Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Distribution Shift and Transfer Learning in Insurance Prediction

Abstract. Distribution shift is a central threat to insurance prediction because the insurance portfolio is simultaneously a statistical population, a contractual selection process, a time-dependent exposure system, and a regulated set of decisions. A claims-frequency model trained in one insurer, territory, distribution channel, underwriting period, or product generation can experience changes in the mix of policyholders, the conditional relationship between predictors and claims, the reporting process, and the financial meaning of its predicted outcome. This monograph investigates the conditions under which predictive models can be transported across these settings. It distinguishes covariate shift, label shift, concept drift, selection effects, measurement changes, and changes of contractual estimands; develops the mathematics of density-ratio estimation, importance weighting, target risk, effective sample size, transfer learning, calibration and stress testing; and analyzes insurer-to-insurer, time-series, regional, and jurisdictional transfers. Two transparent numerical examples exhibit the difference between a reweightable change in exposure composition and a nontransportable change in conditional loss. Comparative assessment integrates scientific evidence with actuarial professional standards and international, European and U.S. supervisory publications, each with its actual jurisdictional status. The resulting analytical synthesis concludes that insurance model transport requires verification of the prediction target, support, observation mechanism, conditional response, and decision costs, followed by target-domain validation. No proprietary insurer data, new named theory, or observed empirical performance gain is claimed.

Keywords: Distribution shift; dataset shift; covariate shift; label shift; concept drift; insurance prediction; actuarial modeling; domain adaptation; transfer learning; importance weighting; target risk; external validity; insurance portfolios; exposure measurement; model calibration; model governance; negative transfer; jurisdictional variation; insurance data quality; uncertainty; reproducibility.

Introduction: Object, Method, and Analytical Position

An insurance prediction is an estimate of a specified future or outstanding quantity for a specified exposure population at a specified decision time. That definition carries two independent requirements: a statistical description of how outcomes arise and a contractual description of which outcomes count. A model of reported motor claims is useful only insofar as the policies, exposure measures, claim definitions and future operating conditions used at inference resemble those required by the decision. When any of these change, the predictive problem changes even if the fitted model parameters and software remain fixed.

The scientific concept of dataset shift refers to changes between the joint distributions generating training and target observations, as described in the edited volume Dataset Shift in Machine Learning (Quiñonero-Candela et al., 2009). In insurance this is more than a conventional statistical nuisance. Underwriting decisions alter who enters the insured population; claim-handling rules affect which incidents become recorded claims; rating changes can influence renewal; geographic relocation changes hazards; inflation and policy limits change nominal severity; digital distribution modifies both risk signals and observed behavior. A risk prediction deployed across these conditions inherits the consequences of each selection and measurement step.

The insurance-specific research problem is therefore expressed as follows: given labeled source observations S, an identified target population T, and a decision-relevant loss function, determine whether a predictor fitted on S remains valid on T, whether adaptation can improve validity, and which assumptions cannot be assessed using the available observations. The core inferential distinction is between predicting correctly conditional on recorded variables and transporting that predictive relation to a new environment.

The general transfer-learning literature, including Pan and Yang's survey (2010), distinguishes adapting across distributions from transferring parameters, representations or tasks. The mathematical theory of learning across domains (Ben-David et al., 2010) establishes why a small observed source error and modest empirical domain divergence do not alone ensure low target error: irreducible differences between labeling relationships remain. This is particularly relevant where insurance exclusions or claims adjudication differ between contracts.

The method used in this monograph is deductive and comparative. First, define training and target risks and identify the statistical assumptions supporting any transport. Second, derive useful formulas, clarify their conditions, and construct reproducible hypothetical numerical examples. Third, compare research on distribution detection, weighting, transfer representations and uncertainty assessment. Fourth, map those inferences onto insurance frequency, severity, pricing, claims operations, reserving and portfolio-risk use. Finally, consider professional standards and supervisory guidance according to their institutional status. The authorial contribution consists of integrating these sources, specifying failure modes, offering transparent mathematical examples, and deriving practical implications from the established evidence. It does not introduce a new theoretical framework under a proprietary name.

A preliminary restriction is necessary: evidence on machine-learning benchmarks in natural images, documents, medical or satellite datasets illustrates general mechanisms. The WILDS benchmark (Koh et al., 2021) documents losses under real-world distribution shift, but its results are not measurements of actuarial prediction improvement. Insurance-specific claims about performance require data and external validation in the insurance population under consideration.

Covariate and Label Shift

The target risk and the relevant probability laws

Let X denote predictors available at a stated decision time, Y the decision-relevant insured outcome, S the source training regime and T the intended future regime. For loss function L and fitted predictor f, define target predictive risk by

RT(f)=E(X,Y)∼PT[L(f(X),Y)].R_T(f)=\mathbb{E}_{(X,Y)\sim P_T}[L(f(X),Y)].

Ordinary empirical risk minimization approximates R_S(f) from source observations. Its relevance for target use depends on the relation between P_S(X,Y) and P_T(X,Y), the dependence structure of exposures and the definition of Y. For a policyholder-level binary claim indicator, a candidate loss is the log loss. For pure premium a squared-error, Poisson, Tweedie or another justified proper scoring measure may be relevant. A model selected by classification accuracy can fail a premium-estimation task: a high frequency of correctly classified zero-claim policies does not imply an accurate portfolio aggregate.

