Their scientific object is a conditional relationship between damaging physical intensity and an economic or physical consequence. The model's applicability follows from the hazards, asset types, geographical and engineering contexts, cost definitions, and observation periods used to establish the function.

The GEM global seismic vulnerability documentation sets out an explicit connection among hazard, building classification, structural response, fragility and vulnerability. Its structural typologies distinguish material, load-resisting system, design code, earthquake-resistant design, and building height. These characteristics control how a nominally similar ground-motion level can produce very different levels of damage. An intensity measure is a variable such as peak ground acceleration or another ground-motion metric, defined with physical units and measurement assumptions. Its relation to damage requires a vulnerability specification for the class of asset under investigation.

A fragility function gives the conditional probability that a defined physical damage state is exceeded at a specified level of hazard intensity. A vulnerability function describes the conditional expected fraction of economic loss or a distribution of loss ratios, potentially obtained by integrating several damage states and their consequences. The GEM fragility formulation commonly uses a lognormal curve with a median intensity and a dispersion parameter for each threshold. This representation encodes an engineering interpretation of structural variability and uncertainty in the intensity-damage relation. A fragility exceedance probability must not be read directly as an expected monetary loss fraction.

The GEM vulnerability model explicitly obtains expected loss ratio through damage-state probabilities and damage-to-loss consequence ratios. A conditional mean is only one feature of this distribution. Two building classes can share the same expected damage ratio but have different tail probabilities of complete destruction, and their insurance costs may differ substantially because deductibles, limits and catastrophe aggregation are nonlinear. An insurer using only a smooth mean curve thus risks missing the contribution of dispersion and concentration to high-severity portfolios.

The operational meanings of loss must be explicit. Physical damage can mean cracking, deformation, inundation, component failure or functional interruption. Engineering replacement cost is a valuation base. Repair expenditure depends on labor, materials, access, inflation, emergency measures, code upgrades and reinstatement decisions. A covered insurance loss additionally depends on policy wording, applicable exclusions, insured values, deductibles, sublimits and adjustment arrangements. A paid claim may be recorded many months after the original hazard and may be revised or litigated. Collapsing these distinct observational outcomes into a single generic damage curve obscures the source of model error.

The Oasis catastrophe modeling platform demonstrates the separation of hazard and vulnerability model components in an interoperable engine. Its open model architecture allows scientific functions, exposure data and financial calculations to be updated independently. Such modularity supports auditability when the version of a vulnerability library changes without a corresponding change in geographic hazards or policy data. It also makes the validation problem more precise: an apparent improvement in modeled portfolio losses may arise from hazard data, replacement-cost values or new vulnerability functions rather than from an improved fit of vulnerability alone.

Construction of Intensity-to-Damage and Damage-to-Loss Curves

Vulnerability can be estimated from engineering simulations, observed event losses, analytical structural models, expert-elicited functions, or combinations of these. A useful construction begins by defining the random physical damage state D conditional on intensity I and asset taxonomy T. For damage states indexed by j, let q_j(I,T) denote the probability that the asset occupies state j and let m_j(T) be the expected loss-cost ratio assigned to that state. The expected ground-up loss ratio is then

E[LR | I,T] = Σ_j q_j(I,T) m_j(T),

provided the state definitions are mutually exclusive, collectively exhaustive, and the consequence ratio definition is consistent across states. The identity follows directly from the law of total expectation. It does not presume that the probabilities can be estimated from a single simulation or historical storm.

When the underlying fragility functions supply exceedance probabilities F_j(I,T) = Pr(D ≥ j | I,T), the probability of occupying exactly state j is F_j - F_{j+1}, with the highest state probability equal to its exceedance probability and the undamaged probability given by one minus the first exceedance probability. The exceedance functions must form coherent nested events: F_1 ≥ F_2 ≥ ... ≥ F_J. Crossing fragility curves can produce negative implied state probabilities and can invalidate a vulnerability calculation. An analyst should check coherence after fitting or interpolating fragility functions rather than assuming it from independent fits.

A fully hypothetical example makes this construction reproducible. At a selected hazard intensity, suppose the probabilities of no damage, slight damage, moderate damage, extensive damage and complete damage are 0.40, 0.30, 0.20, 0.08 and 0.02. Assign respective illustrative loss ratios of 0, 0.05, 0.20, 0.60 and 1.00. The expected loss ratio is 0.30×0.05 + 0.20×0.20 + 0.08×0.60 + 0.02×1 = 0.123, or 12.3% of replacement cost. The probabilities sum to one. For a hypothetical $1 million replacement-cost building, the expected ground-up loss is $123,000 under these assumptions. The example shows how a loss ratio can be computed from damage-state probabilities; it is not a calibration of any published GEM class.

