RESEARCH DOSSIER
Spatial Statistics and Geographic Risk Prediction for Insurance
Abstract. Geographic location influences insured losses through hazard regimes, physical vulnerability, social and economic environments, infrastructure, market selection, claims administration, and the spatial footprint of shared events. Spatial statistics provides a language for representing these relationships in observable exposures and probabilistic forecasts. The central inferential task is to estimate a specified future insurance-loss distribution at a particular location, for a specified population and coverage period, using incomplete, spatially dependent, and frequently selected observations. Location therefore acts simultaneously as a coordinate, a proxy for covariates, a grouping mechanism, and a carrier of dependence. These roles demand separate treatment. This monograph analyzes geocoded exposure and claim observation, spatial autocorrelation, regional aggregation, Gaussian and non-Gaussian geostatistics, spatial generalized linear and hierarchical models, spatial validation, changes in risk surfaces, geographic fairness, and uncertainty communication. It develops insurance-specific formulas for exposure-adjusted claim counts, shared-field dependence, portfolio aggregation, and predictive scores, and provides independently derived synthetic examples. The literature is examined as a collection of established statistical methods and institutional evidence sources; no new named theory, mathematical family, or doctrine is proposed. The resulting analytical synthesis identifies the conditions under which a geographic risk prediction is statistically meaningful, operationally testable, and suitable for insurance research.
Keywords: Spatial statistics; geographic risk prediction; insurance; exposure; geocoding; spatial autocorrelation; Moran's I; geostatistics; Gaussian processes; conditional autoregression; territorial ratemaking; spatial cross-validation; predictive calibration; catastrophe accumulation; geographic fairness; model risk.
Introduction
Insurance losses are observed at policy, property, person, vehicle, claim, event, administrative-area, and portfolio levels. Any geographic analysis must specify which of these units generates each record. A roof located at a street address, a motor policy registered at a mailing address, an accident reported on a road segment, and a flood footprint reconstructed from remote sensing are different geographical objects. Assigning all four the same meaning because their data contain latitude and longitude creates a measurement problem before a statistical model is fitted.
Spatial statistics studies the distribution, dependence, and prediction of phenomena indexed by location. Cressie's systematic treatment of spatial data (1993) distinguishes point-referenced measurements, areal observations, and point patterns. The Handbook of Spatial Statistics (Gelfand et al., 2010) places these alongside space–time methods. Those distinctions are particularly consequential in insurance because an insurer typically observes claims from a changing portfolio rather than every potentially damaging physical event.
The object of geographic insurance prediction can be written as a conditional forecast
where is a precisely defined future loss outcome, a location or spatial support, coverage and exposure attributes, insured time or insured value as appropriate, and information genuinely available at the forecast date. The event set may describe a claim, a loss interval, a tail exceedance, or aggregate claims across several locations. A forecast of expected claim frequency, a predicted conditional severity, and a geographic dependence structure answer different questions and require different observation models.
The central historical sources establish complementary inferential approaches. Moran's continuous stochastic phenomena analysis (1950) helped formalize global spatial association; Geary's contiguity-ratio investigation (1954) emphasized local contrasts among neighbors. Anselin's local indicators of spatial association (1995) developed interpretable local diagnostics. Besag's lattice-system statistical analysis (1974) gave a conditional-model route to dependent areal effects. Their historical scope is spatial inference generally; their relevance to insurance is established here by the explicit observation and prediction problems, not by attributing actuarial applications to every original author.
The research methodology combines independent mathematical derivation, comparison of original statistical publications, use-specific review of government and supervisory data documentation, and a constructed insurance example that permits numerical checks. Publisher pages and official institutional sources were inspected for source identity, descriptions, methods, and dates; restricted monographs and paywalled historical articles were not represented as fully read. Every illustrative number is hypothetical. No proprietary insurer dataset, confidential policy record, individual rating decision, or prospective field validation underlies the examples.
The scope is international. Insurance arrangements vary by country, product, pricing law, hazard, claims convention, and data availability. Institutional materials from Europe or the United States are identified as geographically limited evidence rather than universal regulation. This research is an encyclopedia monograph, with conceptual and methodological conclusions rather than a personalized insurance recommendation.
Geocoded Insurance Exposure
Spatial support, policy time, and outcome definition
A geocoded exposure is a risk-bearing unit with a location assignment, an insurable interest, applicable contractual coverage, an observation period, and an exposure measure. For a residence, a parcel polygon or building footprint may be the relevant spatial support; the centroid of a postal code is usually a coarser substitute. For a moving vehicle, mileage, driving routes, garages, registration addresses, and accident locations differ. For crop insurance, fields, farming operations, soil units, weather stations, and administrative reporting areas need explicit crosswalks. A spatial model should record the operational definition and precision of each support rather than treating coordinate pairs as self-explanatory facts.
Consider an annual claim count , exposure duration , location , and feature vector . A baseline frequency model is
The offset is meaningful when exposure enters multiplicatively and the unit rate is approximately constant over the counted exposure. The choice of denominator matters: insured years, vehicle-years, vehicle-miles, acres, buildings, and sums insured are not automatically interchangeable. A claims-made liability record may require a reporting and development model before any geographic comparison can be interpreted as ultimate-loss risk.
For a monetary loss , the research design should distinguish total paid amount, incurred amount, ultimate loss, gross loss, net loss after reinsurance, loss per exposure, and severity conditional on a claim. Claim frequency and severity may have different spatial relationships. A territory with higher reporting propensity may exhibit more recorded claims even if underlying damage incidence is comparable; a region with higher rebuilding costs may produce larger insured severities under equal hazard intensity. Geographic estimation must locate the response in the claim-production chain.
