RESEARCH DOSSIER
Property Catastrophe Accumulation and Spatial Correlation
Abstract. Property catastrophe accumulation arises when an insurer's geographically distributed commitments can suffer losses from a common damaging event, a connected sequence of events, or mutually dependent hazards. Its quantification requires the joint representation of spatial exposure, hazard footprints, engineering vulnerability, insurance terms, event occurrence, and financial recoveries. This monograph examines the scientific foundations of accumulation from insured-site records through spatial random fields, correlated loss distributions, reinsurance, and solvency. It distinguishes a concentration of insured values from dependence among damage indicators, from the dependence of financial claim amounts and from tail capital. A spatially extensive peril can generate high portfolio loss despite modest individual marginal loss probabilities; location precision, building attributes, shared infrastructure, event definitions, and policy conditions determine the conversion. Original transparent numerical examples show that identical marginal risk and expected aggregate loss can produce different extreme-loss quantiles and that identical annual gross loss can yield different net loss when occurrence structure differs. The study critically compares earthquake, tropical cyclone, flood, wildfire, and severe convective storm mechanisms, examining data standards, catastrophe modeling workflows, geographic selection bias, model validation, climate nonstationarity, and international supervisory practices. The principal synthesis is an auditable chain linking individual location and contract, joint hazard realization, joint damage, covered loss, event aggregation, recoveries, and capital. The analysis integrates established probabilistic methods and independently identified primary sources; it establishes neither a new named theory nor a proprietary catastrophe model. Numerical portfolios are hypothetical, and reported industry figures retain their original geographic, temporal, and accounting boundaries.
Keywords: Property insurance; catastrophe accumulation; spatial correlation; geospatial exposure; catastrophe models; event footprint; common shocks; ground-motion fields; natural hazards; geocoding uncertainty; vulnerability functions; conditional dependence; tail risk; occurrence exceedance probability; aggregate exceedance probability; probable maximum loss; reinsurance; solvency capital; model validation; climate nonstationarity.
Introduction
The underwriting unit in property insurance is usually an identifiable insured interest, but the unit of catastrophe production is an event whose footprint may cross thousands of insured locations and several administrative boundaries. Consequently, the risk of an insurer's book cannot be inferred simply from its number of policies, the distribution of premiums among regions, or the average expected loss of separate properties. A shared coastal windfield may strike insured dwellings, factories, and warehouses; an earthquake may generate correlated shaking across a metropolitan region; precipitation over a river basin may drive both local and downstream inundation. A second event before recovery may further change damage, costs, claims handling, and reinsurance response.
The available empirical record shows both the importance and the limits of insurer-reported catastrophe losses. Swiss Re Institute's 2025 natural catastrophe assessment estimates global insured natural catastrophe losses at USD 107 billion in 2025. The estimate describes insured natural catastrophe losses aggregated under the source's event criteria and excludes much of the economic damage borne outside insurance. It is informative about the scale of realized catastrophe claims, yet supplies no universal empirical correlation coefficient applicable to a particular insurer's properties. Swiss Re's account of the losses of 2025 also identifies materially different loss experiences among wildfires, convective storms, floods, and other perils.
The underlying environmental science concerns physical processes. The Intergovernmental Panel on Climate Change (IPCC, 2021) examines simultaneous and sequential hazards in its assessment of compound extremes, including spatially compounding events. This scientific category matters because an insurer may be vulnerable both to the geographical reach of one event and to combinations of events that affect separate regions during a common financial period. Climatic attribution at peril and region level, however, must be distinguished from changes in insured exposure, reconstruction costs, policy penetration, and contractual coverage. Financial accumulation is a modeled consequence of all these components.
The subject of this Research Dossier is the loss-bearing property portfolio. Its principal Science is Property Insurance, with established relations to Catastrophe Risk and Solvency and Capital. Approved related Science also includes Automobile Insurance, Agricultural Insurance, and Marine and Aviation Insurance, whose catastrophe-exposed interests can overlap a physical footprint while retaining distinct contract-specific losses. Cross-line dependence is examined here only insofar as it changes the interpretation of property accumulation.
This monograph advances six connected tasks. It identifies the legal and physical units of observation; establishes a measurable framework for geospatial accumulation; derives dependence consequences for event and annual losses; relates the loss distribution to retained risk and capital; tests the evidentiary limitations of the available data; and synthesizes comparative findings across perils and institutions. It uses documented source interpretation, elementary probability identities, transparent synthetic arithmetic, and critical assessment of publicly described catastrophe-model architectures. It does not claim proprietary insurer exposure records, confidential contract wordings, a fitted global correlation function, or original catastrophe event backtesting.
Definitions, Historical Context, and Primary Source Foundations
The insured object and the catastrophe event
An insured property is a location-dependent asset or group of assets capable of producing covered financial loss under a particular insurance contract. An insured location is an address, coordinate, parcel, building, facility, campus, or declared risk unit linked to the exposure record. The recorded site may be a proxy for a geographically extended asset. An exposure amount can mean replacement value, declared sum insured, total insured value, probable recovery cost, business interruption value, or an exposure-specific limit. These values are not interchangeable. A portfolio may be aggregated on a location basis, a policy basis, a corporate group basis, or a legal-entity basis, each with a different denominator.
A catastrophe hazard is a physical process with a defined intensity variable, spatial domain, and temporal evolution. Wind speed, earthquake ground-motion intensity, water depth, hail size, heat and dryness, and fireline exposure are examples of hazard measures. A hazard realization is not itself a property loss: damage depends on construction, occupancy, defenses, maintenance, elevation, surrounding structures, and the peril's temporal path. Ground-up physical loss includes the direct damaged-asset cost. Economic loss may additionally include interruption, displacement, foregone production, cleanup, and indirect effects. Insured loss applies contract coverage, conditions, deductibles, exclusions, limits, coinsurance, and special clauses. Net insurer loss further accounts for ceded recovery, costs of recovery, and potentially counterparty default or timing differences.
Accumulation is the simultaneous or temporally connected materialization of loss potential across an identified portfolio. Spatial accumulation describes the geographic concentration of exposure susceptible to common hazard fields. Statistical dependence concerns the joint law of losses. Contractual aggregation determines whether individual damages constitute one occurrence, multiple occurrences, one policy aggregate, or recoverable losses under particular reinsurance layers. A single portfolio can be geographically concentrated but well protected against one specified peril, or geographically widespread but strongly dependent on one long-range event, common infrastructure system, or aggregate recovery limit.
The World Meteorological Organization's disaster-loss Atlas for 1970–2021 documents the international scale of reported weather, climate, and water-related disasters. The underlying disaster statistics contain reporting, definition, inflation, and coverage limitations. They support the historical significance of extreme events, not a direct reconstruction of an individual insurer's annual aggregate loss distribution. The appropriate source for an actuarial portfolio calculation is an exposure inventory and a hazard- and contract-consistent event representation, independently checked against observations.
Mathematical objects and observation levels
Consider a fixed valuation date and portfolio of insured locations. Let denote the location or uncertain geographical support of site , its relevant exposure valuation, and its contract characteristics. For a catastrophe occurrence with hazard field , define site ground-up loss , where denotes observed building characteristics and covers vulnerability and process uncertainty. The corresponding insurer-funded claim is , where is a contract-specific financial transformation. Event gross insurer loss is before treaty recoveries; annual gross insurer loss is for a specified treatment of claims, event attribution, and annual policy limits.
The dependence structure enters at several points. A common field affects many . Conditional damage fluctuations can be correlated beyond the shared field through common building stock or construction defects. The mapping introduces dependence through policy-level aggregate limits, shared insured values, or coverage terms. The occurrence set may exhibit clustered events; treaty limits connect event losses financially. Treating these operations as an unordered sum generally changes the result whenever caps, deductibles, reinstatements, or aggregate exhaustions apply.