Under covariate shift, P_S(X) differs from P_T(X), while the conditional law P_S(Y|X)=P_T(Y|X) remains stable on relevant common support. This is the setting of importance-weighted cross-validation (Sugiyama, Krauledat, & Müller, 2007). With an integrable loss and a density ratio w(x)=p_T(x)/p_S(x), if target support is contained in source support,

RT(f)=EPS[w(X)L(f(X),Y)].R_T(f)=\mathbb{E}_{P_S}[w(X)L(f(X),Y)].

The identity follows by substituting p_T(x,y)=p_T(x)p_S(y|x) into the target-risk integral. It is an equality about population distributions, not an automatic guarantee that estimated weights, finite data, and a misspecified model yield a reliable fitted target predictor.

In insurance, a simple example is a changed mixture of urban and rural exposures when within-stratum annual claim risk remains the same and records are measured consistently. Reweighting the source by the new geography mix can recover the target aggregate expected frequency under those assumptions. The same calculation fails when the road network, telematics instrumentation, weather pattern, coverage terms, or conditional driver behavior has changed. A variable called “territory” does not necessarily absorb all the causal and administrative changes that accompany relocation.

Label shift and prevalence change

Label shift assumes that P_T(Y) changes but P_T(X|Y)=P_S(X|Y) is invariant. The black-box label-shift analysis (Lipton, Wang, & Smola, 2018) gives identification and correction procedures under an appropriate invertible confusion matrix. For a binary fraud flag or claim-occurrence classifier, prevalence may differ between pools while predictor measurements conditional on the label are stable. In that restricted case class-prior adjustment can be considered.

Insurance outcomes are often more complicated than categorical labels. Claim amount is a mixture of zero, positive continuous, capped, and delayed payments; the threshold for large-loss labeling may be indexed to inflation or attachment point. A label shift defined on “claim above 10,000 monetary units” is not invariant if the currency or deductible changed. Moreover, label prevalence is often not observable at deployment because claims develop over years. When target labels are unavailable, assumptions about stable class-conditional predictors must be evaluated indirectly and later backtested.

Neither label shift nor covariate shift should be used as a convenient name for every environmental change. Both are specific mathematical restrictions on the data-generating process. The basic statistical object, P(X,Y), can also change because P(Y|X) changes; this is commonly called concept drift or conditional shift. The survey of concept drift adaptation (Gama et al., 2014) documents the relevance of this phenomenon for online supervised learning and temporal model monitoring.

Selection and observation are components of the shift

Introduce an observation indicator A, equal to one if a risk is included in the recorded training portfolio. Then the data available for estimation follow P(X,Y|A=1), not necessarily the distribution of all risks eligible for the contemplated insurance product. Underwriting acceptance, premium sensitivity, marketing reach, geographic eligibility and policy cancellation all affect A. If A depends on unrecorded determinants of Y after conditioning on X, no density ratio based solely on observed X can identify the full target outcome law.

Claims observations introduce another selection process. Let N denote the event of an insured incident, R its reporting and C its coverage recognition. A recorded claim indicator may be represented as Y_observed=N×R×C in a simplified binary setting. The expected rate of Y_observed changes when reporting or adjudication changes, even if physical incident frequency is stable. In practice payment timing, recoveries, limits, reserves and re-openings add further stages. Transfer must therefore specify which stage the model predicts.

Delayed target labels create asymmetric information: target inputs may be available immediately, but ultimate claims require development. A short-term distribution detector can assess P_T(X); it cannot establish stability of ultimate P_T(Y|X). This asymmetry is an identification limit, and prospective validation must be designed around it.

Common support and untransportable regions

The density ratio exists as a usable source expectation only when every material target covariate region has positive source probability. If a carrier expands into flood-prone construction types absent from its historical property portfolio, no amount of finite importance weighting can reconstruct the missing conditional claim law. Out-of-support predictions require new data, credible structural information, justified expert assumptions, or explicit abstention from automated reliance. A sophisticated learned representation can smooth across observed neighborhoods, but it cannot prove that wholly unobserved contractual or hazard relationships are stable.

A further complication is policy heterogeneity. A carrier might record a feature named “sum insured” in the same currency across products, while one product has replacement-cost coverage and another applies actual cash value. The data share a feature vocabulary but not an economically identical Y. Transportability analysis therefore begins with semantic reconciliation of exposure and outcome definitions.

Portfolio and Jurisdiction Change

Portfolio composition as an endogenous result

An insurer's portfolio is shaped by prices, coverage, capital appetite, underwriting rules, distribution channels, competitors and policyholder choice. The target population after a rate change is not necessarily the same population with a changed numerical premium. Higher-risk insureds may disproportionately renew or leave, and cross-subsidization may change through new acquisitions. Portfolio composition therefore may depend on the model's own decisions. A model trained under one pricing policy can encounter a changed insured mix under the pricing policy it subsequently enables.