An exposure-based identity clarifies a frequent aggregation error. If one district has 15 claims in 100 insured years and a second has 90 claims in 900 insured years, the observed district rates are 0.15 and 0.10; the combined rate is , whereas the unweighted district-rate average is . The pooled exposure-weighted rate and the average district rate describe different estimands. Neither provides a causal explanation for the difference between districts. These hypothetical figures are fully specified to demonstrate that the geography of denominator weights changes a reported risk surface.
What a coordinate records
Coordinate systems carry units and assumptions. Earth-surface great-circle or projected local distances may be appropriate for physical proximity; the shortest route over a road network may better characterize vehicle access, repair time, or emergency response. Latitude-longitude measured in degrees should not be inserted as though degrees were homogeneous kilometers. A projected coordinate reference system introduces distortion that must be assessed over the study region; boundary-crossing and global portfolios may require spherical or non-Euclidean treatments.
Geocoding uncertainty also changes risk classification. If an assigned parcel point lies close to a mapped inundation boundary, a modest positional error can shift a policy from one polygon to another. A sufficiently honest data record retains a coordinate precision or uncertainty region and, when material, computes a distribution over candidate hazard values rather than declaring one exact intersection. A postal-code centroid and an exact rooftop coordinate offer different information content, especially for coastal floodplains and wildland–urban interfaces. Precision in a software field is not equivalent to location accuracy.
An insurance data audit therefore records the address source, geocoding procedure, match confidence, coordinate reference system, boundary vintage, map version, coverage vintage, start and end dates, location-change history, and the party authorized to use each field. If a location is updated after a claim, the historical exposure must retain the pre-event information state. A seemingly innocuous overwrite can leak an event's outcome into a model advertised as prospective.
Observed loss and selection
The insured portfolio is a selected population. Underwriting acceptance, policy price, affordability, distribution access, insurer withdrawal, reinsurance availability, and government participation all influence which exposures become observed. A frequency model estimated only on accepted policies predicts experience conditional on that selection; extrapolation to uninsured properties or newly entered territories requires an explicit transport assumption.
Let denote that an exposure enters the observed insurance data. Then generally differs from if selection depends on unmeasured hazard or vulnerability. A mapping system that displays empirical insured losses as geographic physical hazard may therefore confuse claim selection with hazard incidence. The correct research object can remain the insurer's selected portfolio—provided that selection is named and new-business forecasts are evaluated against the relevant prospective mix.
Reporting practices add further nonrandom measurement. A weather-event database can provide valuable exogenous hazard indicators, but those recorded events are not insurance policies or complete claims. NOAA's Storm Events Database documents significant weather events in the United States and adjacent waters; its observation threshold, documentation process, and spatial metadata differ from insured claims. The Storm Data preparation instructions distinguish reported damage estimates from cases with missing or approximate valuations. Such amounts cannot simply be treated as audited insurer-level payments.
Hazard products also have version and interpretation constraints. FEMA's National Flood Hazard Layer guidance describes effective US flood hazard mapping and related revisions. A flood-zone designation is a specific cartographic and administrative product; it is not, by itself, the full conditional distribution of an insured building's future loss. The hazard map, insured building characteristics, contract, claims process, and calendar period must be connected deliberately.
Environmental covariates and granularity
Environmental features can be sampled at very different resolutions. Copernicus describes ERA5 climate reanalysis on regular atmospheric grids with a stated temporal and spatial structure. Assigning the same grid value to many nearby policies produces shared covariates, so treating policy rows as independent replicates of a precisely measured local hazard can generate overconfident inference. Sub-grid variation, terrain effects, and local construction quality remain unresolved even when the reanalysis is internally consistent.
Elevation products illustrate the same issue. The USGS 3DEP accuracy documentation distinguishes terrain-model products and reports geographically variable vertical accuracy. A model for flood depth sensitive to a few tens of centimeters cannot ignore elevation error solely because a raster contains detailed decimal values. Spatial resolution, source vintage, horizontal registration, vertical datum, interpolation, and physical meaning should be part of the data dictionary.
A practical provenance table has, for each field: source and edition; coordinate system; spatial and temporal support; extraction algorithm; missingness and positional uncertainty; permitted legal use; and whether the feature was available as of the intended forecast date. This is a reproducibility requirement and a direct defense against geographic leakage.
Spatial Autocorrelation
Dependence as a property of outcomes and observation
Spatial autocorrelation describes statistical association between measurements indexed by related locations. Close locations may share unobserved characteristics, the footprint of one event, social and infrastructural systems, or a reporting regime. A positive coefficient indicates similarity under a chosen spatial weight system; it does not identify which mechanism created similarity. Distinguishing observation-process clustering, latent hazard dependence, and genuine event dependence is particularly important when the purpose is risk aggregation.
Define observations , mean , nonnegative adjacency or distance weights with , and . The familiar global statistic is
Moran (1950) provided the classic basis for global measurement through spatially indexed stochastic observations. Under a suitable random-permutation null and zero diagonal, the familiar expectation is ; an observed value near zero is not proof of independence or of a correct spatial prediction model. Weight construction, permutation design, heterogeneity, and the insurance exposure denominator determine interpretation.
For areas that share boundaries, a contiguity matrix may be justified; for continuous distance, inverse-distance or kernel weights may be useful. Row-standardizing weights changes the statistic's comparative meaning, particularly when islands and sparsely connected regions are present. Road-network adjacency and geographic adjacency can also yield different structures. No single weight matrix is a neutral representation of insurance geography.