For random variables with finite second moments, . This exact identity demonstrates why ignoring positive pairwise covariance understates variance. It does not imply that variance determines a catastrophe tail: severely skewed or mixture distributions may have similar variances and materially different exceedance probabilities. Embrechts, McNeil, and Straumann (2002) explain the limits of linear correlation as a measure of joint extremes.
A probability of exceedance belongs to a defined random quantity and observation period. An occurrence exceedance probability (OEP) curve concerns the maximum event loss in a year, (with in event-free years). An aggregate exceedance probability (AEP) curve concerns total annual loss . At threshold , OEP is and AEP is . Their return-period labels are reciprocal annual exceedance probabilities under a stationary annual experiment; a modeled 1-in-200 annual exceedance probability is 0.5% in one modeled year, not an event scheduled to occur every 200 years. Return periods depend on loss basis, current portfolio, treaty design, model edition, and distribution assumptions.
The evolution of catastrophe modeling as an auditable architecture
Historical hazard maps and engineering reconnaissance preceded integrated event-based loss estimation. Contemporary catastrophe modeling brings together event catalogs, spatial footprints, exposure records, vulnerability distributions, contract financial modules, and exceedance-probability calculations. FEMA's Hazus loss-estimation system provides a public U.S. example of geographical hazard and damage analysis. Its planning-related economic impacts are distinct from the insurer's policy-specific liability. The Global Earthquake Model community's OpenQuake event-based workflow expressly combines exposure, vulnerability, and stochastic earthquake events to calculate losses across assets and portfolios.
An important architectural achievement is transparency between model components. The Oasis Loss Modelling Framework publishes Open Exposure Data standards to reduce dependence on incompatible provider-specific records. Its financial module documentation distinguishes ground-up, insured, and net-of-reinsurance results. These stages form separate audit objects. Moving a model result from one stage to another without recording the terms and assumptions can alter accumulation estimates while leaving the mapped hazard unchanged.
Risk governance adds a second primary-source tradition. Lloyd's Realistic Disaster Scenarios framework uses specified catastrophe events to stress-test individual syndicates and the market, including sequences such as two Atlantic seaboard windstorms. Event scenarios make spatial concentration visible even where a fitted historical probability is contested. The International Association of Insurance Supervisors' Insurance Core Principles framework identifies enterprise-wide identification and management of connected risks as a supervisory concern. The scope and binding force of such expectations depend on national implementation and the regulatory jurisdiction of the undertaking.
Fundamental boundaries of inference
Five inference boundaries organize the research. First, a geographic heat map of values measures concentration and may reveal co-exposure, but it is not a probability distribution. Second, the correlation of physical hazard intensities is not automatically the correlation of damaged property values, because asset vulnerabilities and repair costs differ. Third, joint damage is not automatically joint insured loss; a peril exclusion may remove one component and a shared deductible may alter another. Fourth, modeled insurer claims are distinct from timing and recoverability of reinsurance proceeds. Fifth, a modeled solvency quantile is a conditional result of its scenario generator, data, and accounting basis, not proof of physical certainty.
The analysis is internationally comparative. Its legal examples concern named regimes only. Institutional datasets differ in coverage and purpose: NOAA cyclone records identify tracked meteorological events, earthquake engineering datasets quantify recorded or modeled shaking, European protection-gap summaries compare economic and insured losses, and insurer catastrophe models calculate portfolio-specific distributions. Joining datasets requires explicit keys, event dates, valuation and inflation conventions, spatial coordinate systems, and loss definitions.
Geospatial Accumulation
Exposure inventory as a scientific measurement
The portfolio location record is the primary observational foundation of geospatial accumulation. At minimum, the analyst must distinguish location identifier, full address, geocode coordinates and quality, building footprint, insured interest, occupancy, construction and height, year built or retrofitted, flood elevation or earthquake site class where relevant, replacement value, currency, valuation date, policy limits, deductibles, coverage periods, and the insured legal entity. Property-level characteristics can be uncertain even when street coordinates are exact; industrial sites can extend beyond a single point, and infrastructure networks demand line or area geometries. The exposure database also needs stable identifiers for duplicate detection, shared insured complexes, policy endorsements, and changes over time.
A practical geospatial accumulation map overlays an exposure layer with peril-specific hazard layers and selected buffers, footprints, or catastrophe scenarios. The result can be a table of insured value within a flood zone, a wildfire interface, a wind corridor, or an earthquake shaking band. Two equivalent exposure totals have different implications if one is concentrated in a single watershed while the other lies across hydrologically distinct catchments. Conversely, distance alone can be misleading if a tropical cyclone, atmospheric river, earthquake, or synoptic storm drives hazard over a large territory. Peril-specific connectivity is more informative than universal distance thresholds.
Geocoding precision must be recorded as a statistical attribute. Rooftop coordinates, interpolated street addresses, postal-code centroids, municipality centroids, and unidentified addresses do not describe equivalent locations. A building placed at a postal centroid may move from a floodplain to a higher-elevation road or from one side of a firebreak to another. Positional error can be systematically related to locality, property size, and economic value. A representative source of exposure database interoperability is the Oasis Open Exposure Data specification, which permits structured exchange while leaving the quality of the underlying building facts to the data provider.
Concentration measures and their limits
Let be the share of insured value at location . The Herfindahl-type concentration index summarizes dispersion over the chosen unit partition. If the number and definition of location units change, changes even when the physical insured portfolio is unchanged. If all exposure units are placed inside the same large coastal hazard footprint, a low location-level index may coexist with extreme peril concentration. A geographic grid or administrative region therefore defines a reporting lens, not a physically privileged catastrophe unit.
For a peril , region , and intensity threshold , an exposure-at-risk measure may be written when a hazard model supports those exceedance probabilities. This is a hazard-weighted exposure total, not expected insurer claim, because it omits vulnerability, contracts, and joint event occurrence. A footprint scenario produces a different quantity: . Its result is conditioned on a specified event and need not have a calibrated annual frequency.
Spatial scale is a major source of aggregation bias. Postal zones, census tracts, floodplain categories, building footprints, and raster pixels divide the same landscape differently. Area averages may hide localized wind speed, levee overtopping, soil amplification, or elevation changes. When an insurer moves to a coarser grid, an apparently smooth exposure surface may understate maxima in critical small cells. Comparisons should report exposure-weighted and site-level statistics alongside grids, with results recalculated at more than one support resolution.
Hazard fields and their spatial dependence
Represent intensity at location during event by . The spatial covariance function depends on peril, site, scale, and conditioning information. In earthquake applications, a ground-motion prediction equation describes central intensity and dispersion, while residual spatial dependence governs the joint uncertainty across sites. Jayaram and Baker (2009), in their spatial ground-motion correlation study, estimated correlations from past earthquakes and described their decay with distance and variation with spectral period and site conditions. Their empirical earthquake formulation cannot be transferred mechanically to flood, hail, or wildfire.
The U.S. Geological Survey's ShakeMap software documentation explains how observations, ground-motion models, and intensity reports are combined to generate shaking estimates and uncertainty. Engler, Jaiswal, and Ganesh (2024) examine spatially correlated ShakeMap realizations and show why uncertainty in shaking fields can change aggregate impact measures. Such research addresses the physical field before an insurance policy is applied. Loss analysis must subsequently condition on local building and coverage characteristics.
Windfields can be strongly coherent over substantial distances yet change sharply near a storm track, coastal boundary, or topographic feature. NOAA's National Hurricane Center archive provides HURDAT2 best tracks and tropical cyclone reports for the Atlantic and eastern/central Pacific; a six-hourly track is an input to event reconstruction, not a complete high-resolution wind loss footprint. The archive itself is revised as historical information is reanalyzed. A model reconstructing insurer risk from hurricane catalogs must identify the archive revision, storm intensity and size assumptions, local wind modifiers, event season, and missing observation periods.