A second hypothetical calculation illustrates dispersion. Let the loss ratio take only two values, zero with probability 0.9 and one with probability 0.1. The expected ratio is 0.1, but the probability of complete loss is 10%. A smooth deterministic curve assigning a fixed 0.1 loss ratio at the same hazard intensity has the same mean but a zero probability of complete loss. Portfolio exceedance tails differ sharply under the two representations, particularly when losses share a common catastrophe event. Conditional variance and a distributional family must therefore be treated as scientific inputs rather than optional aesthetic detail.

The OpenQuake vulnerability input specification illustrates how numerical implementations bind intensity measure levels to mean loss ratios, coefficients of variation, and distributional choices. This is also a provenance constraint: a code string identifying building taxonomy and a file denoting structural loss cannot be treated as interchangeable with an occupancy category or a contents-loss function. The same physical intensity may give separate structural, nonstructural, contents and business-interruption losses, and any aggregation must avoid double counting.

An alternative structural engineering approach uses the distribution of engineering demand measures such as interstory drift or floor acceleration rather than a preselected set of broad damage states. The documented GEM modeling workflow derives vulnerability functions using structural response, fragility, consequence and component-specific loss relationships. Its numerical framework depends on assumptions about construction types, ground-motion records, building height, code design and damage thresholds. Transferring the resulting curve to a different construction stock requires a documented mapping and independent evidence about those variables.

Peril Specificity, Exposure Taxonomy, and Engineering Transfer

Earthquake, flood, hurricane wind, coastal surge and wildfire cause damage through different physical mechanisms. The intensity variable appropriate to one peril can be meaningless for another. A seismic vulnerability curve may use spectral acceleration or another calibrated ground-motion parameter, whereas river flood loss models frequently relate damage to inundation depth relative to a specified floor datum. Wind damage may depend on sustained or gust wind speed, direction, duration, envelope components and local construction practice. Multi-peril modeling cannot establish comparability by merely drawing curves on the same horizontal scale.

Flood loss is particularly sensitive to the definition of elevation. Water depth measured relative to exterior ground differs from depth above the first finished floor, and structure foundations and basements alter damage at a given local flood level. The FEMA Hazus risk-data program includes general building stock, flood, hurricane and earthquake modeling information for baseline community risk assessment. Baseline data make modeling possible where individual building information is sparse, while creating aggregation and classification uncertainty. Occupancy class, foundation, stories, first-floor height, component vulnerabilities and replacement cost require local assessment when used for a specific insured asset.

A simple illustrative flood comparison can be expressed without claiming an empirically fitted flood-damage function. At the same outside ground-water elevation, Building A with a floor at ground level may experience substantial interior inundation, while Building B with a raised occupied floor may experience much less interior damage. The relevant intensity at the asset must be computed relative to its own datum before either curve is applied. This is a basic reason why geocoding, elevation resolution and exposure-inventory quality influence losses even with an unchanged vulnerability library. It also explains why an apparently incorrect modeled loss cannot always be attributed to the damage curve itself.

The FEMA Hazus software resources are relevant because loss analysis combines inventories, hazard extent and intensity and damage functions. FEMA model generations and supported hazard modules change over time. Research results must therefore specify the exact Hazus generation and technical method used, and retain a compatible archival record when newer software changes the supported function. A curve obtained from a previous implementation should not be presented as automatically current after a platform migration.

An earthquake building taxonomy is likewise an evidence-based mapping, rather than a harmless code label. The GEM taxonomy discussion distinguishes building material and lateral system, code level and height, among other features. A portfolio database that contains only a broad occupancy description such as "commercial" cannot uniquely determine those structural characteristics. A modeler must use supplementary inspection, probabilistic class allocation, local engineering inventories or transparent assumptions. Classification uncertainty may be more important than small differences between two otherwise similar curves.

The geographical transfer problem is acute in international insurance portfolios. Construction eras, local building codes, retrofits, labor costs, repair practicability and inflation differ. A published vulnerability function derived for one structural typology and region may serve as a scientific starting point for another, but its parameter transfer requires justification and out-of-sample testing. When economic consequence values are expressed relative to a local replacement-cost base, the choice of currency and price index must be documented. Actual historical insurance payments also reflect policy limits, deductible structures and claim settlement practices. Without normalizing these, fitting a loss-ratio curve to paid claims may mix vulnerability with contractual effects.