A scientifically credible transfer compares at least four objects: the eligible population; the quoted population; the bound and exposed portfolio; and the subset that reports or settles claims. Each object has its own inclusion probabilities and time index. Using only policy-level training rows risks treating underwriting selection as if it were random sampling from the eligible population. This is particularly severe when a model is moved from an established carrier with several years of renewals to a new entrant whose business comes primarily from first-year digital acquisition.

Selection can also arise through different channel agreements. A broker may send certain risks to particular carriers, or an insurer may receive leads pre-screened by aggregators. Similar policy feature distributions do not imply identical unobserved risk appetite or claim behavior. Where the channel changes, the transported model requires fresh validation of the conditional response and inclusion mechanism.

Time and hazard regime

At time t, denote the predictive response by m_t(x)=E[Y_t|X_t=x]. Changing severity inflation, repair costs, judicial claims practices, disease incidence, weather trends and exposure usage can produce m_t(x)≠m_{t+1}(x). A calendar feature may help encode historical patterns but does not establish future stability, especially when extrapolating beyond training periods.

The observation window also matters. A 2023 motor policy could generate bodily injury claims notified in 2024 and developed years later. A 2026 extract of that policy year may contain incomplete ultimate severity, whereas older years have more developed claims. Apparent temporal drift can arise from development maturity rather than the underlying risk. Training on immature target outcomes without development correction creates spurious evidence of improvement or deterioration. Validation should use comparable valuation ages or explicitly model claim development.

Catastrophic hazards amplify nonstationarity. A model might fit past location-level average losses yet omit evolving wildfire fuel, flood defenses, building-code change, demand surge, or correlated event losses. Geographical transfer across coasts cannot be justified by matching a few property attributes. Portfolio accumulation and event dependence must be modeled separately from policy-level expected loss when the decision concerns capital or reinsurance.

Contracts and target semantics

The contractual Y may be incurred loss net of deductible, paid loss after exclusions, frequency of reported claims, number of accepted claims, ultimate claims cost, or a loss ratio. Transferring between these targets requires an explicit bridge. If policy limit changes from M_1 to M_2, then Y=min(L,M)−D with appropriate floor and coverage conditions changes even if physical gross L does not. A severity predictor for indemnity on M_1 is not automatically a predictor for M_2; observed capped payments cannot identify the tail above M_1 without supplementary data or assumptions.

Differences in currency, inflation adjustment, policy year, exposure time, earned versus written exposure, and claim closure are particularly important for multinational comparison. Nominal equality can conceal purchasing-power inequality and distinct contractual obligations. Reconciliation requires a documented unit dictionary, exchange or inflation convention as relevant, and controls for changes of product wording.

An explicit relation between source and target estimands should be recorded: the source conditional expected observed claim and the proposed target conditional expected covered obligation are equal only if the transformations between observation, coverage and accounting are identified. A statistical transportability proof cannot substitute for this insurance-specific bridge.

Jurisdiction and legal environment

Insurance law is predominantly implemented through national, subnational or supranational arrangements; there is no single globally operative underwriting algorithm rule. Transfers between territories may implicate permission to process data, lawful discrimination criteria, rating restrictions, local claims settlement law, insurance supervision and consumer remedies. Institutional differences can alter both the legal admissibility of variables and the empirical outcome process.

The IAIS artificial-intelligence Application Paper (2025) provides international supervisory guidance on governance, robustness, transparency, fairness and proportionality. Its status is guidance within the IAIS supervisory framework rather than a self-executing global statute. The EIOPA 2025 Opinion addresses European supervisory interpretation of insurance-sector AI governance. The NAIC 2023 Model Bulletin is a U.S. model bulletin whose operational force depends on relevant state adoption or supervisory practice. Source and target jurisdiction, date, policy function and applicable instrument must be recorded for each operational transfer.

The consequence is practical: an actuarially accurate transport that imports a prohibited rating variable is unsuitable for that market. Conversely, a jurisdiction-specific law that changes data availability or claims measurement creates a statistical shift that must also be assessed. Legal permission and statistical validity are two necessary inquiries associated with the same deployment.

Domain Adaptation and Reweighting

Estimating the density ratio

Under verified covariate shift, weights may be estimated directly or through a domain classifier distinguishing source X from target X. Consider a pooled dataset with indicator D=1 for target observations and D=0 for source. Let π=P(D=1) be the sampling fraction in the pooled design and q(x)=P(D=1|X=x). Bayes' rule yields

pT(x)pS(x)=q(x)1−q(x)1−ππ.\frac{p_T(x)}{p_S(x)}=\frac{q(x)}{1-q(x)}\frac{1-\pi}{\pi}.

This identity concerns the sampling scheme used to build the domain classifier; if the pooled source/target sample proportions are engineered, the correction factor must match that design. When q(x) approaches one, weights grow large, signaling weak source overlap or domain discrimination. Clipping weights trades some bias for reduced variance; it does not repair absent support.

A useful complementary diagnostic is the effective sample size,

neff=(∑i=1nwi)2∑i=1nwi2.n_{\mathrm{eff}}=\frac{\left(\sum_{i=1}^{n}w_i\right)^2}{\sum_{i=1}^{n}w_i^2}.