Geary's statistical mapping contiguity ratio (1954) gives a different emphasis:
For the conventional normalization a null value near one is often used as a reference, with smaller values indicating positive local similarity. The contrast-based nature of Geary's statistic makes it sensitive to differences between linked neighbors. Because both global measures depend on weights and observation scale, research should present them alongside maps and residual diagnostics, not elevate either to a universal risk score.
Local indicators and multiple comparisons
Anselin's (1995) local spatial association indicators identify local contributions and spatial outliers within an explicitly specified neighborhood definition. An insurer might locate a high-claim-frequency region surrounded by other high-frequency regions, or a district whose observed frequency diverges sharply from neighbors. The diagnostic is exploratory. Repeated tests across many locations create a multiple-comparison problem; sensitivity to weight specification and unstable small exposures can produce spurious clusters.
Raw annual rates are particularly volatile where denominators are small. A single claim in a district with five insured years corresponds to 0.20 observed claims per insured year, whereas the same claim in 500 insured years corresponds to 0.002. A mapped rate cannot be interpreted without accompanying exposure and interval information. Hierarchical shrinkage or carefully disclosed smoothing may stabilize estimates, but it can also suppress a genuine small high-risk pocket if the smoothing structure is unsuitable.
Spatial diagnostics should ideally be applied to residuals after explaining measured building, coverage, vehicle, and exposure covariates. Testing autocorrelation in a raw claim map mostly confirms that unadjusted risk varies geographically; a remaining residual pattern reveals dependence or unmodeled structure that the baseline model has not captured. Residual diagnostics require the correct distribution and comparison scale: Pearson or randomized quantile residuals for discrete counts are not interchangeable with raw currency errors.
Stationarity, anisotropy, and common shocks
A stationary covariance representation assumes that dependence is a function of spatial separation rather than absolute geographic position. Isotropy further assumes that separation direction is irrelevant. Coastal flood paths, mountain valleys, river networks, infrastructure corridors, and prevailing winds can violate these idealizations. Such violations can make a local interpolator appear convincing while distorting long-range prediction.
An event footprint introduces a particularly important common shock. One severe windstorm can generate many simultaneous policy losses that are correlated even when ordinary claims are rare. A local latent field estimated from routine claims can describe gradual background geographic variation while failing to generate event-sized dependence or catastrophe tails. Conversely, an event-driven model may inadequately describe chronic small-claim behavior. Correlation parameters should be selected for the peril, timescale, and question being forecast.
Geographic autocorrelation and temporal persistence also interact. Recurrent flood losses at fixed properties, repeat claims following repairs, regulatory changes, shifts in insured values, and changing climate conditions can create apparent geographic stability that reflects data history. The covariance structure must distinguish contemporaneous neighboring effects, within-policy repeated observations, within-event clustering, and shared calendar shocks.
When geographic smoothing becomes a source of error
Smoothing is an inferential choice about how risk changes with location. A long spatial correlation range borrows heavily across areas and may conceal steep physical boundaries. A short range can overfit noisy locations. A smooth spatial surface fitted to coastline coordinates may falsely smooth across water and land; a road-network or barrier-aware model may represent the exposure process more plausibly.
A comprehensive diagnostic compares at least a nonspatial baseline, a spatially smoothed model, and a relevant event or group structure under the same prospective test. The improvement in likelihood or random-split error is insufficient evidence of geographic transportability if the training set contains neighboring policies or later records unavailable at prediction time. This distinction sets up the validation chapter.
Risk Surface and Spatial Regression
What a risk surface means
A risk surface is a function assigning a target risk measure to spatial support at a specified time and under defined portfolio conditions. Common surfaces include expected claim count per insured year, expected insured loss per insured value, conditional loss quantiles, exceedance probabilities, and spatial covariance. Each has a different mathematical unit. Maps should state the unit, coverage, observation period, population, and uncertainty interval in their legends or accompanying metadata.
A geographic mean surface can be represented by a spatial generalized linear model:
Here is a continuous spatial effect, an optional area-level effect, and a link appropriate to the response. For counts with a log link an exposure offset is included; for severity or paid losses, another likelihood and censoring/development treatment may be necessary. Separating fixed attributes from location can help explain residual geography, but statistical collinearity between geography and measured covariates can prevent unique causal attribution.
The surface may be predictive without being causal. If construction quality, poverty, wildfire vegetation, and claims practices are spatially patterned, can absorb their combined effects. Removing or changing geographic coordinates then alters prediction, but an intervention on a coordinate has no coherent physical meaning. Decisions about prevention or market conduct require mechanism-specific explanatory evidence beyond a geographically accurate fitted effect.
Continuous Gaussian spatial fields
Model-based geostatistics formulates a latent process at continuous locations. Diggle and Ribeiro's model-based geostatistical framework (2007) sets out principles for inference and prediction under such models. A simple observation system is
where has covariance , and represents observation-specific noise, often assumed conditionally independent. If parameters are known and the relevant normal assumptions hold, conditional prediction at a new location has mean
where contains covariances with observed latent values, is the observed latent covariance matrix, and is observation-noise covariance. The corresponding latent conditional variance is . This calculation assumes known covariance and regression parameters; parameter uncertainty, geocoding uncertainty, and future covariate uncertainty require additional propagation.
Kriging is therefore a conditional statistical prediction, not simply an aesthetically smooth interpolation. The choice of covariance family, nugget, anisotropy, measurement noise, and observation sampling affects both mean and interval. Predicting a latent hazard measurement at a coordinate is also not equivalent to predicting a policy's insured loss without a coverage and damage process.