Calibration, Validation, and the Observation Process

Calibration relates a model curve to observed or simulated outcome data. Validation tests its predictive agreement on distinct observations and checks whether the underlying quantities were measured consistently. Engineering model calibration may use nonlinear dynamic analysis and component response; empirical calibration may use post-event damage surveys, property inspection, engineering repair records, event footprints and claims. These evidence types are complementary but have different biases. A simulated building is conditioned on an idealized construction specification; a claims database is conditioned on the presence of insurance, reporting incentives, policy terms and the insurer's settlement process.

For an observed property i, let L_i denote an independently estimated ground-up repair cost and V_i its appropriate replacement value at the hazard date. The empirical damage fraction is Y_i = L_i/V_i when V_i is positive and both measures follow the same valuation scope. If a dataset instead supplies paid insurance amounts, observed Y_i may be censored by contractual limits and affected by deductibles, excluded losses or non-repair settlement; treating those paid amounts as physical loss fractions creates mechanical underestimation. Large losses may also be censored by insufficient cost documentation or post-disaster rebuilding barriers. These defects must be analyzed before training a curve.

The performance metrics should reflect the intended application. Mean squared error evaluates a conditional mean in the observed distribution and is sensitive to large loss ratios. A log score or proper probability score is needed when the output is a full conditional distribution. Calibration by intensity band tests whether observed loss proportions align with predicted values over exposure cohorts. Reliability should be assessed by construction taxonomy, peril intensity, occupancy, geography and vintage. Aggregate observed-to-expected ratios may conceal opposing errors in low- and high-intensity zones, especially if the observed event sample is imbalanced.

Validation design must account for event clustering. Multiple buildings damaged by the same flood or earthquake share hazard reconstruction, weather, soil, construction and repair market conditions. Random splitting of individual properties from one event between training and testing can leak shared event effects and overstate generalization. A stronger out-of-event test holds out entire catastrophe events, and a geographic transfer test holds out regions or building regimes. Temporal tests retain the data and price assumptions actually available at each forecast origin. The OpenQuake calculation workflow distinctions reinforce why the validation object must specify whether the model predicts damage distribution for a scenario, losses for a specified event, or losses under a stochastic event set.

The reporting process also determines what an empirical "ground truth" means. Building-level surveys may prioritize highly affected blocks, accessible structures or severe visual damage; insurers may inspect only reported claims; owner repair records may omit uninsured work or deferred reinstatement. A sampled set of visibly damaged buildings will not generally identify a regional damage-frequency curve without accounting for intact structures. Similarly, a dataset enriched with high-value insured properties may have a different construction and mitigation profile from the overall building stock. The sampling frame, weights and non-observation mechanisms should be retained in the evidence record.

For rapid earthquake impact estimation, the USGS PAGER methodology combines early shaking and exposure information to produce broad impact ranges that evolve with new data. Its publication stresses uncertainty in initial ground shaking and the limitation of approximate early losses. This use case differs from a year-end insurer reserving analysis requiring contract-level coverage, loss adjustment and payment information. Forecast revisions should be preserved instead of retroactively rewriting the original risk estimate; a later accurate PAGER value cannot prove the precision of the earlier warning.

Financial Application and Portfolio Risk Aggregation

The physical vulnerability function produces a ground-up loss distribution. Insurance contracts transform this outcome through legally operative terms. The Oasis financial module distinguishes ground-up losses, insured losses under direct contracts and net losses after reinsurance. These are separate financial perspectives. For a simple hypothetical occurrence policy with a $10,000 deductible and a $300,000 per-occurrence insurer limit, a ground-up loss of $123,000 produces $113,000 in insured indemnity if all damage is otherwise covered and the deductible is applied before the limit. The same rule applied to a $600,000 covered physical loss produces $300,000 if that limit caps the insurer's payment. Exact contract order and aggregation may differ in real policies.

One consequence is the nonlinear effect of vulnerability uncertainty on covered loss. With a fixed deductible d and limit u, an illustrative insurer payment for a covered physical loss L is min(max(L − d, 0), u). The expected payment E[min(max(L − d, 0), u)] generally cannot be calculated by applying the contract function to E[L]. In particular, a mean loss ratio curve without a dispersion specification cannot reproduce the correct frequency of deductible exhaustion or limit attachment. For high layers and reinsurance, tail shape and common event dependence become indispensable.

Portfolio losses depend on how individual asset vulnerabilities respond to a shared catastrophe. The OpenQuake event-based risk calculation samples losses conditional on event ground-motion fields and can incorporate the correlation structure of vulnerability for similarly classified assets. Treating all conditional losses as independent can understate common building response and tail concentration; treating them as perfectly correlated can overstate it. The appropriate assumption depends on hazard-field spatial correlation, shared engineering characteristics and residual vulnerability variability. A distinction between correlated hazard and correlated damage residuals helps avoid counting the same source of dependence twice.