This expression equals n when all weights are equal and becomes smaller as weights concentrate. It is a descriptive measure of weight concentration, not a universal confidence interval or a replacement for clustered exposure, serial dependence and parameter estimation uncertainty. A portfolio with one region accounting for nearly all the weighted loss estimate may have tens of thousands of records but little information relevant to that target region.

Importance weighting can adjust estimation, model selection and validation when its assumptions hold. The Sugiyama et al. (2007) analysis specifically studies the failure of ordinary cross-validation under covariate shift and motivates importance-weighted alternatives. Reweighting a validation set based on estimated target density ratios is useful only when that set contains relevant source cases and leakage from the target evaluation outcomes is prevented.

Distinguishing diagnostics from guarantees

Two-sample procedures can test for evidence that covariate distributions differ. The kernel two-sample test of Gretton et al. (2012) uses maximum mean discrepancy. Domain classifiers measure how separable source and target covariates are; a strong classifier may reveal a substantial difference, but classification accuracy alone does not quantify target prediction loss.

The Failing Loudly study (Rabanser, Günnemann, & Lipton, 2019) compares shift-detection approaches and investigates the malignancy of detected changes. Its empirical results support treating detectors as diagnostic instruments. An insurer should therefore relate detection thresholds to subsequent predictive error, calibration changes, and financial sensitivity in historical or prospective target cohorts. A detector that alarms on harmless shifts in policy identifiers can consume governance resources while missing a modest shift in a high-cost claims subgroup.

A simple algebraic example shows the issue. Suppose the only changing covariate is a marketing-channel code independent of loss conditional on the remaining predictors. P_T(X) may differ markedly from P_S(X), and a two-sample detector may reject equality. Yet a well-specified predictor independent of that code can retain target risk. Conversely, a small shift in a rare high-severity group may have limited statistical power but high capital consequences. Shift detection and decision-impact assessment must be coupled.

Weighted estimation and finite-sample limitations

For a weighted mean target loss one might calculate

μ^T=∑iwiYi∑iwi,\widehat{\mu}_T=\frac{\sum_iw_iY_i}{\sum_iw_i},

where weights are derived from source-target covariate densities. The self-normalized estimator changes the exact unbiasedness properties of an unnormalized inverse-probability estimator but often improves numerical stability. Its reliability depends on estimated weights, positivity, tail behavior and the stability of E[Y|X]. For heavy-tailed insurance severity, rare large payments interacting with large weights can dominate both point estimates and variance.

Fitted propensity ratios and model hyperparameters must be constructed without leaking evaluated outcomes. Cross-fitting can reduce overfitting in nuisance estimates, but it does not make an untestable invariance claim true. Uncertainty should account for estimation of both model parameters and density ratios, including clustering by policyholder, household, intermediary, catastrophe event or exposure group as applicable.

A useful empirical comparison includes unweighted baseline; fixed-source model on target; reweighted source fit; target-only fit on available target labels; source pretrained plus target fine-tuning; and, where justified, a robust or hierarchical model. The desired outcome is target population performance under a prespecified scoring and decision criterion, not the numerical size of the training-data objective.

Label shift correction and its boundaries

For categorical Y, black-box shift estimation can infer new class priors from target predicted-label frequencies and the source confusion matrix. Its identification depends on stable class-conditional feature distributions and appropriate invertibility. The Lipton et al. method does not establish that all insurance fraud, claim or medical-risk classifiers admit this decomposition. In particular, investigator workload, fraud-detection strategy and claim definition can change P(X|Y), invalidating the assumed mechanism.

For regression of claim severity, adapting only class priors is generally an inadequate operation. A continuous severity outcome demands a different model of changed outcome law or a discretization whose information loss and contract boundaries are defended. If a thresholded classifier is used for operational triage, validate that threshold's stability and the costs of false positives and false negatives in the target population.

Transfer Learning of Features

Task and representation compatibility

Transfer learning encompasses multiple operations: initializing target models from parameters learned on source data; reusing a representation of X; adapting a feature extractor while refitting a target output layer; or combining data under regularized multitask estimation. The transfer-learning survey (Pan & Yang, 2010) provides the broad research classification. Each procedure presumes some reusable statistical structure. In insurance, that structure may be a relationship between vehicle characteristics and collision risk, claims text and settlement category, property images and construction features, or exposure seasonality and frequency.

Shared features need not imply shared outcomes. A text encoder trained on claims documents in one language may extract some transferable semantic content, but the meaning of a claim category depends on local legal, policy and administrative conventions. A property-image network may transfer visual patterns yet perform poorly when local building materials, images, hazards or underwriting conventions differ. The representation should be evaluated for predictive adequacy and decision relevance in target cohorts.

Feature transfer is especially attractive when target labels are sparse and source data are abundant. It is also especially vulnerable to negative transfer when abundant source data dominate a small but structurally different target pool. The main empirical question is whether transfer improves target performance over an appropriate target-only baseline and simple actuarial model at the same target-label budget.

Invariance and latent differences

Some domain-adaptation methods seek a transformed representation Z=g(X) for which P_S(Z) resembles P_T(Z). Matching feature distributions is insufficient by itself. If P_S(Y|Z) differs from P_T(Y|Z), a domain-invariant representation can discard the very differences necessary to model target claims. The domain adaptation theory (Ben-David et al., 2010) exposes the role of the best shared hypothesis and domain discrepancy; alignment cannot eliminate irreducible conditional disagreement.