Flood depth is controlled by catchment, rainfall, river connectivity, drainage, topography, and engineered defenses. Adjacent parcels may experience different depths, while distant downstream locations may share a common rainfall episode. Wildfire loss combines fire behavior, fuel, wind, ember transport, structure ignition and suppression, with potential dependence through shared neighborhoods and emergency-resource constraints. Hail and severe convective storms require event-scale geometry: one synoptic weather system can contain multiple localized damaging cells. In every peril, the correlation of hazard measurements depends on the conditioning event and cannot simply be inferred from latitude and longitude.
The Copernicus Climate Change Service's ERA5 global reanalysis provides hourly atmospheric estimates on spatial grids, offering one scientific input for reconstructing broad meteorological conditions. Reanalysis is a model-and-observation synthesis with stated resolution and uncertainty, and it does not resolve every building-level gust, embankment, or convective hail core. Data assimilation and model-based interpolation must be acknowledged when fitting spatial dependence from its fields.
Cross-peril overlays and shared infrastructures
Spatial accumulation can bridge perils. A tropical cyclone may generate wind, surge, rainfall, river flood, and infrastructure interruption; an earthquake may lead to fire following earthquake or ground failure; prolonged drought can change wildfire ignition and fuel conditions. Treating these components as separate independent hazard events can double-count physical damage or underestimate aggregate loss. The IPCC's compound-event assessment provides the physical basis for studying combined hazards, including those that operate in sequence.
Insured property losses may also share a dependency network outside the hazard footprint. A damaged substation can interrupt several factories beyond the site of physical damage; a port closure may affect trade flows; a single major complex can host assets insured under many schedules and contracts. Analysis needs a clear loss perimeter: contingent business interruption is contract-specific and can be excluded, limited, or triggered by a different event definition. Economic connectivity should be represented explicitly only when supported by coverage and evidence; otherwise mapped physical-damage sums and modeled downstream consequences should be reported separately.
Illustrative geocode uncertainty calculation
Consider a hypothetical warehouse with a USD 10 million replacement value. Two plausible geocode candidates remain after address reconciliation. The conditional expected annual insurer loss at a verified low-flood site is USD 0.10 million; at a floodplain site it is USD 0.80 million. An audit determines a probability weight of 0.80 for the first candidate and 0.20 for the second. Marginalizing over location uncertainty produces $0.8(0.10)+0.2(0.80)=0.24$ million expected annual loss. Assigning the postal centroid to the low-flood site alone would give USD 0.10 million, 58.3% below the mixture estimate. These conditional loss amounts and weights are expressly constructed for explanation; no actual warehouse, insurer rate, flood map, or empirical geocoder error distribution is represented.
The arithmetic demonstrates a more general effect: positional uncertainty changes loss estimates if the conditional loss surface varies across plausible sites. A portfolio containing many similarly coded postal centroids may have dependent geocoding errors, since all addresses inherited a shared spatial approximation. In that case independently perturbing individual coordinates understates the uncertainty in aggregate concentration. A robust workflow stores candidate locations, positional confidence, polygon support when relevant, and alternative loss calculations before selecting one apparently precise coordinate.
Peril-specific accumulations
Earthquake loss can concentrate in dense cities; its dependence has a measurable source in ground-motion field correlation and shared vulnerability. Earthquake fire, soil liquefaction and infrastructure disruption can produce additional sources of loss covariance. Windstorm loss often spans broad coastal and inland corridors, and regional diversification within the same storm track can be weak. Flood loss is strongly shaped by topography, catchments, engineering protection, and high spatial heterogeneity at small scales. Wildfire loss can be concentrated within an urbanized interface and may escalate through neighborhood-to-neighborhood ignition and suppression constraints. Severe convective storm accumulation is sensitive to the spatial-temporal event definition and the recurrence of multiple mesoscale storms within one underwriting year.
A credible insurer exposure report therefore records a hazard-consistent scenario footprint, chosen intensity metric, geographic support, insured value basis, hazard threshold, and the level of financial loss being displayed. It does not equate a single high-value-zone ranking with the insurer's final retained peak loss.
Event Dependence and Diversification
Conditional dependence generated by a shared catastrophe
Loss dependence is generated by the physical event, the exposure characteristics, the insurance terms, and the financial aggregation of claims. These elements must remain conceptually separate because a calibration error at one layer can be mistakenly offset by a parameter change at another. Consider binary indicators identifying whether individual properties suffer damage above a specified threshold. Even if conditional on a fully specified hazard field each building's residual vulnerability uncertainty were independent, the unconditional indicators can be positively dependent through the probability distribution of the common field.
For random losses and and common explanatory variables , the law of total covariance is
The first term measures remaining association after conditioning; the second measures the dependence induced by the shared condition. An analyst who models only the first term while treating the physical event as fixed may report a low conditional correlation even when portfolio losses are strongly dependent across annual realizations. Equally, an analyst who adds an empirical correlation to hazard fields and then imposes another fully overlapping common-shock process at the damage stage can exaggerate the same mechanism twice.
Different causal mechanisms imply different stochastic structures. An earthquake generates spatially correlated ground-motion intensities that affect building response; an atmospheric river may connect catchments through rainfall timing, antecedent wetness, and flood propagation; a wildfire can spread through neighborhoods with endogenous effects from firefighting constraints. Their commonality lies in the need for a joint hazard-and-damage model, not in a universal spatial covariance shape. The ShakeMap research by Verros and colleagues (2017) directly addresses the computation of correlated shaking fields, providing a clear example of why a spatial component can be essential for credible earthquake loss estimation.
Independence, correlation, and joint-tail probability
A covariance coefficient summarizes one aspect of association under finite-variance assumptions. It does not determine a joint loss law and is sensitive to the distribution, valuation unit, and observational regime. Independence implies factorization of the joint distribution; zero covariance does not generally imply independence. Pairwise independence does not suffice for joint independence among three or more locations. Copulas provide a mathematical representation of the dependence structure connecting marginal distributions, but their parameterization and upper-tail behavior must be separately identified. The fundamental statistical caution in Embrechts, McNeil, and Straumann's risk-dependence analysis is especially important when rare catastrophes provide little direct information on extreme co-movements.
Linear correlation can be high without guaranteeing a particular probability of simultaneous threshold exceedance. Conversely, tail dependence can be material even when an annual sample's ordinary correlation seems modest. To estimate the extreme loss of a catastrophe portfolio, the model requires a credible joint representation at the severity levels that drive capital and treaty exhaustion. Historical correlations of insurer claims may appear to be zero because few severe common events occurred in the observed period or because exposure was acquired, divested, and reinsured over that period.
The elementary variance identity remains useful as a diagnostic. Suppose identical exposures have loss variance and all pairwise correlations equal . Then . At , the standard deviation scales as ; at , it scales as . This contrast is conditional on the common pairwise-correlation specification and does not establish an empirical spatial dependence law. It clarifies that increasing policy count provides different variance diversification under different forms of association. For large , even relatively small positive average dependence can dominate the variance of aggregate loss.
Fully specified two-site tail experiment
Consider two insured sites, each paying exactly USD 4 million if it is struck during a one-year observation period. Each site's marginal probability of such loss is 0.02. The declared marginal loss distribution is therefore and , in units of USD millions. Define .
Under a counterfactual independence assumption, the annual aggregate distribution is , , and . Expected annual aggregate loss is million. The 99th percentile of , defined as , is USD 4 million: the cumulative probability at zero is 0.9604 and at four is 0.9996. The event of both losses in the same modeled year has probability 0.04%.