A simple synthetic example highlights the portfolio issue. Two buildings each have a 10% chance of complete loss under a particular event. If losses are conditionally independent, the probability both are destroyed is 0.01. If the two damage indicators are perfectly correlated, that probability is 0.10. Both models assign the same 10% marginal loss probability to each building, yet the probability of aggregate loss of both replacement values is ten times as large under perfect dependence. The example cannot substitute for calibrating actual spatial correlation, but it proves why asset-level curves alone do not determine portfolio catastrophe tails.

Insured vulnerability is also shaped by coverage availability. The EIOPA protection-gap dashboard combines modeled hazard risk with reported historical insured and economic losses, distinguishing current risk and historical experience across European markets. A protection gap emerges from both hazardous damage and the insurance arrangements that determine which assets and events are covered. A vulnerability model estimating physical flood loss for an entire building stock cannot be treated as a forecast of insured loss without insurance penetration and contractual mapping.

The EIOPA catastrophe data hub provides exposure and incurred-loss information from a sample of European insurer groups and undertakings. Its sample is not a census of the entire market. Validation against that dataset therefore needs to state whether the target is the reporting sample, the insured market or the region's full economic loss, and how the underlying building types and insured values differ. Missing insurance may be correlated with catastrophe vulnerability, income, construction conditions and geographic hazard. The comparison is sensitive to selection before any vulnerability-curve parameters are changed.

The Oasis open data standards support separation of risk exposure, model results and financial processing. Standardization of field labels and schemas is valuable, but semantic consistency must be inspected within each portfolio. A field called "building value" may represent reconstruction cost, insured sum, book value or an estimated market valuation. A vulnerability calculation based on replacement cost becomes internally inconsistent if a different value base is silently substituted during loss aggregation.

Uncertainty, Scientific Synthesis, and Research Implications

Vulnerability-curve uncertainty has several layers. Epistemic uncertainty concerns the choice of engineering model, functional form, structural taxonomy, hazard intensity measure, consequence costs and transfer assumptions. Aleatory variability describes random damage outcomes conditional on a specified model and intensity. Measurement uncertainty concerns estimated hazard fields, geolocation, inventory, building height, construction type, repair estimates and claim records. Their meanings influence how a risk assessment should sample or compare them. Repeating stochastic events under a single fixed vulnerability library estimates variability conditional on that library; it does not automatically represent uncertainty over alternative plausible libraries.

A rigorous sensitivity study varies both vulnerability functions and their inputs, preferably using documented sources and parameter ranges. In a lognormal fragility relationship, the median threshold controls where a probability curve rises, while the dispersion controls its transition width. A change in both may leave an observed mean loss similar at one intensity while yielding substantially different losses at higher intensities. Analysts should evaluate multiple intensity bands, occupancy and construction strata and tail outcomes. Interpolating a curve between calibrated intensities can be useful; extrapolation far outside the data range requires engineering support and explicit uncertainty.

Version control is a scientific requirement. The GEM vulnerability modeling framework and FEMA Hazus data releases make underlying typology and methodology changes observable to model users. For every official risk study an archive should record hazard dataset and version, vulnerability function library and class mapping, inflation and replacement-cost source, exposure inventory, contract transformation rules, random seed and simulation settings, software engine version and analytic date. If any dependency changes, a comparison of catastrophe loss estimates should isolate the contribution of that change before attributing differences to increasing physical risk.

An original integrated interpretation follows from the evidence. Vulnerability curves are conditional transformations within a broader probabilistic and legal loss system. They are scientifically meaningful only when the intensity variable, asset class, damage-state definitions, cost base and conditional distribution are jointly specified. Their practical insurance value depends further on how they interact with contract terms and portfolio event dependence. The strongest validation strategy evaluates each link separately, uses actual out-of-event and out-of-region evidence, and preserves uncertainty rather than collapsing complex claims experience into one average damage percentage.

This interpretation yields a clear research agenda. Multi-peril vulnerability analyses should test the transferability of physical intensity and construction classes, the stability of repair-cost ratios under post-disaster inflation, the alignment of public damage observations with insured claim populations, and the effects of nonlinear deductibles and shared-event correlation on high quantiles. Studies should also compare models under frozen event datasets and harmonized exposure and contract inputs. Such design isolates vulnerability-model skill from unrelated improvements in hazard fields or asset inventory quality.