Insurance covariates can be proxies for distinct latent processes. The same vehicle age may correspond to different traffic environments or repair costs; the same occupation label may be measured under different labor-market taxonomies; a postal area may change boundaries over time. When a transferred representation suppresses source-target differences without explaining outcome mechanisms, a model may appear stable on aggregate yet fail within a newly introduced product or minority subgroup.

Transferability should be tested at multiple feature resolutions. A source encoder can be frozen initially, then compared with partially fine-tuned and fully retrained alternatives using a strictly target-held-out evaluation. Evaluation must record performance by policy cohorts and time, not just overall loss. If target sample size is small, uncertainty intervals and selection-adjusted results matter more than slight nominal ranking differences.

Group and worst-case performance

Aggregate expected loss can conceal weak generalization in vulnerable or commercially significant groups. Group distributional robustness research (Sagawa et al., 2020) studies worst-group generalization and the role of regularization; its empirical benchmark results are not insurer results. A related insurance use is to inspect model performance in defined territory, product, tenure, claim-size and demographic slices where lawful, appropriate and adequately supported.

A group-robust objective might minimize max_g R_g(f) over identified groups. This changes the loss criterion from average performance to worst-observed-group performance. The choice of groups, their representativeness and the policy consequence of balancing accuracy across groups are substantive governance decisions. It is inappropriate to treat the existence of a mathematical robust objective as proof of fairness or legal compliance.

The appropriate evidence for transfer learning is a direct comparison under target holdout data, aligned definitions, and enough labeled observations to distinguish an improvement from sampling fluctuation. Claims of universal cross-country or cross-insurer transport are stronger than available evidence supports unless validated across those contexts.

External Validation

Validation follows the deployment clock

A deployment question is inherently prospective. Random train-test splits that mix policy years, neighboring renewals, duplicate policyholders and reopened claims can leak information across the chronological boundary. A primary evaluation should emulate the intended decision: train only on records known as of time t, select parameters using an earlier validation period, and evaluate claims associated with policies written after t using outcomes developed to a comparable valuation age.

An insurer-to-insurer evaluation should use an actually external portfolio when feasible. A source carrier's random holdout tests within-carrier generalization, not external validity. A transfer claim requires the intended target population or a justified natural experiment that closely reproduces its features and outcome process. When target labels are not yet mature, interim metrics and future revalidation should be declared separately.

The WILDS research (Koh et al., 2021) demonstrates why natural environmental splits can reveal failure that random benchmark splits miss. The lesson for actuarial design is to create environment- and time-aware holdouts aligned with business deployment. It does not justify importing published benchmark scores into insurance forecasts.

Loss, calibration and aggregate accuracy

Performance should be measured according to a defined estimand. For binary occurrence, log loss, Brier score, calibration plots and observed-to-expected ratios can complement discrimination. For severity or aggregate losses, suitable proper loss functions, tail-specific measures, quantile coverage and aggregate bias may be informative. AUC or rank correlation may indicate ordering but can coexist with severe premium-level miscalibration.

Calibration requires attention because the target base rate can change even when rank discrimination remains excellent. The Guo et al. (2017) calibration study evaluates modern neural-network probabilities on noninsurance datasets; its general lesson is to distinguish probability quality from classification accuracy. Recalibrating on a small new target sample may improve estimates but risks instability, particularly for rare claims, and must be validated on a further holdout.

Insurance model assessment should report performance in financial units. If the model predicts mean policy loss, the difference between total predicted and eventual covered loss across a cohort may influence premium adequacy and reserves. The acceptable tolerance depends on the model's role, portfolio scale, aggregation method, regulatory context and asymmetry of decision costs. The same one-percent frequency miscalibration may be financially tolerable for a small routine book and consequential for a high-severity catastrophe concentration.

Predictive intervals under shift

Ordinary interval guarantees based on exchangeability can fail when the target distribution differs. The weighted conformal prediction result (Tibshirani et al., 2019) extends certain coverage results under covariate shift and weighted-exchangeability conditions. It should not be summarized as a guarantee under any arbitrary time, label, concept, measurement or structural change. Weight estimation and support conditions require explicit study.

For uncertain claim severity, intervals should account for changing tails and development lags. An interval with nominal marginal coverage may offer inadequate conditional protection in a small but financially important region, and dependence between exposures can invalidate naïve independent-case reasoning. Reporting uncertainty is therefore multidimensional: conditional estimates, aggregate reserve or premium sensitivity, weight concentration, subgroup calibration and data-development uncertainty deserve distinct treatment.

The uncertainty-under-shift study (Ovadia et al., 2019) compares techniques under distribution changes and reports that some ordinary post-hoc calibration methods can degrade. Such general evidence motivates target-domain stress testing rather than confidence in a one-time source calibration operation.

Monitoring after initial validation

A deployed system needs an observation plan linking inputs, outcomes and decisions over time. Monitor feature shifts, missingness, policy mix, class or outcome prevalence when matured, calibration, financial loss and relevant subgroup metrics. Alarms should have a documented response: investigate data pipeline changes; determine whether the target estimand changed; identify out-of-support risks; sample more target labels; refit or recalibrate where justified; escalate or suspend use when reliability cannot be established.