Non-Gaussian insurance responses
A Gaussian observation model may be useful for a transformed aggregate rate or continuous covariate, yet a claim count is discrete and often sparse. A log-Gaussian Cox construction, rooted in Møller, Syversveen, and Waagepetersen's log-Gaussian Cox-process research (1998), allows a random spatial log-intensity. Applied to discrete policy counts it can take the form
This is a spatial Poisson–lognormal mixture when the log-rate is Gaussian. Conditional independence of counts given the field does not imply marginal independence. A shared latent field generates marginal covariance.
For clarity, suppose all locations share Gaussian marginal variance , and let . For distinct locations , elementary lognormal identities yield
The single-location marginal variance is
These results follow from the laws of total expectation and covariance under the stated conditional-Poisson assumptions; they are not a universal covariance formula for all spatial insurance models.
An independent numerical illustration makes the magnitude concrete. If two policy-group expected counts are each 0.1, spatial field marginal variance is 0.2, and field covariance between locations is 0.1, then cross-location claim-count covariance is . Each marginal variance is , so correlation is approximately 0.01029. Repeated independent spatially correlated pairs in a large portfolio can affect aggregate variation, although the direction and capital relevance depend on the full cross-covariance matrix and tail distribution.
Area effects, geographic regression, and multiscale surfaces
When policy locations are observed only by administrative area, conditional autoregressive (CAR) effects offer an alternative to continuous Gaussian fields. Besag's conditional neighborhood formulation (1974) establishes the broader statistical basis. Besag, York, and Mollié's Bayesian spatial applications (1991) developed a widely influential structured-plus-unstructured approach in disease mapping that can be adapted carefully to insurance areal data.
A stylized area-rate model is
where captures structured neighborhood variation and independent area variation. Intrinsic CAR components may be improper without identifying constraints; sum-to-zero or other priors and constraints must be disclosed. Results vary with boundaries and neighborhood graph. Areas sharing a border are a mathematical construction, not proof that their peril exposures interact.
Brunsdon, Fotheringham, and Charlton (1998) introduced geographically weighted regression for studying spatially varying coefficients. Their later book-length treatment of local regression (Fotheringham et al., 2002) documents the exploratory and inferential issues. Local coefficients can reveal geographic heterogeneity, but data-adaptive bandwidth, correlated local estimates, and repeated significance testing complicate interpretation. A varying coefficient is not necessarily evidence that the underlying physical or socioeconomic causal effect changes by region.
An insurer can compare a global GLM, a structured area model, a continuous smooth, and a local-coefficient model for a single target under identical geographic holdouts. The selected method may differ for infill prediction within known insured regions, interpolation into a lightly insured adjacent neighborhood, or extrapolation to an entirely new country. No model class dominates all targets.
Physical boundaries and extrapolation
Physical risk mapping often creates non-smooth features: river levees, levee-protected regions, building elevations, seismic faults, municipal drainage networks, coastal defenses, or wildfire fuel breaks. A smooth field can approximate broad residual patterns but should not automatically interpolate across a credible structural discontinuity. Mechanistic hazard covariates, discontinuity terms, barrier-respecting covariance, and targeted local models are available when data and physical knowledge justify them.
The model must disclose whether a map is interpolation among observed insured locations or extrapolation outside the covariate and geographic support of training data. Outside the observed domain, uncertainty often grows because prior assumptions and spatial covariance dominate. A plot with a high-resolution color grid cannot create evidence at locations with no credible reference data.
Hierarchical Geographic Effects
The hierarchy of insurance observations
Insurance observations are naturally nested and crossed: claims within policies, policies within buildings or drivers, properties within neighborhoods, neighborhoods within administrative regions, and many locations within the same catastrophe event. An insurer may operate across years and jurisdictions, each with its own contracts, currency, inflation, claims handling, construction standards, and regulatory framework. A useful hierarchy can accommodate these structures without pretending they all represent a single smooth spatial field.
A hypothetical multi-index count model is
Here represents administrative or institutional grouping, continuous residual geography, calendar-period variation, and a possible common event effect. The last term is meaningful only when event identities are observed or explicitly modeled; in routine claim-frequency modeling no artificial event label should be invented for every claim. This equation describes a family of established mixed-effect and spatial constructions, not a new statistical model.
Separating levels helps prevent overconfident standard errors. Ten thousand insured policies exposed to the same localized hailstorm are not ten thousand independent observations about the storm's spatial footprint. Independent policies can still produce conditionally independent idiosyncratic losses given one event, while marginal aggregate loss is dependent through the shared event severity and spatial field.
Hierarchical pooling and sparse territories
Spatially structured random effects permit partial pooling when individual neighborhoods have limited exposure. A small area with one large claim may borrow information from comparable areas through a covariance or prior model. The posterior or penalized estimate reflects both local evidence and the strength of pooling. A richer data region can rely more on its own observations; a sparse region may be driven largely by broader assumptions.
This pooling is valuable for stability and dangerous if poorly targeted. A statistically weak area adjacent to strongly exposed territory may be smoothed toward a neighbor with different levee protection, construction type, or insurance contract. Borrowing should follow a justified risk process, and shrinkage uncertainty should be displayed. Smoothing cannot repair a missing peril-specific hazard variable merely by applying more elaborate covariance.
Prior sensitivity is particularly important for unobserved or nearly unobserved territories. A hierarchical posterior can look precise under a strong smoothing prior while being weakly supported by local insured-loss data. The research report should compare plausible priors, graph specifications, and target areas and disclose where forecasts are prior dominated.