Empirical literature, model scope, and unresolved measurement differences

The relationship between a formally defined fragility model and an empirically valid insurance loss estimate is central to global catastrophe risk assessment. The published global seismic risk model of Silva and colleagues (2020) combines a probabilistic hazard model, geographically distributed building stock and building typology, and engineering fragility and vulnerability functions within OpenQuake. This architecture can generate different statistical measures, including average annualized loss and aggregated loss for specified return periods. Each input carries measurement uncertainty. Better spatial precision in ground shaking does not necessarily resolve the uncertainty in the strength of masonry, construction vintage, replacement cost or the share of a national housing inventory represented in its building taxonomy.

A global asset inventory may be calibrated using national building counts, census characteristics or occupancy classes and then disaggregated to locations where physical attributes are only partially observed. Such allocation can be appropriate for global comparisons, but it does not establish the detailed composition of a single insurer's insured portfolio. Even when each building type's vulnerability curve is credible, assigning that curve to the wrong building class shifts the conditional loss distribution. The global seismic modeling study illustrates the need to treat hazard, exposure and vulnerability as distinct technical inputs that require their own tests. In financial reporting, a change in annualized loss between model revisions must be decomposed before being interpreted as a change in actual physical hazard.

An influential review of flood loss modeling by Gerl and colleagues (2016) compiled 61 professional or published sources and reviewed 47 in detail, yielding nearly one thousand vulnerability relationships. The review found major differences in geographical and sectoral scope, functional specification, inputs, and loss denominators, with explicit validation accompanying only about half of the underlying modeling studies. That conclusion describes the observed literature inventory and the authors' definitions. It does not establish that exactly half of all current operational catastrophe models are invalid or unvalidated. Its scientific implication is that comparing loss curves requires a common observational basis and independent assessment of their scope before deciding which model predicts a target insurance portfolio better.

For flood risk, benchmarking curves from distinct studies can require translating monetary amounts into a consistent price year, harmonizing depth measured above ground or first floor, aligning structural with contents damage, and restricting observations to common occupancy classes. An apparent reduction in residual error after fitting a more complicated curve could simply reflect an unacknowledged change from average household value to replacement-cost value. Empirical comparability is best established through a common evaluation set of observed building and hazard characteristics, accompanied by uncertainty bounds and error decomposition. A model whose parameters were estimated for residential masonry structures cannot be evaluated as if the same curves had been constructed for industrial equipment, infrastructure or business interruption.

Peril-specific and version-specific scientific reproduction

The United States Hazus 7.2 release notice identifies June 4, 2026 as the software release date and documents the addition of the U.S. Geological Survey's earthquake ShakeMap functionality. It also explains the discontinuation of support for the former Hazus 6.1 platform when ArcGIS Desktop was retired in March 2026. This dated primary record is material to scientific reproducibility. The phrase "Hazus earthquake model" can refer to substantially different combinations of hazard inputs and software generations. A documented impact analysis should identify the actual software and vulnerability library used, separate ShakeMap update effects from vulnerability-parameter changes, and retain the original event data needed to reconstruct an earlier estimate.

In the wind-peril literature, the Caribbean wind-damage development study by Vickery and coauthors (2023) addresses a geographic and construction context distinct from a continental United States baseline. Tropical cyclone wind fields, terrain, building envelope connections and roof assemblies may differ. A loss function developed for one territory and building stock may be adapted to another only after documenting the hazard wind-speed convention, construction-era taxonomy, mitigation features, material systems and empirical adjustment. This is particularly important for coastal areas where windstorm and surge may affect the same insured asset and where claims can combine different physical causes subject to different coverage provisions.

The existence of several peril-specific curves for an asset does not imply that their damage fractions may simply be added. Earthquake and subsequent fire, hurricane wind and storm surge, river flood and surface-water inundation, or compound wind and rainfall can cause overlapping physical damage. If a roof fails before heavy rain, incremental interior damage depends on that earlier state. The combined expected loss requires either a joint hazard-damage representation, sequential transition rules, or a defensible allocation of incremental damage. If two hypothetical models predict 30% wind damage and 25% surge damage to a building, adding them to claim 55% combined damage double-counts overlapping repairs when the same component is damaged by both mechanisms. A rigorous multi-peril model identifies the underlying asset components and exposure states before estimating total recovery costs.

A sequential illustrative calculation clarifies the issue. If an initial hazard destroys 30% of replacement value, and a second hazard subsequently damages 25% of the remaining undamaged value, the combined physical loss fraction is 0.30 + (1 − 0.30) × 0.25 = 0.475, or 47.5%, conditional on these synthetic assumptions. A naive sum gives 55%, an overstatement of 7.5 percentage points. The formula does not purport to describe real sequential wind and surge events, for which structural interactions and repair costs may be more complicated. It demonstrates the necessity of clearly specifying overlap and the conditional base of subsequent damage functions.