Under perfect common-shock dependence, in every realization, with and . The marginal distributions and expected annual aggregate loss remain unchanged at USD 0.16 million, but the 99th percentile rises to USD 8 million. Joint loss probability is 2%, fifty times its independence counterpart. For the upper 1% of observations, expected shortfall at 99%, under the standard quantile-integral convention, is USD 8 million in the perfectly dependent model. Under independence, the upper 1% consists of the 0.04% probability of USD 8 million and another 0.96% probability of USD 4 million, giving million.
| Synthetic annual aggregate metric | Independent sites | Perfect common shock |
|---|---|---|
| Each site's loss probability | 2% | 2% |
| Expected aggregate loss, USD million | 0.16 | 0.16 |
| Probability of USD 8 million aggregate loss | 0.04% | 2.00% |
| 99% aggregate loss VaR, USD million | 4 | 8 |
| 99% expected shortfall, USD million | 4.16 | 8 |
This experiment is a constructed probability calculation, independently reproducible from its declared assumptions. It estimates no geographic correlation coefficient, real property loss probability, or insurer capital requirement. It isolates the effect of a dependence specification while holding marginal distributions and expected loss constant. The simplified point-mass distributions also demonstrate why an apparently stable mean loss can coexist with a large change in tail capital. Real models additionally require uncertainty over event severity, varying asset values, multiple occurrences, policy conditions, and model parameters.
Spatial diversification is peril-conditioned
A portfolio's diversification cannot be summarized by separation distance alone. Two inland cities may have low dependence for a local flood but significant co-exposure under one regional drought-related wildfire season or synoptic storm sequence. Two structures inside the same urban block can exhibit contrasting physical vulnerability because of building age, seismic reinforcement, elevation, or noncombustible materials; yet their loss adjustment and repair costs can be jointly affected by local labor shortages. A metropolitan earthquake portfolio may achieve some vulnerability diversification across structural classes while retaining high dependence from the ground-motion field.
The useful scientific question is thus: which common physical events and common contractual or economic conditions can generate large joint loss at the selected portfolio boundary? Alternative peril-specific maps can display exposure concentration and scenario losses separately. Natural-disaster stress portfolios should also record whether events are a single occurrence, a temporally compounding sequence, or separate occurrences whose annual losses interact through treaty limits.
Correlated residuals versus correlated events
A spatial event simulation normally includes at least two forms of dependence. Across one earthquake, residual shaking may be correlated across sites conditional on the rupture. Across years, stochastic catalogs determine which earthquake ruptures occur. Confusing spatial residual covariance with frequency correlation leads to an internally inconsistent model. The OpenQuake scenario risk calculation illustrates the distinction by simulating ground-motion fields for a specified rupture and explicitly considering spatial residual correlation. Its event-based risk calculation adds stochastic-event sets and portfolio loss distributions.
For atmospheric catastrophes, a similar distinction exists between uncertainty in the track or spatial field of an identified storm and uncertainty in annual storm number, track distribution, and climatic conditions. A hazard generator must preserve physically meaningful spatiotemporal organization rather than draw each property's storm intensity independently. Evidence of geospatial dependence can enter from instrumental recordings, physics-based model ensembles, historical event reconstructions, and observed damage patterns, each with distinct sampling biases.
Event attribution and the definition of an occurrence
An occurrence can be defined physically, contractually, or operationally. A windstorm spanning two days may have one storm identifier but damage several times to different properties; a sequence of flood waves can share rainfall sources while falling within different insurer occurrence clauses. Reinsurance treaty wording may include hours clauses and separate occurrence definitions by peril. An insurer can book losses over different accounting periods as claims develop; a reinsurer may recognize an aggregation differently under the treaty.
This distinction is material to spatial correlation research because the same physically dependent losses may aggregate financially under one occurrence or multiple occurrences. Physical modeling should preserve event IDs and times before the contract layer groups claims. Without this separation an observed ceded-loss statistic can be incorrectly used to infer a physical dependence structure. Lloyd's 2026 disaster scenario program emphasizes multiple-event stress as well as single-event footprints, reinforcing the importance of occurrence sequencing.
Climate drivers and dependence change
A stationarity assumption holds the joint occurrence and intensity distribution fixed over the forecast period. Exposure growth, land-use change, changing fire management, coastal defenses, urban infrastructure, and evolving climate conditions can all alter catastrophe accumulation. The IPCC (2021) evaluates observed and projected changes in weather and climate extremes, including combined hazards and concurrent events. These scientific statements do not imply a uniform trend in every regional peril or a known annual percentage change in any insurer's capital requirement. The translation to insured loss depends on the specific geography, building stock, coverage, and adaptation measures.
A defensible nonstationarity analysis compares separate mechanisms: hazard occurrence and severity, conditional spatial dependence, vulnerability, insured value growth, take-up, claim cost inflation, and policy terms. Sensitivity tests should vary them separately where possible before combining them in scenario ensembles. A single aggregated time trend in insurance losses cannot establish which mechanisms changed or whether correlations between losses have strengthened.
Loss Distributions and Capital
From event-level loss to annual loss
An event loss table records a set of modeled catastrophe occurrences , associated rates or probabilities, event-specific losses, and the insured portfolio and financial basis used. A stochastic year is a simulated sequence of occurrences with whatever dependence assumptions govern frequency, seasonality, and event interaction. Annual aggregate loss is the result of applying all relevant policy terms across that sequence. Mean annual loss, OEP and AEP must be calculated consistently from the simulated years rather than mixed across incompatible representations.
For a homogeneous Poisson count with rate and independent event severities , an elementary compound model is , with and, when the second moment exists, . This is a simplifying benchmark. Catastrophe frequencies may be regime-dependent, spatially connected, temporally clustered, or affected by changing hazard conditions. In such cases a mixed or conditional process may be more appropriate, and the compound-Poisson variance identity no longer describes every source of uncertainty.
A return-period calculation must specify whether the reported loss is ground-up, gross insured, or net retained, and whether it represents a single event maximum or total annual losses. The 99.5% quantile of annual net insurance loss can shift when occurrence frequency, treaty reinstatements, aggregate limits, and event ordering are varied. It is methodologically incorrect to sum a collection of separate single-peril quantiles and label the result the quantile of the joint portfolio without evaluating the combined distribution.
Probable maximum loss and the quantile question
Probable maximum loss (PML) is used in different commercial contexts for a selected exceedance probability, event stress loss, or specified scenario severity. A scientifically complete PML statement identifies the exact probability and time horizon, the modeled loss variable, the current portfolio, coverage basis, event set, peril aggregation and reinsurance assumptions. The use of the same phrase across vendors or contracts does not establish that two figures are comparable.
For a nonnegative annual retained loss , define . Expected shortfall may be defined by when the integral exists. Both are conditional on a probability law. Expected shortfall summarizes the mean of the worst modeled probability mass beyond a threshold, with suitable treatment of atoms. Neither statistic represents maximum physically possible damage, and a simulation horizon with insufficient rare-event sampling can yield unstable estimates.
Tail metrics should be accompanied by event-contribution analysis. A portfolio may have a large 1-in-200 annual loss because of one severe hurricane, several moderate catastrophes exhausting annual terms, or a compound physical event. Equal quantiles can conceal different underlying mechanisms and different loss-control options. The distribution of annual event count and the correlation structure of event severities are therefore part of the capital explanation.
Insurance terms as nonlinear mappings
Let denote ground-up physical loss to an insured asset, a deductible, and the available policy limit above the deductible. In a deliberately simplified single-coverage model, covered insurer loss can be . Actual wording may instead apply percentages to total insured value, separate occurrence deductibles, franchises, named-storm deductibles, sublimits, waiting periods, or policy aggregates. Business interruption and additional expense covers can require separate damage triggers and loss periods.
The important consequence is nonlinearity. In general differs from when limits or aggregate deductibles apply. A physical damage correlation calibrated without the financial layer cannot be reused as the correlation of indemnity payouts without additional examination. Deductibles can make smaller correlated physical damages disappear from claims, whereas near-limit losses can become capped at the same contractual maximum. The result depends on the joint distribution relative to attachment and exhaustion thresholds.