Geographic rating and actuarial use
Actual actuarial research demonstrates how location can enter insurance rating. Guven's multivariate spatial territory study (2004), published through the Casualty Actuarial Society, treats geographic relativities using latitude-longitude functions in a multivariate GLM setting. Jennings's geographic territory clustering study (2008) describes an application of cluster analysis to group geographically similar risk units. The papers concern specific pricing methodologies, rather than a general proof that geographic grouping improves every market or meets every legal standard.
A territorial rating factor may be an estimate of expected relative loss, a credibility-adjusted parameter, a product-management boundary, or an implemented filed rate. These are distinct stages. Statistical analysis establishes a conditional loss estimate; the tariff then reflects expenses, capital, coverage terms, legal constraints, and insurer governance. Geographic risk prediction must report which stage the surface belongs to.
An original comparative assessment follows. Continuous latitude-longitude polynomials offer parsimonious smoothness but can behave badly near irregular boundaries and outside training support. Cluster-based territories offer administrative convenience and stable group assignment but create discontinuities and depend strongly on selected features and map units. CAR and Gaussian-process methods represent structured stochastic variation and permit uncertainty intervals but require prior or covariance assumptions. Spatially blocked prospective validation can compare their actual predictive utility for a common insurance estimand.
The CAS studies are valuable historically and methodologically, yet their reported numerical performance, locations, and historical rating settings should not be projected onto an unrelated international portfolio. Evaluation must use the appropriate legal market, insurance line, and future population.
Aggregate loss, accumulation, and reinsurance
Spatial dependence matters for aggregation even if location-level expected losses are estimated accurately. Let portfolio loss be . Its variance satisfies
The second term is a structural accounting of dependence, not a minor adjustment that disappears by collecting more policies. Common perils can produce positive covariance across substantial geographic regions. Diversification depends on the geography of event footprints, the mix of policies and perils, and how losses co-move, not merely the number of coordinates in a database.
For a constructed example, assume two independent equal-variance area losses each have variance 100 monetary units squared. Portfolio variance is 200. If their covariance is 40, portfolio variance is 280. The standard deviations become approximately 14.142 and 16.733, respectively. The 40% increase in variance does not imply a fixed 40% increase in a catastrophe tail quantile, because tail probabilities require a joint distribution beyond the first two moments.
A geographic Gaussian field may reproduce ordinary spatial dependence while understating a rare, high-severity storm system. A catastrophe model may provide explicit hazard footprint simulations and vulnerability relations; a statistical frequency model then serves as a complementary component or benchmark. This dossier addresses spatial inference and prediction. Physical catastrophe models, treaty structures, and contract-specific accumulation are broader specialist topics connected through the existing Science architecture.
Change over time and international transfer
Regional risk surfaces change with climate, land use, infrastructure, exposure values, inflation, property maintenance, claims practices, migration, and insurer underwriting. A geographic factor estimated on the last decade is not automatically a stationary physical constant. Trend decomposition should distinguish physical-hazard change from changes in the insured portfolio and monetary loss definition.
The IPCC's risk-management and adaptation assessment (2022) considers uncertainty, decision making, adaptation, and institutional settings for climate risks internationally. That body of evidence supports scenario-sensitive reasoning and the need to communicate deep uncertainty. It does not supply a universal portfolio-specific probability distribution or validated insurance premium factor.
The EIOPA natural-catastrophe protection-gap dashboard combines economic loss, insured loss, coverage, and modeled risk information for European countries. Its figures are valuable for comparing protection mechanisms and identifying information gaps within the dashboard's definitions. They should not be mistaken for individual-property expected-claim predictions or blindly extrapolated to non-European markets. Spatial scale and economic concept determine the proper inference.
International generalization therefore requires explicit transport checks: covariate overlap, claims definitions, hazard instruments, building and population characteristics, insurance selection, and institutions. An apparently successful model in one region can fail when applied where the underlying peril, reporting, or coverage system differs.
Spatial Cross-Validation
Validation must follow the prospective decision
Predictive evaluation asks how a method performs on the information and geographic units it would encounter after deployment. A random split of claims within neighboring insured homes may be appropriate for some narrowly defined missing-record tasks but can be misleading for new-territory prediction because nearby training observations reveal much of the same latent risk environment.
Roberts and colleagues' structured-data cross-validation analysis (2017) demonstrates that temporal, spatial, hierarchical, and related dependence must inform resampling. Valavi and colleagues' spatially separated folds methodology (2019) provides a practical spatial-block approach, originally developed in species-distribution modeling and transferable as a validation design principle. Ploton and colleagues' large-scale spatial validation study (2020) reports an ecological mapping case where spatial validation materially revised conclusions about predictive performance. None of these papers establishes a numeric error rate for insurance; together they support the methodological need to match testing to spatial deployment.
The validation question should be written before splitting data. At least four targets are distinct: prediction for another policy at a familiar location; prediction for a nearby unsampled location; prediction for an entire unseen district; and prediction in a different future climate or coverage regime. A suitable test design for one does not automatically validate another.
Spatial blocking, buffering, and temporal holdouts
A spatial block split holds out contiguous areas, not scattered random records. A buffer around test locations can reduce leakage when close spatial neighbors share unobserved effects. Buffer size is a modeling choice influenced by dependence range and decision scale; selecting it because it produces an attractive score is evaluation leakage. Spatial blocks may need to span a physical basin or storm corridor rather than a square grid if the dependence mechanism follows hydrology or transportation networks.