Distributional vulnerability and portfolio dependence

The critical problem addressed in Silva's correlation study (2019) is how variability in conditional seismic loss ratios and correlation among assets affects aggregate risk. The original study compares three simulated building portfolios and evaluates the consequences for annualized and exceedance losses. Portfolio vulnerability requires more information than asset-level mean loss curves. Residual vulnerability correlation measures the relationship between asset damages after conditioning on the chosen hazard field; it should be modeled separately from the spatial correlation already encoded in that field. Double inclusion of common-cause variability can distort risk capital or reinsurance-layer prices.

For two buildings, consider a synthetic event in which each has a 10% conditional probability of complete physical destruction. With independent binary damage indicators, the joint probability of destruction is 0.1 × 0.1 = 0.01. With perfectly correlated indicators, the joint probability is 0.10. The mean number of destroyed buildings is 0.20 in either representation. However, the distribution of aggregate losses differs, even though the two marginal vulnerability functions are identical. In the first model, the chance of at least one destroyed building is 0.19; in the second, it is 0.10. A reinsurance layer covering only aggregate loss above the value of one building is sensitive to the joint-destruction probability, not merely to the mean.

A comparison of mean loss and catastrophic tail risk must use the same units and aggregation boundaries. The nominal value of a building may incorporate land, whereas insured reinstatement cost excludes or limits it. An insurer's exposure inventory may aggregate contents under the same location as structure. A catastrophe model generating separate structure and contents losses must apply component-specific terms and then the contract's actual aggregation rules. If a carrier sums modeled building values from one source and insured amounts from another without validating their meanings, modeled exceedance curves may appear more precise than their underlying financial inputs justify.

Probabilistic risk calculations and convergence

The probabilistic earthquake loss analysis by Silva (2018) evaluates the sensitivity of seismic risk outcomes to event-based simulation convergence, epistemic logic-tree choices, loss exceedance construction and mapping, using a Lisbon-region application. In a simulated event set, sampling additional events reduces Monte Carlo uncertainty for a fixed model, but may leave epistemic differences among plausible hazard or vulnerability models unresolved. A convergence graph should therefore report both random-simulation error within each selected model branch and the dispersion across model branches.

For independent simulated annual aggregate losses A_1, ..., A_n with finite variance, the arithmetic sample mean estimates a model-conditional expected annual loss. Its approximate standard error decreases with sample size according to the sample standard deviation divided by the square root of the number of independent annual draws. This standard error is conditional on the assumed event rates, financial terms, exposure inventory and vulnerability rules. It does not include uncertainty about whether a fragility curve fitted in one region is valid for buildings elsewhere. Presenting a very narrow simulation confidence interval as the full uncertainty around catastrophe risk would therefore mischaracterize the scientific evidence.

A numerical thought experiment separates the two sources of uncertainty. Suppose model branch A has a simulated expected annual ground-up loss of $10 million and branch B has $14 million, with each branch's Monte Carlo numerical error already negligible. Assign equal illustrative epistemic weights to the two models. The weighted mean is $12 million; variation between the model means remains $2 million around that mean. Increasing simulated years within either branch cannot eliminate that model choice dispersion. If one branch additionally assumes higher insured building values, its difference needs an exposure-value attribution before being described as vulnerability uncertainty.

Annual-average loss, occurrence exceedance probability and aggregate annual exceedance probability are also different statistical functions. Occurrence exceedance concerns a single-event loss above a threshold, whereas annual aggregate exceedance concerns the sum of all event losses within a year, including multiple events. A catastrophe event catalog and its stochastic frequency specification determine how these distributions relate. A model reporting a one-in-200-year occurrence threshold is not automatically reporting the one-in-200-year annual aggregate threshold. The difference can matter for reinsurance programs, annual aggregate limits and insurer capital analysis.

Insurance contract nonlinearities and actuarial sensitivity

The map from physical loss to insured loss is nonlinear because contracts often include deductibles, limits, policy aggregations, salvage and recovery rules. Consider a wholly synthetic property with $500,000 of replacement value and an all-covered single-event loss that is zero with probability 0.8 and exactly $500,000 with probability 0.2. The mean ground-up loss is $100,000. Under a $25,000 deductible and a $250,000 insurer-payment limit, the insurer pays zero in the first state and $250,000 in the second, giving an expected indemnity of $50,000. Applying the same insurance formula to the $100,000 mean loss would produce $75,000. Thus a mean-only vulnerability curve would overstate expected insurer payment by $25,000 in this constructed example. The calculation proves that the insurer's contract transformation generally must be applied to the conditional loss distribution before expectation is taken.