Monitoring creates a delayed-feedback problem: claim outcomes may not be known for months or years, while new policies are quoted daily. Leading data-quality and input-drift indicators can prompt investigation but cannot establish updated outcome accuracy before labels develop. The distinction between an early warning signal and scientifically confirmed predictive deterioration must remain explicit.

Insurance Transportability Limits

Why statistical matching can fail actuarially

An insurance portfolio is not an interchangeable matrix of predictors and labels. Coverage attaches legal meaning to the loss variable; exposure denominators specify time at risk; underwriting and renewal policies select who is observed; claim reporting and payments introduce delays; capital exposure depends on tail and dependence characteristics. A model may pass a covariate-shift test and nevertheless have a different target owing to altered coverage. Conversely, it may detect a visible change in inputs without a meaningful change in conditional expected claims.

The insurance literature in Predictive Modeling Applications in Actuarial Science (Frees, Derrig, & Meyers, 2014) establishes the actuarial context for model design. This dossier adds a detailed comparison of transportability assumptions, rather than asserting that the cited volume tested the hypothetical examples below.

The first limit is identification. Observing two marginal covariate distributions cannot reveal whether P_T(Y|X)=P_S(Y|X) when target labels are absent. A rich set of proxy variables is still incomplete when important determinants are unrecorded. Domain adaptation can be useful under reasonable assumptions, but usefulness must be differentiated from identification certainty.

The second limit is overlap. Density ratios can reweight observed source risks into target proportions; they cannot create claims outcomes for target risk types that source underwriting systematically excluded. Algorithmic extrapolation into these regions may be useful as a scenario assumption but should be communicated as an extrapolation.

The third limit is outcome definition. A target change from reported count to ultimate paid loss, from gross to net loss, or from one policy form to another requires redefining Y and potentially an actuarial bridge. A transfer-learning method that treats these as identical labels estimates a mixture of distinct financial objects.

The fourth limit is dependence. Large insurance losses can be dependent across insureds through catastrophe events, litigation regimes, supply-chain disruptions or weather. Excellent independent-policy predictive mean performance does not by itself establish accurate aggregate capital, value-at-risk or reinsurance attachment probabilities.

The fifth limit is feedback. Risk scoring influences prices, acceptance, investigation and fraud detection. The deployment itself can alter the distribution subsequently observed. Static source-target descriptions must therefore be periodically revisited when the decision policy changes.

The sixth limit is operational and legal scope. Data shared by two insurers may have different processing rights, definitions, quality and governing law. A model transferred under contract must comply with the relevant restrictions and be documented in the target environment. Legal validity is examined for the deployment jurisdiction and actual use, rather than inferred from statistical precision.

Decision-level sensitivity

Let premium revenue or expected loss decision depend on a prediction μ. If the pricing action is a=g(μ), then target loss from using source model f should be assessed through the consequence of action a, not only generic prediction error. A fixed 5% error in pure premium may affect different products differently if expenses, limits, reinsurance, consumer demand and regulation vary. Model selection can therefore reasonably employ an actuarially relevant weighted objective while also reporting standard statistical metrics.

For a proposed transfer the decision question can be formulated as: under explicitly stated ranges of target shift and model error, which actions remain acceptable? This creates sensitivity tables for prediction error, rate adequacy, expected claims cost, and any legally constrained outcome. Sensitivity analysis is not a substitute for empirical target performance, but it makes decision exposure transparent when labels are immature.

Evidence quality hierarchy

Direct evidence from a target insurer's future holdout cohort is most relevant for a specific deployment if leakage, maturity and selection are controlled. Next are multiple independent external portfolios with suitably aligned contracts and exposures. Internal temporal validation is informative but more limited. General algorithmic research establishes possibility and method properties, not the realized magnitude of effect in a particular insurance system. Supervisory publications specify expectations and controls, not external validation scores. This hierarchy prevents transferring claims of empirical benefit from a noninsurance benchmark into an insurer without corroboration.

Quantitative Illustrations

Illustration A: reweighting a changed regional portfolio

Consider a hypothetical single-year motor book with two mutually exclusive exposure groups, urban (U) and rural (R). Assume each policy contributes exactly one unit of earned exposure, claims are binary for illustration, and the conditional annual probabilities of at least one covered claim are stable across source and target: p(Y=1|U)=0.12 and p(Y=1|R)=0.04. In the source portfolio, 20% of exposures are urban and 80% rural; in the target, 60% are urban and 40% rural.

The source mean occurrence probability is 0.20×0.12+0.80×0.04=0.056. The target mean, by direct stratification, is 0.60×0.12+0.40×0.04=0.088. A source aggregate rate of 5.6% used unadjusted in the target would underpredict the target mean by 3.2 percentage points, or about 36.36% relative to the true target rate, even though the within-group model is perfectly calibrated by assumption.

The density ratios are w(U)=0.60/0.20=3.0 and w(R)=0.40/0.80=0.5. Under an exactly representative hypothetical source sample of 1,000 policies, 200 urban and 800 rural, the weighted expected number of occurrences is 200×3×0.12+800×0.5×0.04=88. The sum of weights is 1,000, so the weighted target occurrence frequency is 0.088, matching the direct target stratification.

The weight-based effective sample size is (1,000)^2/(200×3^2+800×0.5^2)=1,000,000/(1,800+200)=500. This is a loss of half the unweighted sample-size equivalent despite retaining every original source policy. Actual interval uncertainty would additionally depend on Bernoulli outcomes, dependence and estimation error; the number 500 is an illustrative weight concentration diagnostic, not a validated actuarial precision statement.

The example establishes a mathematical possibility, not a general result about region or motor insurance. The correction rests completely on stable conditional claim probability, correct grouping, equal exposure length, and support in both regions. If the target urban conditional risk becomes 0.15 because the driving regime or reporting process changed, target mean becomes 0.60×0.15+0.40×0.04=0.106. Reweighting with the old conditionals still yields 0.088, understating the true target risk by 0.018. A density ratio cannot repair this conditional change.

Illustration B: contractual change and false transport

Consider a hypothetical positive gross property loss L with two possible values: 8,000 monetary units with probability 0.9 and 80,000 with probability 0.1, conditional on a covered occurrence. Suppose the source policy covers gross losses up to 20,000 with zero deductible; the target policy covers them up to 60,000. Both policies are written on the same physical exposure, and assume for illustration there is no difference in incident occurrence, inflation or claim reporting.

The source expected covered severity per occurrence is 0.9×8,000+0.1×20,000=9,200. The target expected covered severity is 0.9×8,000+0.1×60,000=13,200. A model transferring source paid severity unchanged understates the target mean by 4,000 per occurrence. The target covered cost exceeds the source mean by approximately 43.48%.

Here P_S(X) can equal P_T(X) exactly, and the physical gross-loss distribution can also be invariant. Yet the modeled Y—covered payment under the policy terms—differs. The failure therefore arises from a change in outcome semantics rather than a conventional covariate shift. If source paid-loss data are censored at 20,000 and no gross-loss amount above the limit is observed, source data alone cannot tell whether those losses are 30,000, 80,000 or 1 million. The target tail is unidentified without additional information. The numerical probabilities and amounts in this illustration are stipulated, not inferred.

This example explains why transport across jurisdictions or product generations requires a contract-to-data bridge before a statistical algorithm is selected. It also shows why external validation that uses source coverage definitions to score target policies can appear favorable while evaluating the wrong financial quantity.

Illustration C: detecting changed class prevalence

Suppose a source binary operational fraud screen has true-positive rate 0.80 and false-positive rate 0.10, and the target system is assumed to preserve these conditional properties. Let the target fraction classified positive by the model be 0.17. If π_T is the unknown target fraud prevalence, then 0.17=0.80π_T+0.10(1−π_T). Algebra gives π_T=(0.17−0.10)/0.70=0.10. This is an elementary special case of label-prior identification, conditional on invariant error rates and a nonzero sensitivity-specificity gap.

If investigator practices change, or screening itself changes which events are labeled fraud, the assumed rates may fail and the implied prevalence can be misleading. Moreover, a fraud label can refer to suspicion, investigation outcome or legally established fraud. Operational interpretation depends on which event was measured. This is a demonstration of assumptions rather than a factual estimate of insurance fraud.

Comparative lessons

The three examples identify different origins of apparent predictive displacement: changed exposure weights with stable conditional claims, changed contractual payment semantics with identical physical risks, and changed outcome prevalence under a restricted classifier-invariance model. The correct response differs in each. Weighted validation is defensible in the first case, contract-specific outcome modeling is required in the second, and prior adjustment is conditionally available in the third. Treating all three as a generic “retrain the model” instruction obscures the scientific questions that determine whether retraining is valid.

Supervisory, Actuarial, and Institutional Governance

Professional modeling guidance

In the United States, Actuarial Standard of Practice No. 56 (Actuarial Standards Board, 2019; effective 2020) addresses model design, development, selection, modification, use, review and evaluation within its professional applicability. Applying the standard to a transferred actuarial model calls for clear documentation of purpose, model limitations and validation appropriate to intended use. The standard is a professional actuarial instrument with defined scope; it is not a single worldwide insurance-model statute.

A transport review should document the origin and target population, policy inclusion rules, intended claims outcome, valuation date, model version, training data history, available target labels, evidence of overlap, reason for believing conditionals are stable, sensitivity to weighting, and anticipated decision costs. Such a record supports technical review and gives future validators a reproducible reference for monitoring changes.

Model governance should also distinguish ownership. A vendor may supply a model pretrained on broad third-party data while the deploying insurer remains responsible for its own product and jurisdictional use. A vendor's generic benchmark report cannot establish the model's performance on the insurer's data unless actual target validation is performed. Relevant contractual permissions, vendor documentation and data lineage need examination before the model enters operational decisions.

IAIS and supervisory interpretation

The IAIS 2025 Application Paper discusses proportional supervision, governance and accountability, robustness and safety, transparency, fairness and redress. This supports attention to transfer uncertainty and model-use controls. Since the paper is an international supervisory application paper, the operative legal requirements for any insurer remain those of the relevant supervisory authority and domestic or applicable regional framework.

For a large or financially sensitive transfer, governance could require an independent validation team, scenario and sensitivity analyses, escalation when support fails, monitored re-estimation of target calibration, and procedures for corrective action. The threshold and form of these controls depend on the insurer's situation and applicable requirements, rather than being mechanically derived from the existence of distribution shift.

European supervisory perspective

EIOPA's AI governance Opinion (2025) clarifies expectations under the European insurance-sector legislative framework for AI systems within the Opinion's scope. It follows a proportional approach, including data governance, record-keeping, fairness, cyber security, explainability and oversight. It expressly operates as supervisory interpretation rather than independent creation of new general obligations. Transfer of an insurance model between EU contexts therefore requires attention to the model's function and relevant rules, including any separate AI Act requirements where applicable; this monograph does not treat the Opinion as displacing legislation.

Insurance model transfer can raise data minimization, lawfulness, confidentiality, and discriminatory impact issues. Technical accuracy does not resolve all such questions. A regulator may have to understand which variables create the learned relationship and why the transfer remains justified in the local market. Reusing the same parameter estimates across jurisdictions without examining lawful use and underlying coverage definitions is an operational governance failure even if the original model was well validated.

U.S. state insurance supervision

The NAIC Model Bulletin on AI Systems (adopted December 2023) describes expectations for insurer governance, predictive model validation and avoidance of adverse consumer outcomes. It is a model regulatory document addressed to state implementation and supervision, and not a uniform federal law. Transfer documentation must therefore identify the insurance department and actual applicable requirements, distinguishing NAIC model text from enacted or formally adopted local expectations.

The core methodological implications are widely generalizable without asserting a universal regulatory duty: target validation, explainable changes, adequate data documentation, appropriate third-party oversight, and corrective monitoring support reliable transfer. In specific legally governed deployments the final compliance determination belongs to the operative jurisdictional sources and qualified responsible institutions.

Reproducibility and Research Agenda

A reproducible transport study requires an explicit protocol rather than a generic statement that historical records were split into training and test partitions. The protocol should name the source and target insurer or population where disclosure permits, time period, product, exposure definition, policy and claim target, legal coverage conditions, outcome development horizon, predictor availability cutoff, missing-data conventions and unit conversions. It should also preserve preprocessing, feature selection, random seeds, learned weights, model and calibration parameters and scoring definitions.

An academically credible comparison preregisters or freezes the evaluation design before inspecting final target labels. It includes an honest source-only baseline, a straightforward target-only baseline when labels permit, and one or more candidate adaptation procedures. Report score differences and relevant uncertainty in addition to point estimates, with appropriate clustering and awareness of multiple model comparisons. The protocol should include subgroup and decision-cost analyses where the use case requires them.

For insurer-to-insurer transfer, access agreements and confidentiality can prevent public release of observations. Reproducibility can still be strengthened through documented schema, executable pseudocode, de-identified aggregate audit tables, synthetic examples that are clearly labeled as such, controlled-access replication scripts and independent validation records. Synthetic data never establish actual insurance prediction benefit.

Research remains particularly valuable in five areas. First, studies can separate changes of policy definition from statistical covariate and conditional shifts. Second, source-target density ratios and effective sample sizes should be evaluated under realistic policyholder and catastrophe dependence. Third, target-label delay should be incorporated explicitly into real-time monitoring and corrective procedures. Fourth, multimarket transfer needs datasets that preserve jurisdictional metadata and lawful feature definitions. Fifth, studies should report both actuarial and consumer consequences of negative transfer rather than aggregate predictive scores alone.

The strongest future evidence would come from carefully designed international, multi-insurer research with temporally honest external cohorts, transparent product and claims ontologies, mature target outcomes and explicit statistical uncertainty. That aspiration is a research agenda, not a claim that such a comprehensive benchmark currently exists.

Conclusion

Insurance prediction can be transported from a source portfolio to a target portfolio when the relevant statistical relationships, outcome definitions, exposure mechanisms and decision conditions can be defended for the proposed use. Covariate shift permits principled density-ratio reweighting under conditional invariance and common support. Label shift permits restricted prior correction under stable class-conditional data and identification conditions. Concept drift, changed coverage, selection on unmeasured variables and absent support require further target evidence or revised model specification. Transfer of learned representations may save scarce target labels, but improvement is established through external target evaluation rather than through the act of transferring parameters.

The three hypothetical illustrations demonstrate complementary failures and remedies: a portfolio mix shift can alter aggregate claims with fixed within-group risk; a higher contractual limit changes the paid-loss estimand even with identical underlying physical risk; and label-prior recovery is possible only under explicit classifier-invariance assumptions. These are deductions from stipulated conditions and are presented with their full numerical premises.

For insurance science the decisive object is target-domain validity of the specified prediction in its contractual and institutional setting. Model quality is demonstrated by coherent estimands, defensible identification, representativeness, temporally honest external validation, calibration, uncertainty disclosure and decision-aware governance. Academic domain-adaptation research provides the mathematical tools; actuarial practice and applicable supervisory frameworks determine how those tools must be examined and controlled in concrete insurer operations. The monograph's original contribution is the integrated reasoning and transparent illustrations that make these relationships auditable, with no claim of a new theory or empirical discovery.

References