For a future-year prediction, calendar ordering adds another dimension. Train on information available at cutoff , tune within earlier periods with geography-aware splits, and evaluate on future observations at . When claims mature slowly, prevent future claims development or retroactively revised exposure definitions from entering the training features. If an evaluation set includes an event footprint, every policy affected by that event may require a common fold so that the test measures transfer to genuinely unseen events.
Nested evaluation is useful when selecting covariance ranges, smoothing penalties, spatial block widths, model families, and hyperparameters. Outer folds assess prospective performance; tuning occurs only in inner training folds. Preprocessing, imputation, feature selection, geocoding corrections learned from outcomes, and risk-surface construction must be fitted exclusively within training partitions. Otherwise an apparently spatially separated test can remain contaminated by shared learned information.
Scores, calibration, and decision-specific losses
A model's predictive performance should use the complete target distribution where possible. Gneiting and Raftery's strictly proper scoring rules analysis (2007) establishes why suitable scores reward truthful probability forecasts. For a binary claim event with realized outcome and predicted probability , the Brier score is ; lower average scores are better. For count predictions, held-out log score or a proper count-distribution score may be used. For continuous loss distributions, the continuous ranked probability score or quantile scores can assess distributions and decision-relevant tails.
Spatial calibration must be examined by area, peril, exposure band, and forecast horizon. An overall calibration plot can hide systematic underprediction in a floodplain if overprediction elsewhere compensates. Very small subgroups require uncertainty intervals around calibration estimates; an apparent hotspot in raw residual maps may reflect sampling noise. Spatial uncertainty maps and model scores should report denominator support.
A rate model can have strong discrimination across territories while being badly calibrated in absolute claims per exposure. Conversely, a well-calibrated global mean may fail at identifying regional tail concentrations. Ranking statistics, likelihood, scoring rules, economic loss functions, and catastrophe accumulation checks assess different performance dimensions. Selection should follow the specified insurance research purpose.
Constructed validation counterexample
Suppose a region has two geographically coherent blocks with persistent but different claim rates. In a hypothetical training portfolio, block A contributes 100 insured years and 15 claims, while block B contributes 900 insured years and 90 claims. A model with block labels can learn observed rates 0.15 and 0.10. A random policy-level split of the same blocks can reward this memorization because both blocks appear in training and testing.
Now define the deployment task as forecasting a previously uninsured third block. Its conditional rate is not identified merely by knowing the rates of A and B. A block-specific categorical model has no empirically estimated coefficient for the new area. A spatial smoother can forecast only by imposing distance and smoothness assumptions, while a hazard-covariate model depends on transporting exposure and physical relationships. Random-split accuracy on A and B cannot establish success on the new block.
This example makes an epistemic point rather than claiming a measured improvement from a particular algorithm. The forecast's target population and geographic novelty are part of the estimand. The test must withhold data in a manner that challenges the same extrapolation the model will face operationally.
Evaluation under model and location uncertainty
Predictive intervals can fail through covariance misspecification, measurement error, unmodeled event dependence, or parameter uncertainty. A Gaussian-process posterior conditional on known parameters is narrower than a forecast that integrates credible parameter uncertainty and future covariate changes. Confidence or credible intervals for latent mean surfaces are also narrower than predictive intervals for actual claims, which include outcome noise.
Validation should report errors at policy, area, event, and aggregate levels and explain why some metrics receive primary status. A model can pass policy-level mean prediction while missing joint tail behavior. For catastrophe-related applications, scenario testing and held-out event aggregates complement policy-level validation.
A robust audit includes a nonspatial baseline; exposure-adjusted specification checks; several plausible spatial neighborhood or covariance definitions; geographically separated and chronological out-of-sample assessments; uncertainty-coverage diagnostics; and stratified evaluation of sparse and potentially disadvantaged areas. Repeated model revision on the same geographic holdout eventually compromises its independence; a final untouched evaluation sample is advisable where feasible.
Fairness and Risk Mapping
Geographic proxies and unequal observation
Location is statistically useful because physical hazards and other mechanisms vary across space. It is also socially significant because neighborhoods can correlate with income, housing quality, infrastructure investment, race, exclusion from insurance, and uneven claims access. A spatial predictor may capture legitimate hazard information and reproduce inequities that are irrelevant or impermissible for a particular insurance decision. These possibilities require separate empirical and legal review.
The NAIC's 2023 insurer AI model bulletin describes supervisory expectations for AI systems in US insurance, including governance and consumer fairness considerations. It is an NAIC model bulletin, with implementation and enforceability dependent on applicable state supervisory action; it is not a worldwide statute. Geographic model governance must therefore identify the market, product, effective rules, permitted rating variables, and relevant regulatory requirements before implementing a predictor in pricing or underwriting.
Fairness evaluation should examine both prediction errors and decision consequences. For example, underprediction of loss in a vulnerable district can threaten claims-paying adequacy or create hidden cross-subsidies, while overprediction can impede affordability and availability. The ethical assessment cannot be deduced from a colored map or a single groupwise error statistic. It must state the substantive objective and lawful basis of the decision.
Hazard, vulnerability, and insurance access
A geographic risk surface combines observed loss with exposure and vulnerability conditional on coverage. An area with fewer insured claims may have lower hazard, lower insurance take-up, weaker reporting, higher deductibles, or exclusions. Without data on insurance access and contract differences, its lower recorded payments are ambiguous. Likewise an area with high physical hazard but new defenses may need a revised vulnerability function before prospective claims are inferred.
The European insurance protection-gap framework can help researchers distinguish economic loss from insured loss in a defined institutional setting. It demonstrates why an insured-loss map alone cannot measure society's total disaster exposure. For international research, parallel distinctions should be maintained among physical hazard, economic loss, insured coverage, and realized insurer payment.