The Oasis financial module specification operationalizes distinct ground-up, direct-insured, and net perspectives. This provides a reproducibility principle: save event- and location-level outputs before and after each contract transformation, with a version of the policy and treaty wording used to calculate them. A final annual net loss table without that lineage cannot explain whether a change in accumulation arose from hazard science or financial terms.
Reinsurance and occurrence structure: an exact example
Consider a synthetic property insurer with an occurrence excess-of-loss treaty that pays up to USD 3 million per occurrence above a USD 3 million attachment. Ignore treaty premiums, reinstatement premiums, hours clauses, counterparty credit risk, and annual aggregate restrictions for this example. Assume all gross losses are eligible under the treaty.
In Scenario A a single covered catastrophe occurrence produces USD 8 million gross insured loss. The reinsurer pays million, leaving USD 5 million net retained. In Scenario B two separate eligible occurrences produce USD 4 million gross each within the year. Each yields million recovery, for total ceded loss of USD 2 million and annual net retained loss of USD 6 million. Both scenarios have USD 8 million annual gross loss, but annual net differs by USD 1 million because the treaty operates at the occurrence level.
| Synthetic treaty outcome, USD million | Single USD 8m occurrence | Two USD 4m occurrences |
|---|---|---|
| Annual gross loss | 8 | 8 |
| Total occurrence recoveries | 3 | 2 |
| Annual net retained loss | 5 | 6 |
This experiment is exactly reproducible and reveals the need to retain individual event IDs and contract-effective dates. Different treaty attachment, reinstatements, co-participation, annual aggregate limits, or event groupings can reverse or magnify the outcome. The demonstration establishes a mathematical dependence of net loss on occurrence structure, not a market estimate of reinsurance pricing or recoverability.
Sequential catastrophes and annual exhaustions
A portfolio that survives one major storm can incur further losses when a second storm strikes before repairs, defenses, and claims operations have recovered. Damage state can be path-dependent: an earlier event can weaken a structure, consume claim reserves, exhaust a policy aggregate, remove a temporary roof, or increase building-code upgrade costs. Repair inflation may be correlated across insureds due to emergency labor and materials scarcity. Multiple-event stress therefore must consider the temporal sequence and changing asset state, not only the addition of independent event-loss amounts.
Lloyd's Realistic Disaster Scenarios include stress testing of multi-event conditions. Their purpose is supervisory and market exposure management; scenario reporting does not itself define annual occurrence probabilities for a specific insurer. For capital modeling, scenario losses should be evaluated together with plausible frequency, dependence, contracts, and available financing, while retaining uncertainty about rare sequences for which historical observation is sparse.
Reinsurance, capital, and basis risk
Reinsurance can transfer part of the catastrophe distribution, but its contract may attach by event, loss type, geographic zone, hours clause, or insured line. The difference between physical damages and the event definition used in the treaty creates aggregation and basis risks. A modeled payout under a parametric instrument can differ from indemnified property loss because its trigger uses wind, rainfall, seismic intensity, or another index. Recoveries may be delayed or disputed, and a large regional catastrophe can stress several counterparties or the wider claims settlement system.
A complete gross-to-net model must track direct policies, ceded participation, layers and reinstatements, aggregate limits, exclusions, counterparty collectibility, and reinsurance receivables. The issuer's retained loss after contractual recoveries is the relevant underwriting result for many internal capital applications, while liquidity analysis must also consider the period between paying gross claims and receiving recoveries. Credit and liquidity effects should not be quietly folded into a physical spatial correlation coefficient.
International capital regimes and risk governance
Solvency is jurisdictional. In the European Union, Article 101 of the Solvency II Directive sets the one-year 99.5% VaR calibration for the Solvency Capital Requirement in terms of basic own funds. This requirement is not equivalent to declaring a 99.5th-percentile catastrophe event loss the entire insurer capital requirement: market, credit, operational, underwriting, diversification, liabilities, and mitigation interact through the legally applicable approach. The delegated-regulation natural catastrophe risk sub-module addresses windstorm, earthquake, flood, hail, and subsidence under the standard formula. An insurer using an approved internal model and an insurer using a standard formula may represent dependency differently under supervisory rules.
The IAIS Insurance Core Principles and ComFrame provide international supervisory benchmarks for enterprise risk management and assessment of aggregated exposures. These principles are not a single global catastrophe capital formula. The supervisory concern extends to management information, risk appetite, ORSA, concentration, and the interaction of risk-transfer decisions with retained solvency. An acceptable internal model should make assumptions, validation results, and sensitivity to geospatial concentration comprehensible to decision-makers.
The World Bank Group describes disaster risk finance and insurance as one component of a layered financing strategy. Sovereign contingency instruments, insurance, reserves, and credit facilities serve different beneficiaries and economic purposes. A government's financing capacity and the losses retained by a private property insurer should be compared as related parts of catastrophe finance rather than merged into one accounting measure. For the insurer, the central capital question remains whether assets, reinsurance, and contingent resources are sufficient under the selected distribution and legal requirements.
Stress tests and risk decisions
Risk appetite can be expressed through limits on modeled event loss, annual tail loss, geographic concentration, or minimum capital adequacy. Such thresholds are governance choices informed by the distribution, rather than physical properties of hazards. Stress tests should examine concentration across adjacent and remote regions, alternative peril maps, earthquake residual dependence, severe convective storm clustering, partial failure of flood defenses, changes in repair inflation, and simultaneous asset-side market disruption.
A manager can reduce catastrophe risk by declining or repricing exposures, adapting insured structures, increasing deductibles or limits discipline, diversifying across hazard-independent regions, purchasing reinsurance, or building capital and liquidity buffers. The choice has consequences for accessibility and financial protection; a risk measure alone cannot determine an optimal social outcome. The EIOPA protection-gap dashboard demonstrates the distinction between modeled hazard-related economic exposure and observed insurance coverage in Europe. Accumulation management changes the allocation and bearing of loss; it does not remove the underlying peril.
Empirical Verification, Uncertainty, and Competing Explanations
From observations to an auditable test
The validation of property catastrophe accumulation is a test of a chain of measurements and probabilistic assumptions. A source record identifies a building and its contract; observations or physics-based models estimate the hazard at that building; engineering and statistical functions convert intensity into damage; claims and policy records establish the eligible insurer obligation; event attribution and reinsurance contracts produce aggregate retained loss. A convincing validation exercise checks each transformation and the assembled result. Agreement at the annual industry loss total is insufficient: two error-ridden models can obtain the same total through offsetting errors in location, wind intensity, damage ratios, claim inflation, coverage, and event definition.
An auditable data study begins by declaring the target variables. For geospatial concentration, the observed target is the set of properties and exposure values existing on a fixed date, with each site's geocode quality and geographic support. For physical hazard, the target might be observed station wind, water marks, radar-estimated rainfall, intensity measures, shake observations, or remotely sensed burn perimeters. For damage, the target is a clearly classified physical loss state or repair estimate. For insurance, the target is the covered claim amount, including zero claims and settlement development. For dependence, the target must be a joint distribution over comparable occurrences rather than a collection of unrelated marginal records.
The International Association of Insurance Supervisors' Insurance Core Principles require insurers to address enterprise risk and control in a manner sensitive to their exposures and obligations. In research terms, the requirement for demonstrable control translates into versioned data, specified hypotheses, known input provenance, recorded exceptions, and reproducible model runs. These documents set supervisory expectations; they provide no calibrated earthquake residual correlation or universal threshold for acceptable geocoding accuracy.
Source acquisition and data lineage
The portfolio and hazard records have different owners and update cycles. An exposure file may contain a schedule-of-values extraction date, address history, building replacement cost, underwriting class, and coverage changes. A policy administrative system may contain effective dates and amendments, while a claims system may contain notification, paid amount, case reserve, subrogation, and closure. Meteorological and geological repositories generally use different location identifiers and event classification conventions. A reproducible analysis needs a source-of-truth record for each field, spatial reference system, timestamp, currency, and version.
The National Hurricane Center's tropical cyclone data archive illustrates an official observed-event repository, including best tracks and forecast products. The USGS ShakeMap documentation is an example of event-specific ground-motion estimation supported by multiple data inputs. ERA5 single-level reanalysis data is useful for regional meteorological reconstruction. These archives have different spatial supports and observational characteristics. No common numerical resolution can be assumed, and the output of one cannot be substituted for another without changing the estimand.
A claims-based study faces outcome-selection problems. Properties with no insurance, no filing, a denied claim, or unresolved adjustment may be absent from the observed paid-loss records. Closed-claim amounts and incurred estimates reflect different development stages. Paid claims are affected by deductibles, policy limits, legal interpretation, inflation, and adjusting behavior. A spatial model fitted solely to insurer payouts can conflate hazard and vulnerability with purchasing decisions and coverage design. The EIOPA natural catastrophe protection-gap dashboard documents variation between economic and insured losses, illustrating why one denominator cannot be silently substituted for the other.
A defensible claims reconciliation identifies the event, insurer entity, policy vintage, asset geometry, coverage, deductible, reserve development date, ultimate-loss assumption, and gross/net perspective. At a minimum, the study should report missing addresses, duplicate properties, shifted coordinates, policy-term gaps, unobserved claims, and out-of-period settlements. The count of high-quality geocodes is itself an empirical result and should appear in the validation report rather than being concealed by a clean-looking map.
Spatial verification and sample design
A conventional random training/test split across individual properties can provide deceptively favorable performance when adjacent properties share storm intensity and construction practices. Nearby training observations may contain information that would not have existed for an unobserved catastrophe in another region. Spatially blocked validation therefore reserves complete geographic areas or, preferably, whole independent events for out-of-sample assessment. For event-specific predictions, event-blocked holdouts are necessary to test generalization to unseen earthquakes or storms. A model that performs well within one known hurricane footprint does not thereby demonstrate accuracy for a different cyclone season.
A useful protocol compares: (a) independent hazard or damage draws; (b) a physically informed spatial correlation specification; (c) alternative ranges and shapes of dependence; (d) uncertainty in geocoding and exposure attributes; and (e) combined hazard-vulnerability-coverage simulations. The comparison should hold the marginal hazard and vulnerability distributions fixed when the goal is to isolate dependence effects. If independent and correlated scenarios use different mean intensity or exposure totals, the observed change in tail loss cannot be attributed solely to correlation.
The original Jayaram–Baker ground-motion study provides an empirical earthquake example, while Verros et al. (2017) document a computational treatment for ShakeMap. Their methods address the dependence of instrumental ground-motion residuals and their implications for earthquake loss. Validation of a property insurer's loss distribution additionally needs insured construction and contractual data. Importing a published correlation function is a modeling choice whose transferability to the target region, magnitude distribution, site classification, and ground-motion measure must be checked.
Scoring and calibration should be tied to the target. Continuous hazard measurements can be assessed by errors, predictive intervals, and spatial residual diagnostics. Probabilistic exceedance forecasts can be evaluated against observed threshold frequencies when the sample permits. Portfolio claim prediction can be compared through event-level aggregate error, spatial damage concentration, and realized insured-loss ranges after claims development. A high correlation between two historical loss totals is not itself evidence of a calibrated upper tail, and a low average error can coexist with severe underprediction for the largest events.
Dependence identification and limited histories
Identifying joint extremes from observations is difficult because large common events are rare and the insured portfolio is changing. One catastrophic earthquake can provide many correlated site observations, but these are not many independent earthquake realizations. Treating thousands of damaged structures within one event as independent draws falsely increases effective sample size. Distinct events, geographic basins, hazard regimes, and construction classes provide the meaningful replication structure. Estimating event-level tail dependence requires accounting for clustering and the uncertainty of each event reconstruction.
Several competing explanations can fit the same annual insurance-loss trend. An increase in reported insured losses may result from a stronger physical hazard, more buildings in risky areas, rising replacement values, rising insured coverage, cost inflation, changes in deductibles and claims adjustment, or combinations. An increase in local pairwise claim correlation may result from larger or different event footprints, denser development, common coverage terms, or common contractor and supply-chain cost shocks. An apparent reduction may result from excluding higher-risk properties or shifting the geographical composition of the insurer's book. Observed correlation is therefore conditional on the population and measurement period.
The IPCC's assessment of weather extremes provides assessments of physical climate drivers, spatially compounding events and regional variation. It does not identify a unique catastrophe-loss effect for a given insurer. An analysis claiming climate-driven changes in insured-loss correlation must independently establish what changed in the hazard distribution, what changed in geography and insured exposure, and what changed in policy coverage. Historical regression alone is especially vulnerable to simultaneous changes in these variables.
Data uncertainty, parameter uncertainty, and structural uncertainty
Three uncertainty classes should be recorded distinctly. Data uncertainty covers absent or uncertain exposure coordinates, building attributes, insured values, event intensity observations, claim records, and historical reports. Parameter uncertainty concerns quantities estimated within a chosen model, such as vulnerability curve coefficients, frequency rates, or a correlation range. Structural uncertainty concerns whether the specified model family correctly represents the physical and financial system: for example, whether a single correlation kernel can represent multi-storm windfields, whether wildfire losses need contagion dynamics, or whether an occurrence grouping can be fixed independently of treaty wording.
A common Monte Carlo implementation can simulate all three, but each uncertainty source remains scientifically different. Bootstrapping individual location records treats them as exchangeable only if that design is justified; event-block resampling respects clusters more directly. Bayesian or ensemble uncertainty intervals depend on priors and model specification. Perturbing all model inputs simultaneously without attribution can create a wide numerical range that is difficult to interpret. The purpose of an uncertainty budget is to identify which inputs move expected loss, AEP and OEP, reinsurance recovery, and solvency measures most strongly.
The geocode mixture example earlier in this monograph illustrates the importance of marginalizing uncertainty before assigning precision to a loss estimate. In larger portfolios, one can compare high- and low-precision location scenarios, using realistic joint perturbations rather than independent random coordinate noise. A similar paired experiment can vary building replacement costs with catastrophe repair-demand inflation, and can vary contractual coverage while holding the physical event unchanged. The resulting attribution of loss changes is more informative than the final range alone.
Independent model comparisons and counterfactual tests
If competing models disagree, the disagreement is a research object. One model may simulate flood defense failure, while another treats a levee as permanently effective. One may use a high-resolution terrain model with explicit building elevation; another may use a coarser administrative proxy. Both may generate technically plausible but different loss distributions. Neither should be selected merely because its headline PML is lower. The test is whether its assumptions are documented, scientifically consistent with relevant observations, and appropriate for the insured portfolio.
A structured counterfactual experiment fixes exposure and financial conditions, then changes one spatial dependence assumption. A second fixes hazard and contracts, then changes geocode support or vulnerability. A third fixes gross event-loss tables, then changes treaty structure. These exercises establish sensitivity and causal interpretation within the model. They do not prove a real-world physical counterfactual absent external evidence. The distinction between model-based inference and observed empirical regularity is particularly important when a limited history forces the tail to rely heavily on simulated rare events.
The OpenQuake event-based risk methodology supports reproducible stochastic loss calculations with explicit inputs and event-generation choices. The Oasis open data standards improve portability of exposure and result descriptions. Together these open implementations illustrate how a study can reveal, rather than conceal, the points at which uncertainty enters. Open formats alone do not establish accurate inputs, well-calibrated hazard fields, or correct policy terms.
Regulatory scope and legal comparability
An international comparison must distinguish scientific practices from mandatory law. European Solvency II sets a one-year 99.5% capital calibration; the exact insurer result depends on the specified regulatory calculation and its legal scope. Lloyd's realistic catastrophe scenarios serve a market supervisory-stress function and are not a substitute for every legal-entity solvency calculation. The IAIS framework is a shared international supervisory reference, implemented through national regulatory systems. A portfolio model comparing jurisdictions must specify accounting basis, event definition, capital target, applicable regulation and date, and gross-versus-net risk perspective before treating results as comparable.
Environmental and legal sources can be revised. A hazard map may change after new topographic observations, a catastrophe model may be updated after extreme-loss backtesting, and a regulator may amend a capital module. Reproducibility therefore requires edition numbers or archived dates. A research conclusion from one vintage should not silently be used as a conclusion about a subsequently revised model or legal rule.
Research QA criteria and negative findings
A complete substantive QA report for a catastrophe accumulation manuscript tests the following propositions. Every numerical statement has a declared unit, denominator, geography and time. Every probability statement defines the event and sample space. Every correlation statement identifies its random variables and conditioning assumptions. Every capital measure identifies its horizon and financial basis. Every legal claim identifies the jurisdiction and authority. Every external citation supports the sentence in which it appears. Every synthetic numerical example is marked hypothetical and checked independently.
The present study does not establish a reliable worldwide numerical spatial correlation coefficient for insured property loss. Available sources instead establish peril-specific physical structures, methods for calculating loss, and examples of institutional oversight. The absence of original insurer location and claims microdata prevents fitting a new joint insured-loss distribution or a novel global historical estimate. That boundary is an essential empirical finding about the evidence available to this study, rather than an unresolved contradiction in its mathematical results.
Integrative Comparative Scientific Synthesis
The relation between geography and financial accumulation
The preceding chapters support an integrated reading of catastrophe accumulation. Geographic information determines which property interests may be jointly exposed to a hazard event; physical and engineering information determines the dependent damage outcomes; policy terms determine which damages create insurer obligations; event sequence and reinsurance determine the retained annual financial burden. Errors and uncertainty propagate through this sequence, often nonlinearly. A highly precise physical hazard model cannot compensate for mislocated properties or incorrect financial terms, and a carefully coded insurance treaty cannot correct a loss distribution that assumes independence across a common catastrophic footprint.
The essential distinction is between measuring potential common exposure and estimating a joint loss distribution. The sum of insured values in a region is a measure of economic concentration on a chosen exposure basis. A scenario loss is a conditional consequence of one specific hazard realization. A PML is a loss statistic tied to a probability model or specified scenario convention. These objects can support each other in governance, but none determines the others without the additional assumptions examined throughout the study.
A second distinction concerns the spatial scale of the event. An insurer may classify risk by administrative area for underwriting and accumulation reports, yet the event's true geometry is controlled by storm track, fault rupture, watershed, wildfire spread, or storm-cell pattern. Concentration by postcode or county is useful for monitoring but requires peril-conditioned validation. The correct aggregation unit may vary across wind, flood, earthquake, and wildfire while the insured contracts remain unchanged.
Comparison of principal catastrophe perils
| Peril | Spatial dependence mechanism | Critical exposure information | Major accumulation uncertainty |
|---|---|---|---|
| Earthquake | Shared rupture and correlated ground-motion residual fields; possible liquefaction and fire-following impacts | Seismic site class, structure, age, retrofit and occupancy | Rupture and intensity field, structural fragility, secondary impacts |
| Tropical cyclone wind | Spatially extensive meteorological circulation and local wind modifications | Coastal geography, building envelope, construction and coverage | Track, storm size and intensity, local gusts, concurrent surge |
| River and surface flood | Catchment-scale rainfall, topography, hydraulic connectivity and defenses | Elevation, floor height, drainage, flood protection, site polygon | Terrain resolution, defense performance, rainfall and hydrology |
| Wildfire | Fire spread, embers, wind, fuel and shared suppression capacity | Wildland-urban interface, roof and surrounding vegetation, access | Ignition/spread dynamics, suppression, seasonal clustering |
| Severe convective storm | Localized hail, tornado and wind cores within evolving storm systems | Roof vulnerability, location accuracy and asset distribution | Sub-grid intensity, storm-cell geometry, repeated events |
The table is a conceptual comparison of well-established mechanisms, not a newly proposed peril taxonomy or a quantitative ranking. Each peril creates distinct validation targets. Earthquake studies can exploit instrumental ground-motion records and event-based shaking simulations. Tropical cyclone studies can use observed tracks and reconstructed windfields. Flood studies require hydrological and terrain evidence. Wildfire and convective-storm studies require spatial-temporal event identification at appropriate resolution. These distinctions preclude a single universally valid distance-correlation parameter.
Capital consequences of the evidence
The two-site numerical experiment showed that identical marginal loss probabilities and identical expected aggregate loss can produce very different upper-tail loss when dependence changes. The reinsurance experiment showed that identical annual gross loss can produce different retained annual loss when occurrence structure changes. Together, these results establish a substantive methodological consequence: dependence assumptions and contract grouping rules are separate drivers of financial tail risk. Both must be preserved in any credible source-to-capital audit trail.
Under the European Solvency II system, the statutory capital measure is linked to the basic-own-funds 99.5% one-year VaR standard. A catastrophe loss model may inform one part of this calculation; its own event PML is not equivalent to the entire capital requirement. Reinsurance can alter retained tail loss, but it may introduce counterparty, liquidity, and basis risk. The scientific evaluation must therefore identify what has been transferred and what remains at the relevant legal entity and time horizon.
Consequences for underwriting and research practice
An underwriter reviewing a new risk should be able to identify its location uncertainty, value and building characteristics, peril footprint membership, marginal loss potential, event contribution to portfolio tail, insurance terms, and reinsurance effects. An accumulation manager should be able to trace a reported concentration to the contributing insured interests and back from a capital scenario to the generating events. A model validator should be able to reproduce the same run from frozen inputs, software and contract versions, then test location, vulnerability, correlation, event frequency, and treaty assumptions separately.
Portfolio diversification decisions are then conditional and transparent. Geographic spread can lower the dependence of some hazards without eliminating multi-region storms or linked infrastructure losses. Engineering resilience can reduce physical damage while leaving exposure to repair-price inflation or utility failure. Reinsurance can reduce retained loss while leaving uncovered physical risk and contractual basis effects. These consequences follow directly from established physical, actuarial, and financial mechanisms and do not require inventing a separate proprietary doctrine.
The strongest practical research program combines independent hazard observations, engineering-based vulnerability, complete insured-location records, policy and treaty contract testing, and a deliberately plural set of severe-event scenarios. Its credibility depends on reporting disagreements and the sensitivities of tail estimates. As the IPCC compound-extremes assessment and the USGS spatial shaking research illustrate in their respective fields, common physical drivers and conditional spatial dependence must be treated as substantive scientific objects.
Open empirical questions within the established science
Several empirical questions remain scientifically significant. How accurately do current property records place facilities near small flood features, terrain breaks, and wildfire interface boundaries? How stable are spatial correlations estimated from historical earthquakes when transferred between tectonic regions and building inventories? How does multi-storm clustering change the aggregate exceedance curve after applying real treaty reinstatements? Which changes in insured loss arise from physical climate conditions and which from the geography of insured development, inflation, and coverage selection? Which catastrophe models can demonstrate consistent event-level predictive skill across entirely held-out hazard events?
Answering these questions requires specific datasets, identified uncertainty models, and reproducible validation. They are research questions grounded in the existing scientific literature, not announcements of new theories. The most informative future studies will separate observed hazard, observed damage, economic loss, covered insurance loss, and net retained loss so that results are comparable and uncertainty remains traceable to evidence.
Conclusion
Property catastrophe accumulation is the financial consequence of joint physical hazard, vulnerable insured assets, contractual coverage, and the insurer's aggregation and risk-transfer structure. Its spatial component begins with the accurate geographical identification of insured interests and their susceptibility to hazard footprints. Its statistical component concerns the dependence of hazard intensity, damage, and resulting claim distributions. Its financial component concerns how event losses become annual insured obligations and retained losses after policy and reinsurance terms. These components are causally and mathematically connected, while each requires independent documentation and validation.
The research establishes that geographically dispersed insured properties can remain strongly exposed to common events, that geographic concentration alone does not determine joint-loss probabilities, and that marginal expected losses cannot uniquely determine extreme portfolio loss. The reproducible two-site example produced the same USD 0.16 million expected annual loss under independent and perfectly common-shock assumptions, yet doubled the 99th-percentile annual loss from USD 4 million to USD 8 million. The separate treaty example produced different net losses from identical annual gross loss because the same amount arrived in one or two occurrences. These results show precisely how model dependence and contractual grouping enter capital analysis without treating constructed numerical values as empirical observations.
The comparison of earthquakes, tropical cyclones, floods, wildfires, and severe convective storms establishes the importance of peril-specific spatial structure, appropriate observational resolution, event timing and conditional vulnerability. Recorded insured loss incorporates additional effects from policy wording, values, inflation, claims development, protection penetration and risk-transfer arrangements. Climate change may alter relevant hazard processes; attribution of changes in an insurer's observed or modeled tail risk requires controlling for the evolving geography and terms of coverage.
The central research conclusion is that credible accumulation management depends on a reproducible chain from source geocode and insured contract through common hazard, dependent physical damage, insured claim, event aggregation, reinsurance recovery and annual financial distribution. The chain should be interrogated at each stage with observed evidence, plausible competing explanations, explicit uncertainty and jurisdiction-correct regulatory treatment. The scientific contribution of this monograph is its integrated source-grounded assessment, its explicit separation of distinct dependence objects, and its transparent worked calculations. It claims no new named theory, doctrine, global empirical coefficient, proprietary model or external peer review.
References
Bogdanova, A. (2026, November 7). Catastrophe Risk. Insurance by Bogdanova. Published by Aisentica Development. https://insurancebybogdanova.com/science/catastrophe-risk/
Bogdanova, A. (2026, November 7). Property Insurance. Insurance by Bogdanova. Published by Aisentica Development. https://insurancebybogdanova.com/science/property-insurance/
Bogdanova, A. (2026, November 7). Solvency and Capital. Insurance by Bogdanova. Published by Aisentica Development. https://insurancebybogdanova.com/science/solvency-and-capital/
Copernicus Climate Change Service. (n.d.). ERA5 hourly data on single levels from 1940 to present. Climate Data Store. Accessed November 16, 2026. https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels
Embrechts, P., McNeil, A. J., & Straumann, D. (2002). Correlation and dependence in risk management: Properties and pitfalls. In M. A. H. Dempster (Ed.), Risk Management: Value at Risk and Beyond. Cambridge University Press. https://doi.org/10.1017/CBO9780511615337.008
Engler, D. T., Jaiswal, K. S., & Ganesh, M. (2024). Evaluating the impact of uncertainty in ground motion forecasts for post-earthquake impact modeling applications. Earthquake Spectra. https://doi.org/10.1177/87552930241283201. USGS publication record: https://www.usgs.gov/publications/evaluating-impact-uncertainty-ground-motion-forecasts-post-earthquake-impact-modeling
European Commission. (2015, as amended). Delegated Regulation (EU) 2015/35, Article 120, Natural catastrophe risk sub-module. EIOPA Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2557_en
European Insurance and Occupational Pensions Authority. (2025). Dashboard on insurance protection gap for natural catastrophes. EIOPA. https://www.eiopa.europa.eu/tools-and-data/dashboard-insurance-protection-gap-natural-catastrophes_en
European Union. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 101, Calculation of the Solvency Capital Requirement. European Insurance and Occupational Pensions Authority Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2188_en
Federal Emergency Management Agency. (n.d.). Hazus: FEMA hazard and loss estimation tools. U.S. Department of Homeland Security. Accessed November 16, 2026. https://www.fema.gov/flood-maps/tools-resources/flood-map-products/hazus
Global Earthquake Model Foundation. (n.d.). Event-based risk. OpenQuake Engine long-term-support user guide. Accessed November 16, 2026. https://docs.openquake.org/oq-engine/manual/LTS/user-guide/workflows/event-based-risk.html
Global Earthquake Model Foundation. (n.d.). Scenario Damage and Risk. OpenQuake Engine development documentation (version 3.27.0 as inspected). Accessed November 16, 2026. https://docs.openquake.org/oq-engine/master/manual/user-guide/workflows/scenario-risk.html
Intergovernmental Panel on Climate Change. (2021). Weather and climate extreme events in a changing climate. In Climate Change 2021: The Physical Science Basis, Working Group I, Chapter 11. Cambridge University Press. https://www.ipcc.ch/report/ar6/wg1/chapter/chapter-11/
International Association of Insurance Supervisors. (n.d.). Insurance Core Principles and ComFrame: Standards and guidance. IAIS. Accessed November 16, 2026. https://www.iais.org/activities-topics/standard-setting/icps-and-comframe/
Jayaram, N., & Baker, J. W. (2009). Correlation model for spatially distributed ground-motion intensities. Earthquake Engineering & Structural Dynamics, 38(15), 1687–1708. https://doi.org/10.1002/eqe.922
Lloyd's. (2026). Realistic Disaster Scenarios. Market resources: Underwriting. Lloyd's of London. https://www.lloyds.com/market-resources/underwriting/realistic-disaster-scenarios
National Oceanic and Atmospheric Administration, National Hurricane Center. (n.d.). NHC data archive: Tropical cyclone best tracks and reports. Accessed November 16, 2026. https://www.nhc.noaa.gov/data/
Oasis Loss Modelling Framework. (2026). Financial Module. Oasis LMF Documentation, version 2.5.9. https://oasislmf.github.io/2.5.9/oasislmf/explanation/financial-module.html
Oasis Loss Modelling Framework. (n.d.). Open data standards: Open Exposure Data and Open Results Data. Accessed November 16, 2026. https://oasislmf.org/open-data-standards
Swiss Re Institute. (2026). Global natural catastrophe losses in 2025. In Natural catastrophes in 2025: The persistent rise of wildfire and storm risk (Sigma 1/2026, chap. 2). https://www.swissre.com/institute/research/sigma-research/sigma-2026-01-natcat-2025-wildfire-storm-risk/global-natcat-losses-2025.html
Swiss Re Institute. (2026). Natural catastrophes in 2025: The persistent rise of wildfire and storm risk. Sigma 1/2026. Swiss Re Institute. https://www.swissre.com/institute/research/sigma-research/sigma-2026-01-natcat-2025-wildfire-storm-risk.html
U.S. Geological Survey. (n.d.). ShakeMap software documentation. Geologic Hazards Science Center. Accessed November 16, 2026. https://ghsc.code-pages.usgs.gov/esi/shakemap/index.html
Verros, S., Wald, D. J., Worden, C., Hearne, M., & Ganesh, M. (2017). Computing spatial correlation of ground motion intensities for ShakeMap. Computers & Geosciences. https://doi.org/10.1016/j.cageo.2016.11.004. USGS publication record: https://www.usgs.gov/publications/computing-spatial-correlation-ground-motion-intensities-shakemap
World Bank Group. (n.d.). Disaster Risk Finance and Insurance. World Bank. Accessed November 16, 2026. https://www.worldbank.org/ext/en/topic/financial-sector/disaster-risk-finance-and-insurance
World Meteorological Organization. (2023). Atlas of mortality and economic losses from weather, climate and water-related hazards (1970–2021). WMO. https://wmo.int/publication-series/atlas-of-mortality-and-economic-losses-from-weather-climate-and-water-related-hazards-1970-2021