For an insured asset with loss L, deductible d and upper insurer-payment limit u, the simplified indemnity is g(L)=min(max(L−d,0),u). In general, E[g(L)] differs from g(E[L]). The size and direction of the difference depend on the loss distribution, position of the deductible and limit, and the meaning of covered loss. When a policy instead has a limit on ground-up loss before deductible or a percent-of-value deductible, a different transformation is required. Reinsurance may impose attachment points and occurrence versus annual aggregate limits above the primary policy structure, creating additional nonlinear layers.

The same principle applies to model validation. A model can overestimate moderate physical losses while underestimating the frequency of extreme losses, yet produce an apparently accurate average paid loss under a narrow contract and valuation sample. A validation study should compare ground-up loss distributions before and after applying contracts, and should distinguish physical model errors from policy-data errors. An insurer evaluating an alternative catastrophe model needs matched exposures, identical financial terms and the same simulated events to isolate the incremental predictive merits of the vulnerability library.

Scientific reproducibility protocol and validation strategy

A defensible international comparative assessment can record a versioned input matrix with fields for peril and intensity measure, units, exposure geography and taxonomy, insured building attributes, replacement-cost base, construction year, hazard version, vulnerability model version and variance specification, contractual financial terms, observation outcome, date, and verified source. For every event, the audit should preserve both preliminary and revised hazard observations and treat claim development as a time-dependent process. An event loss estimated immediately after a catastrophe may rely on coarse ShakeMap or rainfall products; the subsequent insured claim total may reflect months of inspection, reopening and settlement. Scientific model skill should be assessed against information available at the prediction origin, or the use of retrospective revisions should be explicitly declared.

The flood-model benchmarking review emphasizes harmonizing variables, functional forms and validation scope across available flood loss curves. A strong validation protocol holds out whole events to avoid sharing hazard reconstruction errors between training and evaluation; also holds out regions, construction classes or event intensities to test transportability. Its metrics should include predictive uncertainty and tail outcomes, rather than only an aggregate mean-square error. Failure in an extrapolation test is evidence about scope, not necessarily a failure of the physical modeling assumptions within their calibrated range.

Finally, model governance has a substantive actuarial dimension. A technically robust vulnerability function may still be unsuitable for a portfolio if the firm's exposure taxonomy is unreliable or the policy forms exclude the modeled peril. Conversely, a relatively simple mean curve can be useful for an explicitly bounded screening application when uncertainty and transfer limitations are documented. The appropriate institutional question is which scientific and financial quantities must be correct for the intended decision. This disciplined question connects the engineering study of vulnerability to insurance pricing, aggregation, risk transfer, regulatory capital and protection-gap assessment without converting a single statistical performance score into a universal model verdict.

Scientific uncertainty decomposition and empirical calibration design

A rigorous validation dataset should follow the entire sequence from individual insured building to repair observation, rather than presenting a loss ratio divorced from its construction and reporting history. Its minimum analytical keys comprise the property identifier, construction and occupancy classification, hazard event identifier, recorded location, first-floor or terrain elevation where relevant, hazard intensity and version, insured structure and contents values, date of observation, source of the repair estimate, and policy response. The property identifier must remain stable across claim reopening, adjustment revisions, ownership transfers and duplicate inspections. Geographic privacy controls may prevent public release of address-level records; a reproducible scientific study can still publish its data dictionary, transformation rules, sampling design and aggregate diagnostics without identifying insured households.

A useful outcome is the deviation of observed repair fraction from the predicted mean loss ratio within a prespecified intensity and building class. Such deviations should be summarized separately by event, peril, construction group and region. A model whose residuals average approximately zero in one pooled sample may systematically overestimate residential loss in one construction class while underestimating commercial loss in another. If the model is intended to support insurer tail risk, evaluation should include extreme-event observations, loss-ratio dispersion and joint portfolio aggregates where actual independent loss data permit. A calibration curve fitted entirely to the same storm's post-event surveys represents an in-event fit; the stronger test holds out the storm, its hazard reconstruction and its damage observations.

The flood vulnerability benchmarking study offers a practical conceptual lesson: curves developed from different countries or sectors cannot be ranked reliably before their variables are harmonized. Two empirical data sets with identical reported flood depth may differ because one measures water above exterior ground while the other measures water above the first inhabited floor. Some models report the cost to reconstruct a damaged portion, and others the entire building cost. Classification and data transformations should be recorded before comparing prediction errors. Changes made during harmonization should be treated as part of the scientific method rather than as invisible data preparation.

Epistemic uncertainty can be represented by a set of alternative coherent vulnerability curves with transparent weights or by parameter distributions reflecting estimation limits. Aleatory loss variability conditional on a chosen model remains distinct. In a simple sensitivity experiment, consider a building with replacement value $1 million and three plausible conditional expected loss fractions of 8%, 12%, and 16%. If the curves receive equal illustrative model weights, the mean expected ground-up loss at that hazard intensity is $120,000. Its epistemic between-model standard deviation using probability weights is approximately $32,660, calculated from the square root of the average squared deviation of the three possible dollar losses from the common mean. This calculation does not supply conditional damage variability within each curve or the probability of the underlying hazard intensity itself. Those are separate components of uncertainty.

A genuine international analysis should distinguish the frequency of exposure to damaging physical intensities from vulnerability conditional upon those intensities. Flood-protection works may reduce the frequency of inundation in particular areas, while building retrofitting may change the damage distribution conditional on inundation or wind speed. A change in annual aggregate losses can originate from either mechanism, from increasing insured value or from claims inflation. Reconstructing the difference between two risk estimates is most informative when the analyst freezes all but one input group at a time. A simple decomposition may be path-dependent when hazard and vulnerability interact, so the analyst should report the chosen sequence or use a clearly defined averaging scheme.

Insurance governance, disclosure, and scope of scientific validity

The use of a catastrophe model inside an insurer introduces additional conditions beyond the engineering evidence. Governance must identify who maintains the vulnerability library, its relationship to the primary risk engine, its data licensing and version, relevant exclusions in policy files, and the authority responsible for approving material parameter changes. A model may be integrated through standard file schemas but still be unsuitable for a particular insurer because it lacks the structural detail required to classify local building stock. A reproducible review requires a mapping from each exposure record to its assigned vulnerability and an exceptions report for unresolved, unknown or conflicting building classifications.

The insurer should retain both modeled and observed loss perspectives. A ground-up loss estimate represents modelled repair or replacement cost before the contract; an insured gross result adds insurer-specific deductibles, limits and claims response; a net result reflects applicable reinsurance. The Oasis financial module documents these separate calculations, and a released internal model should preserve the order of operations, especially when per-risk and occurrence aggregations interact. A risk assessment that is technically accurate at ground-up level can give a misleading underwriting answer if values, policy conditions or cessions are incorrectly recorded.

An assessment of geographical protection gaps requires a further population distinction. The EIOPA catastrophe data hub represents specified sampled European undertakings and event types. It can assist calibration or interpretation of the insured portfolio population represented in that sample. It cannot automatically identify the full value of uninsured or underinsured buildings in the same geography. Comparing a modeled regional economic loss to a database of incurred insured losses requires explicit accounting for policy penetration, covered perils, reporting dates, insured value distributions and changing monetary prices. The strength of the conclusion comes from the correctness of these population and contract definitions.

Finally, an actuarial risk analysis should record what was known at the time of the decision. Immediate post-event estimates are made before damage surveys and claims mature, while insurer ultimate losses may become observable only after long claim development. A later accurate estimate derived from revised hazard fields or complete claims does not demonstrate that the earlier model possessed equivalent foresight. Prospective evaluation freezes its input data at the prediction origin, retains later revisions as a separate series, and tests performance across future events. Such disciplined provenance supports model improvement, capital estimation and public interpretation with a clear boundary between observed evidence, scientific inference and hypothetical example.

Conclusion

Property insurance catastrophe model vulnerability curves convert specified hazard intensities and asset characteristics into conditional physical and economic loss distributions. Their foundations in engineering fragility, damage-state consequence assignment, empirical calibration and stochastic event modeling are well established in the reviewed institutional literature. Their appropriate use requires scientifically coherent state probabilities, transparent loss ratio denominators, peril-specific physical inputs, validated construction taxonomies and an explicit distinction between ground-up economic loss and financial insurance obligations.

The analysis establishes that curve precision alone cannot determine portfolio risk. Exposure data quality, cost inflation, reporting selection, spatial and residual dependence, contract terms and historical version changes affect the outputs of a catastrophe model. An international research program can compare these mechanisms through existing scientific methods, with documented sources, reproducible calculations and substantive uncertainty analysis. The resulting knowledge supports more transparent catastrophe risk estimates and more defensible application of vulnerability science to insurance without assuming that public damage rates directly equal insured claims.

References