The coordinate of an insured object should not be used as a substitute for measured hazard mechanisms when defensible physical inputs are available. Explicit flood depth, construction elevation, building material, fire fuel, security infrastructure, and coverage terms can improve interpretability. Yet these inputs are also uncertain and context-dependent, so their use requires versioned source validation rather than the presumption that physically named variables must be unbiased.
Mapping uncertainty instead of false precision
Every geographic risk map should state its intended estimand, unit, coverage and forecast period, support resolution, interpolation assumptions, and uncertainty. A choropleth can conceal internal variation because all properties within an area share the displayed average. A gridded raster can look more precise than the input data support. High-resolution color maps are communication devices; the underlying uncertainty, not the pixel size, determines inferential precision.
Where useful, a map may show expected claim rate and predictive interval in separate layers, alongside count of insured years, missingness, model extrapolation flags, and hazard-data version. Disclosure of low-support areas is as important as highlighting visually extreme predicted rates. For rare losses, display probabilities over a specified horizon rather than converting them into intuitive but misleading return-period shorthand without assumptions.
Borders require explicit handling. Reprojecting cross-border data into a common map may conceal changes in insurance law, claims data definitions, currency, inflation, perils, and environmental measurement practices. An attractive continental map should not imply that each country contributes an identically constructed response. Hierarchical and country-specific components may be necessary, with cross-jurisdiction comparisons reported conditionally.
Original analytical synthesis: conditions for responsible geographic prediction
The research supports a connected set of methodological conclusions. First, a geographic risk estimate is interpretable only relative to a defined insurance outcome, exposure denominator, temporal window, and spatial support. Its meaning is controlled at the data-generating stage, before covariance or machine-learning complexity enters.
Second, the same observed spatial pattern can be generated by physical hazard, vulnerability, insurer selection, contract design, reporting processes, common-event dependence, or data error. A model that predicts one selected portfolio can remain useful for that task, but its surface does not automatically identify physical risk or justify policy intervention. Inference must keep observation, risk, and decision stages connected yet conceptually distinct.
Third, spatial dependence changes both uncertainty and prospective evaluation. Neighborhood similarity supports local information borrowing, but it also means that randomly separated policy rows may not be independent tests. Dependence that improves interpolation can undermine a claimed ability to forecast unobserved territories. Performance claims must report their exact geographic transfer target.
Fourth, area averaging, map grids, coordinate precision, and territorial clustering are substantive modeling choices. Geographic aggregates can have different denominators and boundaries. Their sensitivity should be studied rather than concealed in a single apparently definitive risk map.
Fifth, local expected-loss accuracy does not guarantee aggregate-loss adequacy. Portfolio variance and tail risk depend on cross-location covariance and event structure. Spatial smooths, event footprints, catastrophe models, and actuarial loss distributions contribute different information to portfolio decisions; the scientifically valid combination depends on the peril and the decision horizon.
Sixth, fairness and reliability are inseparable from legal and institutional context. A model can be numerically accurate on an insurer's historical portfolio and still be misleading for populations excluded by selection or for uses restricted by applicable law. Validation must therefore include sample selection, coverage distinctions, subgroup calibration, and the consequences of geographic classification.
These conclusions are an original synthesis of established statistical and insurance scholarship. They do not purport to establish a new named principle or exclusive theoretical system.
Reproducible research design
A defensible reproducibility package would specify a frozen and permissioned portfolio extract, an exposure data dictionary, geocode precision and source identifiers, historical claims-development and reporting assumptions, physical hazard dataset editions, spatial covariance and neighborhood definitions, and clearly time-stamped model inputs. It would include source code and environment specifications, a deterministic synthetic example for testing formulas, fixed geographic test folds, tuning records, and archived predictions for each holdout.
The full experiment should compare nonspatial GLM, flexible smooth, area-level random-effects model, continuous spatial field, and an event-sensitive model only where the event data justify one. For each, report predictive target, fit convergence, sensitivity to sparse areas and projection, held-out proper scores, spatial residual diagnostics, and portfolio aggregation. No performance claim should be made without the actual datasets and analyses.
The cited official geographic sources are potential inputs rather than evidence that their data have already been joined to an insurer's records. NOAA, FEMA, USGS, Copernicus, IPCC and EIOPA describe different measurement systems and scales. An implementation would retrieve the version valid at the forecast cutoff, document source licences and access rules, and independently confirm geographic coverage and units before constructing predictors.
Conclusion
Spatial statistics is indispensable to rigorous geographic insurance risk prediction because insurance observations are embedded in heterogeneous physical environments, selected portfolios, administrative systems, and shared loss events. A credible prediction begins with a precise target: a defined claim or loss outcome, an insured population, an exposure measure, a location or areal support, and a forecast period. The statistical analysis must then represent the observation process, geographic dependence, and uncertainty at the scale of the intended decision.
Global and local autocorrelation measures describe spatial patterns; geostatistical, generalized linear, and hierarchical models estimate conditional relationships and propagate specified uncertainty; catastrophe-event structures address dependence beyond routine local variation. These methods are complementary, and each requires assumptions that determine its validity. Spatial blocking and chronological validation are central for judging prediction in genuinely unseen areas and periods. Predictive scoring, calibration, and aggregate-tail analysis should be matched to the loss quantity at stake.
The original analytical contribution of this monograph is its integrated account of how support, selection, dependence, validation, and institutional use jointly govern a geographic insurance prediction. This account identifies replicable requirements for research without claiming a newly invented model or historical priority. The resulting scientific standard is explicit: every mapped estimate should be traceable from exposure and location data, through a declared statistical target and held-out validation design, to a clearly bounded interpretation in insurance research.
References
Moran, P. A. P. (1950). Notes on Continuous Stochastic Phenomena. Biometrika, 37(1–2), 17–23. https://academic.oup.com/biomet/article-abstract/37/1-2/17/194868
Geary, R. C. (1954). The Contiguity Ratio and Statistical Mapping. The Incorporated Statistician, 5(3), 115–146. https://doi.org/10.2307/2986645
Anselin, L. (1995). Local Indicators of Spatial Association—LISA. Geographical Analysis, 27(2), 93–115. https://onlinelibrary.wiley.com/doi/10.1111/j.1538-4632.1995.tb00338.x
Besag, J. (1974). Spatial Interaction and the Statistical Analysis of Lattice Systems. Journal of the Royal Statistical Society: Series B, 36(2), 192–225. https://academic.oup.com/jrsssb/article/36/2/192/7027394
Besag, J., York, J., & Mollié, A. (1991). Bayesian Image Restoration, with Two Applications in Spatial Statistics. Annals of the Institute of Statistical Mathematics, 43, 1–20. https://doi.org/10.1007/BF00116466
Cressie, N. A. C. (1993). Statistics for Spatial Data. John Wiley & Sons. https://onlinelibrary.wiley.com/doi/book/10.1002/9781119115151
Gelfand, A. E., Diggle, P. J., Guttorp, P., & Fuentes, M. (Eds.). (2010). Handbook of Spatial Statistics. CRC Press. https://www.routledge.com/Handbook-of-Spatial-Statistics/Diggle-Fuentes-Gelfand-Guttorp/p/book/9781420072877
Diggle, P. J., & Ribeiro, P. J., Jr. (2007). Model-Based Geostatistics. Springer. https://link.springer.com/book/10.1007/978-0-387-48536-2
Møller, J., Syversveen, A. R., & Waagepetersen, R. P. (1998). Log Gaussian Cox Processes. Scandinavian Journal of Statistics, 25(3), 451–482. https://onlinelibrary.wiley.com/doi/10.1111/1467-9469.00115
Brunsdon, C., Fotheringham, S., & Charlton, M. (1998). Geographically Weighted Regression. Journal of the Royal Statistical Society: Series D, 47(3), 431–443. https://doi.org/10.1111/1467-9884.00145
Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2002). Geographically Weighted Regression: The Analysis of Spatially Varying Relationships. John Wiley & Sons. https://www.wiley.com/en-us/geographically-weighted-regression-the-analysis-of-spatially-varying-relationships-p-9780471496168
Roberts, D. R., Bahn, V., Ciuti, S., et al. (2017). Cross-Validation Strategies for Data with Temporal, Spatial, Hierarchical, or Phylogenetic Structure. Ecography, 40(8), 913–929. https://onlinelibrary.wiley.com/doi/full/10.1111/ecog.02881
Valavi, R., Elith, J., Lahoz-Monfort, J. J., & Guillera-Arroita, G. (2019). blockCV: An R Package for Generating Spatially or Environmentally Separated Folds for k-Fold Cross-Validation of Species Distribution Models. Methods in Ecology and Evolution, 10(2), 225–232. https://doi.org/10.1111/2041-210X.13107
Ploton, P., Mortier, F., Réjou-Méchain, M., et al. (2020). Spatial Validation Reveals Poor Predictive Performance of Large-Scale Ecological Mapping Models. Nature Communications, 11, 4540. https://www.nature.com/articles/s41467-020-18321-y
Gneiting, T., & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437
Guven, S. (2004). Multivariate Spatial Analysis of the Territory Rating Variable. Casualty Actuarial Society Discussion Paper Program, 245–260. https://www.casact.org/abstract/multivariate-spatial-analysis-territory-rating-variable
Jennings, P. J. (2008). Using Cluster Analysis to Define Geographical Rating Territories. Casualty Actuarial Society Discussion Paper Program, 34–52. https://www.casact.org/abstract/using-cluster-analysis-define-geographical-rating-territories
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
Intergovernmental Panel on Climate Change. (2022). Chapter 17: Decision-Making Options for Managing Risk. In Climate Change 2022: Impacts, Adaptation and Vulnerability, Working Group II contribution to the Sixth Assessment Report. Cambridge University Press. https://www.ipcc.ch/report/ar6/wg2/chapter/chapter-17/
Federal Emergency Management Agency. (2022). Guidance for Flood Risk Analysis and Mapping: National Flood Hazard Layer. FEMA Risk MAP. https://www.fema.gov/sites/default/files/documents/fema_national-flood-hazard-layer_112022.pdf
National Oceanic and Atmospheric Administration, National Centers for Environmental Information. (n.d.). Storm Events Database. NOAA. https://www.ncei.noaa.gov/access/storm-events-database/
National Weather Service. (2016). Storm Data Preparation: NWS Instruction 10-1605. National Oceanic and Atmospheric Administration. https://www.ncei.noaa.gov/stormevents/pd01016005curr.pdf
Copernicus Climate Change Service. (n.d.). Climate Reanalysis: ERA5. Copernicus / European Centre for Medium-Range Weather Forecasts. https://climate.copernicus.eu/climate-reanalysis
U.S. Geological Survey. (n.d.). 3D Elevation Program: Frequently Asked Questions. USGS. https://www.usgs.gov/3d-elevation-program/science/faqs
National Association of Insurance Commissioners. (2023, December 4). NAIC Members Approve Model Bulletin on Use of AI by Insurers. NAIC. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers