Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Explainability and Fairness in Actuarial Models

Abstract. Actuarial models translate uncertain future insurance losses into estimates, classifications, prices, underwriting decisions, and capital or claims-management signals. Their growing statistical complexity creates a compound problem: the insurer must understand what the model estimates, establish why its variables and relationships are suitable for an insurance purpose, measure how its outcomes vary among relevant populations, and provide a defensible process for reviewing consequential individual decisions. Explainability and fairness therefore require distinct statistical, institutional, legal, and procedural tests. This monograph develops a comparative analytical account of those tests through the foundations of generalized linear models, dependent-feature explanations, local attribution, counterfactual analysis, calibration, error distributions, proxy discrimination, group disparity, decision review, and actuarial model governance. It examines insurance-specific scholarship alongside primary regulatory materials from the United States, European Union, United Kingdom, and Canada. Transparent hypothetical calculations show why a model may be calibrated while producing different errors across groups, why correlated proxy variables can generate premium disparities without measured differences in underlying expected claims, and why adjustments that improve a group metric may change the model's actuarial and legal properties. The central conclusion is that explainability provides evidence about model operation, fairness criteria evaluate selected properties of distributions and decisions, and adjudicable insurance decisions require a traceable connection between reliable evidence, authorized use, and practical correction. These are original analytical findings assembled from established research, without asserting a new actuarial theory or an empirical discovery. The monograph specifies its mathematical assumptions, the bounds of comparative law, and requirements for reproducible audit.

Keywords: Actuarial models; explainable artificial intelligence; insurance pricing; generalized linear models; underwriting; algorithmic fairness; actuarial equity; proxy discrimination; disparate impact; demographic parity; calibration; prediction error; SHAP; LIME; counterfactual explanations; model governance; insurance regulation; contestability; data quality; fairness auditing.

Introduction

The actuarial use of statistical models rests on a concrete institutional responsibility: a prediction about future loss or behavior becomes consequential when it shapes the price, availability, benefit, or administration of insurance. Such predictions arise from historical records, exposure definitions, distributions, observations subject to selection, and judgments about the period in which the model will operate. Their credibility can be evaluated empirically. The acceptability of applying them to particular people and purposes also depends on the product, governing law, permissible data, contracting conditions, and available review process. A scientifically defensible interpretation must retain each of these components.

The mathematical history is instructive. The generalized linear modeling architecture established by Nelder and Wedderburn (1972) expresses a conditional mean through a link function and specified covariates, permitting interpretable actuarial frequency and severity models. Contemporary boosting, ensembles, neural networks, and vendor-produced scores expand the range of functional relationships, interaction effects, and data types. New predictive capacity enlarges the importance of validation and meaningful accounts of how data reach a decision. The insurer still faces the original actuarial task of estimating future obligations, now with additional sensitivity to representation, causal interpretation, and delegated decision systems.

Insurance-specific scholarship establishes the necessity of separating different meanings of fairness. Lindholm, Richman, Tsanakas, and Wüthrich (2024), in their analysis of proxy discrimination, formally distinguish the use of proxy information from demographic disparities of outcomes and show that meeting one criterion does not imply satisfying the other. Xin and Huang (2024) survey regulations, fairness criteria, and models in the specialized setting of insurance pricing. These studies locate the problem at the intersection of predictive accuracy, actuarial classification, protected attributes, and the rules under which an insurance contract is offered.

The research question is therefore precise: under what evidentiary and operational conditions can an actuarial prediction be understood, evaluated for legally and ethically relevant fairness, and challenged or corrected when it affects the insured? Addressing that question requires defining the population and target first, specifying the meaning of each fairness metric, analyzing the explanation method's assumptions, and identifying the institutional decision maker. A technically accurate attribution may be inappropriate as an answer to an individual claimant; a policyholder-facing explanation may be understandable yet omit important statistical uncertainty; and a group statistic may flag unequal outcomes without identifying their cause. These differences organize the chapters that follow.

This monograph concerns actuarial prediction for insurance and related underwriting, pricing, valuation, and claims decisions. It surveys methodological principles with the deepest applied focus on classification and personal-lines or life/health pricing. Jurisdiction-specific rules are reported according to their stated scope, rather than generalized into one international prohibition. All numerical portfolios in the text are constructed thought experiments; they represent no insurer's experience, identified individual, or estimated market statistic.

Research Object, Terminology, and Scope

An actuarial model is a documented mapping from data and assumptions to quantities relevant to insurance obligations or decisions. It may estimate frequency, severity, pure premium, mortality, morbidity, lapse, retention, or claims propensity. For observation i, the prospective expected claim cost can be represented by μ_i = E[L_i | X_i], where L_i denotes loss measured on a specified exposure and time basis and X_i the variables available at the underwriting or valuation decision. In practical pricing, the final customer premium can additionally include expenses, taxes, risk margins, demand or renewal factors, regulatory constraints, and portfolio-level adjustments. The equality of a model prediction and a quoted premium therefore cannot be presumed.

Interpretability concerns how readily a model's internal relationship between variables and outputs can be understood. Explainability concerns a representation of model behavior or a particular prediction that another person can inspect and test. Model transparency includes data provenance, assumptions, versions, algorithmic specification, testing, decision authority, and records of overrides. These terms describe overlapping but distinct objects. An interpretable formula with undocumented data quality is insufficiently transparent; an explainability plot generated for a complex model offers a partial view of the prediction mechanism; a contractually consequential explanation must relate that mechanism to the decision actually taken.

Fairness in the actuarial setting is a specified property of treatment, distribution, opportunity, or procedure. Statistical definitions include equalized errors, parity of acceptance or premium distributions, and calibration within defined groups. Legal definitions depend on protected categories, sector-specific prohibitions, permitted distinctions, and any duty to justify differential treatment. Economic assessment may examine claims-cost adequacy, adverse selection, long-run pooling, availability, cross-subsidy, and affordability. Procedural fairness evaluates disclosure, correction of mistakes, review of exceptions, and accessible remedies. These dimensions sometimes reinforce one another and sometimes compete for the same scarce information or premium capacity.

A predictive factor's statistical association with observed loss has three separate interpretations. It may be a reliable forecast signal in the observed sample; it may reflect a causal exposure or an administratively correctable risk; and it may be a factor the law permits to determine terms of insurance. One inference does not logically entail the others. The New York insurance AI circular letter consequently asks for actuarial support, proxy assessment, examination of differential effects, and insurer responsibility for outsourced systems. This regulatory pattern is particularly useful as an operational illustration of why statistical prediction alone cannot settle acceptability.

Model development and model use must also be distinguished. The raw training data can contain attributes retained for statistical control, protected-class assessment, sampling analysis, or bias detection, while the approved rating function may be restricted from using particular attributes. Such data separation raises governance requirements: lawful access, purpose limitation, restricted use in decision engines, role-based protections, and traceable deletion or correction. A group attribute necessary for auditing is not automatically an admissible rating factor. Conversely, prohibiting its direct use in the rating engine does not establish that the remaining variables are free from proxy influence.

Risk Variables and Proxy Effects

The structure of an actuarial signal

A rating variable may be an exposure measure, a direct physical risk characteristic, a historical response variable, a behavior indicator, a geographical aggregate, or an inferred profile. The statistical estimator depends on data-generation mechanisms. Observed claim count is shaped by exposure duration, coverage terms, reporting practices, retention and deductible selection, claim settlement, inflation, and whether an applicant entered the insured portfolio at all. This creates a distinction between the true loss distribution in the target market and the distribution observed among existing policyholders. A model trained on the latter can inadvertently learn the insurer's earlier underwriting and claims practices.

Consider a Poisson frequency model: N_i | X_i ~ Poisson(e_i λ_i), with log λ_i = β_0 + Σ_j β_j x_ij and e_i an offset for exposure. For a severity model, a gamma conditional distribution with log link may predict the mean of positive loss. The resulting pure premium under specified frequency-severity independence assumptions is E[N_i | X_i] × E[S_i | X_i]. This factorization is useful because its interpretability follows from component definitions, yet the model's consequences still depend on how covariates were coded, whether interactions were omitted, the experience period, and the changing contract. The classical generalized linear model formulation provides the statistical architecture and does not itself certify a permissible social use.

Risk variables often carry information about the insured through systems unrelated to the claimed underlying exposure. Postal geography can encode road conditions, catastrophe susceptibility, service access, wealth, past redlining, and local demographic composition. Credit-derived measures can encode financial shocks, data errors, and group histories. A device or mobility signal can reflect both contemporaneous behavior and unequal access to tracking technology. The question for an auditor is how a variable contributes to anticipated contractual cost and whether its observed association is stable, causally interpretable where needed, and permitted for the particular decision.

Direct discrimination, proxies, and mediated associations

Let A denote a protected characteristic and X a set of allowed inputs. Direct use occurs if the decision function receives A as a feature or a derivative explicitly constructed from A. Proxy dependence arises where model output depends on features that reveal information about A even when A is not provided as a named input. Group disparity refers to a difference in distributions or treatment conditional on A, whether or not A is used in the algorithm. These descriptions answer different questions. The actuarial study What Is Fair? demonstrates their mathematical non-equivalence and warns against collapsing fairness criteria into one label.

A causal diagram helps organize a proxy investigation: A may be associated with Z through historical and structural pathways; Z may influence actual insured risk, access to coverage, claim reporting, or only the sample's composition. If Z predicts reported claim costs because previous insurance access varied with A, then the regression can reproduce a historical selection mechanism. Removing the name of A from the feature list leaves that mechanism intact. On the other hand, an association between A and a physically meaningful and legally permissible factor may require contextual interpretation rather than a blanket inference of illegality. Proxy identification needs conditional analysis, comparisons of suitable alternatives, and an explicit legal framework.

The methodological question is also about measurement. If investigators use imperfectly inferred protected attributes for fairness audits, group-specific rates inherit estimation error. For a misclassified group indicator A*, a measured disparity E[Ŷ|A*=1] − E[Ŷ|A*=0] need not equal the disparity under true A. Misclassification may attenuate, exaggerate, or reverse comparisons depending on group prevalence and the error structure. The uncertainty from group inference should be reported rather than buried in a deterministic fairness dashboard. Restricting direct access to audit attributes while preserving auditable aggregate computations helps separate validation from pricing use.

Selection, leakage, and historical feedback

A model is exposed to leakage if a feature contains information that would not have been available at the intended decision time or reflects the target through later administrative events. A claims-fraud model that relies on post-investigation documentation to predict whether a new claim should be investigated may exhibit spectacular validation performance and fail at deployment. A renewal model may learn intervention policies that created observed retention outcomes. Historical premium or prior refusal can embed legacy classification decisions, so the downstream model may effectively inherit a prohibited proxy without reproducing the underlying raw attribute.

Actuarial data-quality requirements apply here before fairness metrics are interpreted. The Actuarial Standards Board's ASOP No. 23 on Data Quality addresses selection, review, and use of data, including material limitations and appropriate disclosures. Its professional scope must be respected, but its diagnostic questions travel well across practice: Where did the observations come from? What exclusions or missingness apply? Are definitions consistent across periods? Can the analyst recognize obvious anomalies, assess the data's suitability for the intended purpose, and explain limitations to intended users? A fairness result computed from unstable labels and unresolved selection biases has weak evidentiary value regardless of mathematical sophistication.

Stability as a fairness condition

When the environment changes, the relation between X and L changes. A driver risk score may incorporate telematics that behave differently after a device upgrade; a health pricing model may encounter diagnostic coding changes or treatment innovation; catastrophic loss assumptions may move with climate exposures. A fairness audit performed on a historical cross-section may no longer describe the deployed system. The appropriate unit of observation is not only the algorithmic function but the versioned algorithm, feature pipeline, population, contractual decision and assessment date. This foundation will later determine the monitoring design.

Actuarial Fairness Criteria

Formal criteria and their denominators

Let A index groups, Y be an observed outcome, s(X) a continuous risk score and D an operational decision. For a binary decision, demographic parity is P(D=1 | A=a) equal across named groups. Equal opportunity may be specified as equality of true-positive rates conditional on Y=1. Equalized odds additionally requires corresponding false-positive rates to be equal conditional on Y=0. Predictive parity or calibration may require P(Y=1 | s(X)=p,A=a)=p across groups. The choice of Y and the meaning of D are essential: approval, declination, fraud referral, expected claim, actual claim, and renewal premium are distinct targets.

Insurance pricing commonly models loss cost as a continuous, often zero-inflated, heavy-tailed quantity rather than a binary classification label. In this setting, one calibration property is E[L | μ̂=m,A=a]=m, assessed in sufficiently populated bins or with smooth calibration curves. For a subgroup g, an aggregate actual-to-expected ratio R_g = Σ_{i∈g} L_i / Σ_{i∈g} μ̂_i offers a familiar actuarial diagnostic, provided the denominator is stable and exposure periods match. A ratio near one is compatible with substantial individual error and unequal model consequences. It is a necessary conditional check for many pricing purposes, not a complete fairness evaluation.

A demographic price parity condition E[P | A=a]=c has a different meaning from equalized pure premium, proportional loss ratios, or equal claims-paying opportunity. Equal mean prices may conceal different coverage limits, deductible levels, expense loads, or underlying insured hazards. When the actuarial target itself differs across groups, parity of prices and parity of expected loss relativities cannot both be guaranteed without additional constraints or pooling arrangements. The relevant normative choice must therefore be recorded in the model's business and legal mandate.

Conflicting objectives and mathematical incompatibilities

Fairness constraints can conflict even in well-specified predictive models. If groups differ in the prevalence of an outcome, perfect group calibration and equality of all error rates can be mathematically incompatible except in special cases. The general algorithmic fairness debate, illustrated by Chouldechova's analysis of disparate impact, concerns such incompatibilities for binary outcomes. A separate actuarial translation is required before using the result to condemn or certify a continuous premium model: claim frequency, severity, aggregate calibration and individualized rating introduce other objectives and loss functions.

Insurance-focused research makes that translation explicit. The pricing fairness criteria survey by Xin and Huang (2024) evaluates distinctions among regulatory approaches and formal metrics. Lindholm and colleagues demonstrate that a model can avoid a particular definition of proxy discrimination while still producing demographic disparities and that group parity transformations can depend explicitly on protected-group information. These are structural limitations of chosen objectives, not defects that disappear by selecting a fashionable fairness score.

Fairness constraints also depend on the estimation scale. A model with equal mean residuals within each group may show unequal tail-risk error. A relative fairness measure such as (P−μ̂)/μ̂ can place exceptionally large weights on low-risk exposures. Absolute premium parity can overlook different policy limits. Cross-group comparisons with severe sample imbalance can be dominated by a handful of catastrophic claims. Good analysis specifies the loss function, unit of money, exposure normalization, grouping rule, relevant confidence interval and denominator before reporting a fairness difference.

Fairness under pooling and contractual design

Insurance prices fund obligations within an interdependent portfolio. Suppose a portfolio's indicated pure premiums are μ_i, and the insurer quotes P_i = μ_i + e_i + k_i, where e_i allocates legitimate expenses and k_i contributes to risk margins and capital. Different cost allocations can produce different observed fairness metrics while leaving underlying claim predictions unchanged. A customer may see a higher final premium because of fixed expense allocation rather than predicted loss. Disparities must therefore be decomposed between technical loss estimate, regulatory or underwriting adjustment, expense loading, margin, taxes, discounts, and the offered terms.

Fairness is also conditioned on what the institution promises. In some legally regulated or socially designed insurance lines, mandatory coverage or solidarity-based pooling requires purposeful cross-subsidy. In others, risk-related price distinctions are central to equilibrium and solvency. A purely model-internal audit cannot decide that institutional allocation question. It can show the cost and beneficiaries of alternative rules, reveal unintended consequences, and measure whether the approved rule was consistently executed. The final decision belongs to the applicable insurance regime and accountable decision makers.

Counterfactual and individual notions

Counterfactual fairness asks whether an individual's decision would change under a hypothetical intervention on a protected characteristic while appropriately holding specified causal mechanisms constant. It requires a causal model, a conception of legitimate and illegitimate pathways, and assumptions about which features would change in the counterfactual world. In insurance, age, geography, occupation, and historical claims often participate in complex pathways; a simplistic operation that changes one column while leaving all correlated attributes unchanged may produce an impossible observation. The causal question should be framed with documented structural assumptions and sensitivity tests.

Individual consistency asks whether comparable persons receive comparable treatment. The similarity metric itself may embed substantive judgments: should two drivers with equal anticipated loss but unequal income be treated alike, and should two policyholders with equal premium but different coverage limits count as comparable? Fairness claims based on similarity must define the metric and justify the protected and operational variables within it. The insurer's promise of equal treatment of comparable risks becomes meaningful only when the comparison class is specified.

Model Explainability

Explanation as a tested account of prediction

An explanation of a model can concern global behavior, local prediction, training process, or the actual administrative decision. Global explanation describes relationships averaged across a population, such as a partial dependence or accumulated-local-effects curve. Local explanation attributes one output to factors around an observation. Process explanation documents preprocessing, fitting, tuning, manual overrides, and authorization. Decision explanation connects the modeled quantity to the final quoted price, decline, or investigation. The last two become central when users seek recourse: a graph of global predictor importance cannot by itself identify the reason a particular application was rejected.

Classical additive actuarial models lend themselves to algebraic decomposition. With log link, log μ_i = β_0 + Σ β_j x_ij; a unit increase in x_j, with other covariates and interactions fixed, multiplies the modeled conditional mean by exp(β_j). This is a conditional model statement. It is not necessarily a causal account of what would happen to the insured if x_j changed, especially when covariates are correlated, selected, or measured after treatment. Interactions, transformations, caps, offsets and postmodel premium adjustments must be disclosed to interpret the actual output. A compact coefficient table is informative only when linked to the exact deployed formula and encoding.

LIME and local surrogate explanations

Ribeiro, Singh, and Guestrin (2016), in their original LIME study, approximate model behavior near one observation with a locally interpretable surrogate. The method depends on the definition of local perturbations, the neighborhood kernel, feature representation and surrogate fidelity. An actuary evaluating LIME should ask whether perturbations correspond to plausible policyholders. Independent random variations of age, tenure, geography, coverage and claims history can construct combinations absent from the portfolio. Apparent explanation stability may disappear when the reference neighborhood or random seed changes.

Local surrogate fidelity should be measured against the black-box model on a specifically described neighborhood; it should not be presented as proof of statistical calibration or fairness. A stable surrogate can faithfully reproduce an undesirable function. A poorly fitting surrogate may still tell a persuasive but false story. The audit should record the target prediction, comparison group, fidelity statistic, perturbation restrictions and reproducibility seed. Explanations for materially different underwriting actions should be rerun under stable versions and tested for sensitivity to permitted changes in applicant data.

Shapley attributions, dependence, and reference populations

Lundberg and Lee (2017) develop a unified feature-attribution approach using Shapley values to distribute a prediction's difference from a selected baseline across features. For a given game definition, contributions sum to the explained prediction difference. That accounting identity must not be confused with a causal effect of a feature, a regulatory justification, or the source of a customer's loss. A contribution depends on background data, feature grouping, and whether coalitions are interpreted through interventional replacement or conditional expectations.

The dependence problem is acute in insurance because covariates are rarely independent. Aas, Jullum, and Løland (2021) demonstrate, in their dependent-feature Shapley analysis, that independence-based feature substitution can produce unrealistic observations and misleading attributions even for simple linear predictors. An insurer should therefore distinguish the explicit computational assumption from the factual dependence in its portfolio. It should test representative alternative backgrounds, characterize correlations and report the stability of the main explanatory drivers.

The explanation baseline also has normative implications. If a customer's premium score is explained relative to the current insured portfolio, the reference incorporates past underwriting selection and portfolio demographics. An explanation relative to all applicants would use another baseline and potentially yield different signs or magnitudes. Neither baseline intrinsically defines a fair comparison. The question to be answered is whether the baseline supports the decision context: underwriting at entry, renewal, regulatory evaluation, pricing committee oversight or consumer challenge.

Counterfactual explanations and actionability

A counterfactual explanation identifies a hypothetical nearby input x' for which the model output would change to a stated threshold, such as eligibility for a lower premium. Wachter, Mittelstadt, and Russell (2018) propose counterfactual explanations for automated decisions as a practical route to understanding and potentially contesting outcomes. In insurance, a useful counterfactual requires eligibility constraints, feasible intervention costs, and stability of the quoted contract. A finding that a person would obtain a lower premium with a different date of birth or childhood postal code is an explanation of model sensitivity, not an actionable recommendation.

Meaningful recourse may instead concern correction of wrong driving history, duplicate claims, an inaccurate property characteristic, an obsolete health code, or a misclassified business use. In these cases the counterfactual change corresponds to a possible data correction and can be verified. For inaccessible or immutable attributes, the explanation may still reveal discriminatory dependence, but the appropriate remedy is institutional review or model revision rather than a request that the applicant change an impossible fact. Explanations should label mutable, conditionally mutable, and immutable factors explicitly.

Interpretable models, fidelity, and practical trade-offs

Rudin (2019), in the argument for interpretable models, emphasizes the advantages of transparent predictors in consequential settings when they can achieve competitive performance. This raises a genuine selection question for insurers: compare a credible GLM or generalized additive model with a more complex system on the same prospective holdout, including calibration, tail behavior, drift, fairness diagnostics, and documentation burden. Complexity is warranted when it contributes materially to fit for purpose. A simplistic insistence on linearity would overlook actuarially meaningful nonlinearity, and unexamined confidence in post hoc explanation would overlook unstable surrogates.

Explanation should be treated as a testable model artifact. Its technical quality can be assessed by fidelity, sensitivity to retraining and background choice, reproducibility, feature-dependence handling, domain plausibility and accessibility. Its institutional quality depends on whether the explanation addresses the question asked and connects to the real operational decision. This distinction supports strong explanations for both relatively simple and complex models without equating pleasant visuals with accountability.

Bias Testing and Validation

An audit begins with a decision inventory

A fairness audit begins by identifying all outputs that actually affect people. In insurance this may include eligibility, risk class, declined coverage, premium, renewal adjustment, claims settlement queue, suspicious-claim referral, or the probability of a manual review. Each output needs an explicit outcome definition, assessment population and legally relevant comparison. A claims severity prediction can appear neutral in a model notebook while an operational threshold routes particular groups to materially different claims services. The chain of analysis must extend through rules and overrides to the consequential decision.

The inventory should map raw sources, feature transformations, training and testing samples, labels, class membership data used only for permissible review, model version, downstream thresholds, rate tables, underwriting authority and appeal mechanisms. Every comparison depends on a denominator: all applicants, accepted policyholders, exposure-years, policies offered at a stated coverage, or claims submitted. Changing the denominator can reverse the apparent direction of a disparity, particularly when underwriting selection influences the observed claim population. A meaningful dashboard is inseparable from its cohort definition.

The CAS research program on bias collects actuarial approaches to disparate impact, proxy variables and practical model assessment. Its value lies in adaptation to insurance loss and rating structures rather than mechanical application of generic classification measures. The Society's 2026 indirect-discrimination toolbox announcement reflects the continuing technical development of scalable tests. Neither a toolbox nor a benchmark creates a universal substantive definition of fairness; the audit must record why its tests answer the relevant insurance question.

Design and sampling

The dataset should be separated into development, tuning, validation and prospective monitoring periods. Time-aware splits help identify leakage and nonstationarity. When groups differ materially in exposure duration, claim reporting lag, coverage amount or retention, diagnostics require comparable conditioning and credible standardization. For rare outcomes, bootstrap intervals or exact or suitable asymptotic methods are preferable to visual interpretation of noisy small-cell ratios. Bootstrap resampling must respect temporal dependence, policyholder clustering and catastrophe-event accumulation when relevant.

If a fairness study focuses on a protected characteristic unavailable in operational data, its audit protocol should describe the legal basis for access, the provenance or inference method, measurement uncertainty, restrictions on use and retention. Sensitivity analysis can report findings across plausible misclassification rates. The goal is to determine whether a potential adverse effect is robust to uncertainty in the group definition; purported exact group results from weak inferred demographics do not meet that requirement.

Audit design must consider selection into insurance. A premium model developed only on policyholders accepted under a previous screening rule may underrepresent declined applicants and those who did not shop due to price. Claims outcomes are only observable under the coverages actually purchased and under the settlement practices then in force. Weighting, external validation or partial identification may address some selection but require assumptions that cannot be established by a model score alone. The resulting fairness finding should name the target population and explicitly state whether it concerns observed customers or the broader applicant market.

Calibration and residual examination

Portfolio-level observed-to-expected ratios should be computed by calendar period, business segment, coverage, geography, claims-size layer and legally relevant groups where measurement is permitted. Group mean residuals E[L−μ̂|A=a] can conceal variation at low versus high predicted loss and within intersections of groups. Calibration curves and distribution-sensitive loss functions reveal more. For frequency models, compare observed counts with expected counts on exposure; for severity, examine losses conditional on report and the tails. A premium model can have equal overall loss ratios but systematically overestimate the low-risk end of one group and underestimate its high-risk end.

For a continuous prediction, define signed calibration error C_g = Σ_{i∈g}(L_i−μ̂i)/Σ{i∈g}e_i and an absolute deviation criterion B_g = Σ_{i∈g}|L_i−μ̂i|/Σ{i∈g}e_i. These serve different purposes. The first detects directional aggregate errors; the second responds to magnitude and volatility, including irreducible loss randomness. A disparity in B_g is not by itself proof of discriminatory data use: more heterogeneous loss distributions can yield greater irreducible prediction error even under a correct conditional mean. Compare against appropriate uncertainty estimates and baseline error ceilings.

In claims fraud detection, examine false-positive referrals, investigation duration and resulting burden across groups, while controlling appropriately for observed claim circumstances and review selection. In underwriting eligibility, examine decisions at thresholds, adverse-action explanations and error costs. In life or health risk scoring, pay special attention to long observation horizons, changes in medical practice and rights under the relevant legal framework. The mathematically sound metric is one that matches the actual decision and harms in scope.

Detecting proxies and residual disparity

A practical proxy study can combine correlation analysis, conditional mutual information, feature removal, subgroup calibration, counterfactual stress tests and model retraining under constrained features. Each method answers a limited question. Marginal correlation of a factor Z with a protected attribute A may be high although Z contributes little to the deployed score. A low marginal correlation can hide nonlinear interactions that predict A effectively. Conditional importance can change because other variables substitute for the removed feature. Feature deletion with no retraining may understate the model's ability to recover protected information through correlated covariates.

A responsible investigation therefore compares at least the full model, a model without suspect attributes, and a retrained model using justified alternative features. It records predictive degradation and changes in calibrated risk, adverse decision rate, error distribution, premium distribution and customer access. The legally relevant question is whether the challenged factor and resulting differential treatment can be justified under the applicable rules; statistical evidence helps answer it but does not replace legal interpretation. The New York guidance on proxy assessment explicitly integrates actuarial validity, disproportionate effects, justification and consideration of alternative practices.

Intersections, multiple testing, and uncertainty

A single protected-group comparison can miss concentrated harm to an intersection such as age by geography, disability by occupation, or sex by product subtype. As the number of groups grows, sampling uncertainty and multiplicity become significant. Exploratory subgroup scans should distinguish prespecified tests from discovered patterns, report confidence intervals or uncertainty bands, and control the interpretation of multiple comparisons. A predeclared minimum materiality threshold can help prioritize review, but it should never replace mandatory legal thresholds or an assessment of the seriousness of harm.

A good fairness audit can report a finding as uncertain, attributable to measurement, or robust across plausible assumptions. It documents mechanisms to investigate, model alternatives and the consequences of intervention. The strongest evidence joins prospective outcomes, reproducible computation and a recorded substantive rationale. It does not rely on a solitary parity percentage or a declaration of compliance in a vendor slide deck.

Decision Review and Contestability

From a modeled estimate to a challengeable decision

A person encounters insurance through an offer, rate change, refusal, policy term or claims response, not through an estimator's latent parameters. To make that decision contestable, the record must identify its effective date, the operative policy or application, the model or rule version, the material inputs, the applied overrides, the reasons for departure from an ordinary comparator and the route for correcting errors. A consumer may reasonably ask whether a data source is accurate, whether a factor may lawfully be considered, whether an excluded factor was used indirectly, or whether the explanation given matches the actual decision. Each question has a different evidentiary answer.

An adverse underwriting decision based on an erroneous motor-vehicle record is a data-correction problem. A disparity produced by an approved but legally impermissible rating factor is a rule-governance problem. A claim adjustment assigned to a high-risk investigation queue on unstable or biased labels is a model-performance and procedural problem. A premium that is accurate as a prediction of loss but breaches a binding prohibition is a legal-compliance problem. A review mechanism should classify and route these questions instead of offering only generic algorithmic descriptions.

Practical explanation for a policyholder

A decision notice can state the product and coverage, the decision made, the principal reason categories, the significant individual facts and their sources, how corrections are requested, the authority responsible for review, and an expected procedural next step. The level of technical detail depends on the applicant's needs and applicable disclosure law. Where information cannot be distributed openly because it reveals other policyholders' data or trade secrets, the institution can still maintain complete regulator-accessible and reviewer-accessible records and provide an accurate decision-specific summary. Commercial confidentiality should not erase the ability to identify material grounds for an adverse determination.

The New York DFS circular letter provides a concrete regulatory example: for specified AIS/ECDIS underwriting situations, it sets expectations concerning notices, access to data used in decisions, correction processes, recordkeeping and vendor accountability. These expectations derive from New York's insurance framework and cannot be transferred unchanged to a different jurisdiction. The general institutional lesson is that explanations should connect to verifiable underlying facts and a real review pathway.

Counterfactual review and human intervention

A reviewer must be empowered to do more than repeat a model score. The operational review may verify input records, identify out-of-scope use, test whether an output depended on a suspect proxy, apply an authorized exception, request new evidence, reassess coverage terms and document the disposition. Where automated profiling triggers legal protections, individuals may have specific procedural rights. The EU General Data Protection Regulation, particularly Article 22 and associated provisions, establishes a conditional regime for certain solely automated decisions with legal or similarly significant effects, including relevant exceptions and safeguards. It does not supply a blanket identical explanation procedure for every insurance score or every insurer worldwide.

An effective human review has defined authority, training, information access, independence appropriate to its scope, and an auditable record of how the conclusion was reached. If an operator lacks the right to depart from the algorithm or the evidence needed to examine it, the review risks becoming ceremonial. Conversely, arbitrary manual overrides can themselves generate unequal treatment. Override patterns therefore belong in periodic model and fairness monitoring, with reason codes and differential outcomes available for independent analysis.

Institutional remedies

Possible remedies include correcting inaccurate data, rerunning the same approved model with corrected inputs, removing an unjustified rule, recalibrating an affected segment, compensating policyholders as applicable, revising a communication, restricting a vendor feed, and retraining or withdrawing a model. The choice depends on the failure. A local model explanation that identifies the influential feature does not automatically determine the appropriate remedy. The insurer should retain sufficient transaction records to show whether remediation changed the actual decision and whether similar cases were addressed consistently.

Review and challenge processes also create information for scientific evaluation. Error-correction rates, overturned decisions, types of data defects, complaint patterns and time to resolution can reveal operational failure modes that conventional model metrics miss. Their interpretation should account for unequal awareness of rights and access to complaints mechanisms. A low complaint rate is compatible with satisfaction, limited awareness, and prohibitive contest costs; it is not independent proof of fairness.

Regulatory and Ethical Comparisons

The United States: state-specific obligations and supervisory models

The United States operates through state insurance regulation with line-specific and jurisdiction-specific rules. The National Association of Insurance Commissioners adopted the AI Systems Model Bulletin in December 2023 as a supervisory model addressing insurers' development, acquisition and use of AI systems. Model bulletins and association materials must be distinguished from legally operative requirements adopted or issued by a particular state. Their common emphasis is insurer responsibility, appropriate governance, documentation, risk mitigation and adherence to existing insurance law.

New York's July 11, 2024 Circular Letter No. 7 extends its articulated expectations to AIS and ECDIS used in underwriting and pricing, including systems that do not rely on external consumer data in the narrow sense. It addresses generally accepted actuarial standards, correlation with protected characteristics, disproportionate adverse effects, lawful rationales, alternatives, vendor assurance, testing, record retention and customer disclosures. This is a significant example of a regulator translating fairness from a general ethical concern into a documented sequence of technical and administrative assessments.

The scientific lesson is that a model's lack of direct access to a protected field proves only an input restriction. The regulatory examination can ask whether outputs have adverse effects, whether correlated variables operate as substitutes, whether the intended purpose is legitimate, and what alternatives preserve legitimate underwriting information. When an insurer uses a third-party model, the procurement contract and secrecy of the algorithm do not eliminate the insurer's obligation to account for the decision within its legal regime.

The European Union: data protection, insurance equality, and AI rules

The Gender Equality in Goods and Services Directive provides a distinct insurance-specific rule in Article 5 governing consideration of sex in premiums and benefits for new contracts, subject to the legal history of its former exception. It illustrates that a protected characteristic can be statistically predictive and nevertheless legally excluded from individual premium differentiation. The justification is grounded in the legal norm, not in a claim that the feature lacks statistical information. A comparative insurance model must therefore encode applicable restrictions as constraints on decision rules, even where its loss-cost estimate retains predictive relationships requiring audit.

The GDPR regulates personal-data processing and establishes information, access and other rights subject to its provisions. Relevant Article 22 safeguards depend on whether an individual is subject to a decision based solely on automated processing that produces legal or similarly significant effects, and on the applicable exception and safeguards. A model used as a recommendation inside meaningful human judgment raises different factual questions from a fully automated denial. The classification of the real processing chain is decisive.

The consolidated EU Artificial Intelligence Act expressly identifies certain AI systems for risk assessment and pricing in relation to natural persons in life and health insurance as Annex III high-risk applications. That classification does not automatically extend to every actuarial calculation, insurance product or internal statistical model. According to the consolidated text of July 27, 2026, the main regulation generally applies from August 2, 2026, while specified Chapter III provisions for Annex III high-risk systems are scheduled to apply from December 2, 2027, and those associated with Annex I from August 2, 2028. Those future dates are material for accurate November 2026 analysis; one cannot describe all of the later high-risk obligations as already effective.

The European Insurance and Occupational Pensions Authority's 2025 AI governance Opinion explains supervisory expectations under insurance-sector legislation in a risk-based and proportionate manner. It concerns the interaction of existing insurance governance with AI use and must be read in light of its stated scope, including the boundary with AI Act categories. Across these instruments, the EU approach combines sector-specific equal treatment, general data protection, AI classification, and supervisory expectations; the precise obligations still depend on the system, decision and applicable legal provision.

The United Kingdom: outcomes and fair value

The UK Financial Conduct Authority's Consumer Duty materials approach fairness partly through whether the product and its distribution deliver fair value to defined customer groups. The FCA's Price and Value Outcome guidance, updated July 10, 2026, expressly recognizes that differential pricing is not categorically prohibited, while emphasizing justified differences and fair value for each group. It also warns against relying on inertia or behavioral biases to maintain pricing differences. A model can therefore pass a claims-cost calibration test while raising issues about the value customers receive relative to premiums, coverage and fees.

This example shows why insurance explainability must identify the object being explained. A pure-premium risk model addresses expected claims. A pricing or renewal engine may incorporate demand elasticity, loyalty behavior or distribution channel. The fairness investigation of the final consumer price requires tracing these components separately. Presenting risk factors as the complete cause of an outcome would misdescribe a system in which commercial overlays materially affected the decision.

Canada: enterprise model risk governance

Canada's Office of the Superintendent of Financial Institutions published Guideline E-23 on Model Risk Management on September 11, 2025 with an effective date of May 1, 2027. Its published scope includes federally regulated insurers and describes risk-based model inventories, independent review, lifecycle controls, documentation, explainability and monitoring. As of November 14, 2026, the revised 2027 guideline remains prospective, while its content provides a well-specified reference for governance design.

The Canadian example connects fairness to enterprise model control: the institution must be able to identify which models exist, who owns them, how they were approved, what limitations apply, and what triggers validation or retirement. A technically impressive fairness study cannot be operationally decisive if the reviewed model is not the version used in live decisions. Conversely, comprehensive version and ownership records make retrospective analysis and correction possible.

Comparative synthesis

The legal regimes differ in their operative definitions and sanctions. The United States example emphasizes state-specific insurance antidiscrimination and supervisory governance; EU law combines binding protected-characteristic rules, data protection, and staged AI regulation; the UK fair-value approach highlights consumer outcomes and pricing design; Canada's revised guidance details lifecycle risk control. Their shared practical demand is a traceable relationship among data, model, authorized decision and review. Their differences prevent a single international fairness score from functioning as a legal clearance certificate.

A transnational insurer should accordingly maintain a jurisdiction-by-product decision matrix identifying allowable inputs, restricted outputs, required validations, disclosures, complaints rights, effective dates and supervisor expectations. A model may be developed centrally while its decision rules and monitoring differ by country. Such localization is a factual condition of lawful use and a substantive research dimension, not merely a documentation exercise.

Numerical Illustrations and Methodological Tests

Illustration: risk-based pricing and group parity

Consider a hypothetical one-year portfolio with two groups of 500 exposures each, identical limits and contract design. Assume group A's expected annual claim cost is $200 per policy and group B's is $300. Assume all expenses and margins are represented by a constant multiplicative factor of 1.25 applied to pure premium, and that there are no behavioral responses, taxes, policy limits differences or distributional changes. Under fully differentiated technical pricing, the quoted prices are $250 and $375. The mean across the equally sized portfolio is $312.50. The expected total claims are $250,000 and the aggregate modeled premium is $312,500.

If a parity constraint instead requires one common quoted premium for both groups at the same aggregate target, each pays $312.50. Group A then pays $62.50 more per exposure than its differentiated premium, while group B pays $62.50 less. Across 500 exposures, the transfers are +$31,250 and −$31,250 relative to differentiated prices. The expected total premium remains $312,500. These figures follow immediately from the stated assumptions; they represent no real-world pricing recommendation or evidence of the merits of parity. Their significance is conceptual: demographic price equality, claims-cost risk differentiation and an aggregate premium requirement define different allocations of the same illustrative expected portfolio resources.

This example cannot determine whether the original group distinction was permissible, whether differences reflect prohibited proxy use, or whether pooling improves overall access. Those questions require the causal and legal content omitted by design. It also assumes a fixed portfolio; with adverse selection, changing coverage purchases or exit by lower-risk consumers, the common-price arithmetic no longer fixes the future composition. The example therefore demonstrates the precise limits of a deterministic fairness metric as well as the arithmetic of cross-subsidy.

Illustration: an uninformative proxy generates a disparity

Consider a second artificial population with a binary protected grouping A and a binary observed feature Z. Let half the population belong to each group. Let P(Z=1|A=1)=0.8 and P(Z=1|A=0)=0.2. Suppose actual prospective expected claim cost is exactly $200 for every policyholder, independent of both group and Z. Imagine that an incorrectly fitted or historically contaminated model estimates the score s=200+80(Z−0.5), assigning $240 to Z=1 and $160 to Z=0. Its average score is $224 for group A and $176 for group B: the group difference is $48, despite identical stipulated true expected losses of $200.

The calculation is direct: E[s|A=1]=0.8×240+0.2×160=$224, while E[s|A=0]=0.2×240+0.8×160=$176. The model produces a predictive group difference entirely from the feature's unequal prevalence and its unsupported estimated coefficient. Its global mean score is $200, matching the portfolio's true mean, but this aggregate equality masks group-specific overprediction of $24 and underprediction of $24. Thus portfolio calibration at one broad aggregation level cannot establish group calibration; explainability showing that Z raises the score would identify the mathematical channel but would not itself supply actuarial justification for Z.

A sound response is to inspect how Z acquired an estimated association with loss, verify exposure and claim labels, assess selection and leakage, and retrain or revise the model if the coefficient lacks legitimate support. Removing A from the feature set changes nothing in this invented example because A was never directly supplied. The example isolates the distinction between prohibited direct use, proxy association and demonstrable predictive validity. It also illustrates that a measured disparity is an investigative finding whose mechanism must be established.

Illustration: calibration and unequal error

Take two groups with ten illustrative observations each. Every observation has a model prediction of $100 and each group's total realized loss is $1,000. In group A, every observation produces exactly $100 of loss; in group B, nine have zero and one has a $1,000 loss. The predicted total and realized total agree for both groups, so the aggregate actual-to-expected ratio equals one in each. Yet the mean absolute prediction error for group A is $0 and for group B is (9×$100+$900)/10=$180. The equality of aggregate calibration coexists with different error magnitude because the distributions have different variances. Since these are realized observations rather than samples from established true conditional means, the example cannot prove bias; it establishes that equality of one summary statistic leaves another property unrestricted.

For insurance, this matters because a model can have stable aggregate calibration and still be unreliable for very different operational purposes. Tail error is important for capital, reserve uncertainty and catastrophic loss; false-positive referrals matter for investigative burdens; premium level error matters for policyholder affordability; and decision thresholds can magnify small score errors. Evaluation must match the use rather than treat one statistic as a universal certificate.

Illustration: explanation versus permissible intervention

Suppose a policyholder's model score is $420, with a selected reference score of $300, and a local explanation assigns +$70 to geographic coding, +$40 to claims history and +$10 to an interaction allocation. The arithmetic sums to $420 relative to the baseline, but the decomposition's meaning depends on its specified attribution algorithm. If the geographic code is erroneous, its correction is actionable. If the code is accurate but acts as an illicit proxy in the applicable regime, a policyholder's moving home cannot be treated as the proper remedy. If geography captures genuine catastrophe exposure and is permitted, then its relevance must still be documented. An attribution answers what the selected explanation algorithm allocates; an administrative review decides which facts may be used and how an error is repaired.

Reproducibility checks

All monetary examples can be independently reproduced with elementary arithmetic. The first: 500×200 + 500×300 = 250,000; 1.25×250,000 = 312,500; 312,500÷1,000 = 312.50; 500×(312.50−250)=31,250. The second: 200 + 80×(1−0.5)=240 and 200+80×(0−0.5)=160; weighted group values 224 and 176; aggregate mean 200. The third: absolute group-B deviations consist of nine errors of 100 and one of 900, sum 1,800, mean 180. These are deductive demonstrations conditional on explicitly invented populations, not reported observed effect sizes or estimated confidence intervals.

Model Governance Across the Actuarial Life Cycle

Specification and institutional ownership

Model governance should begin with a documented decision purpose, target population, permissible application, approval authority and expected materiality. The development specification should identify response definitions, exposure measures, available-time constraints, risk and commercial objectives, restricted variables, data-source licenses, and anticipated model users. The same artifact should state whether the model is an estimator, a rule used automatically, an advisory score, or one component of a multi-stage decision system. These categories determine what can be explained and what must be tested.

The Actuarial Standards Board's ASOP No. 56 on Modeling supplies a professional reference for modeling practice and communication in its jurisdictional context. Its relevance is methodological: the responsible practitioner considers model purpose, complexity, data, assumptions, limitations, validation, and communications. The comparable ASOP No. 23 data-quality standard frames the reliability of model inputs. Together they establish a useful baseline of accountable modeling without offering a universal fairness algorithm.

A model register can record version hash, feature dictionary, dataset lineage, approved applications, independently reviewed limitations, performance targets, fairness tests, responsible decision owner, last calibration and next review trigger. Strong governance also documents model dependencies. A seemingly small change to a geocoding provider or claims classification system can change the deployed feature distribution and the fairness implications of an otherwise untouched model formula.

Independent challenge and validation

An independent validation function should assess design appropriateness, conceptual soundness, data validity, cross-validation and prospective holdout behavior, subgroup uncertainty, model comparison, explainability fidelity, and stress performance. Independence here means sufficient capacity and authority to challenge developers' assumptions and release decisions. Reviewers may be actuarial, statistical, legal, compliance or technical specialists depending on the issue, with clear responsibility for integrating findings. A fairness or explainability approval is meaningful only for the specified model version and intended use.

A strong validation record includes negative findings and remedial evidence. An important adverse disparity should trigger a documented investigation into data, causal mechanisms, alternative specifications, permissible treatment and downstream decision effect. Reviewers may conclude that a finding is statistically unstable, explained by a documented legitimate risk process, remedied by data correction, or unresolved. Each disposition needs a reproducible basis. Unresolved substantive legal or scientific questions belong in a review hold for the affected model or use case, rather than being silently relabeled as accepted risk.

Vendors and outsourced prediction

Third-party systems may expose only a score or a decision API. Contractual access should specify model-purpose disclosures, input provenance, validation materials, update notification, test access, audit rights and incident procedures appropriate to applicable law and institutional risk. When full source access is impracticable, challenge tests can examine model behavior through controlled input variation and outcome analysis. Such testing is useful but cannot reproduce unavailable training data or prove internal mechanisms. The resulting evidentiary limitation should be made explicit in the regulator-facing review record.

The New York insurer-vendor accountability requirements show the weakness of relying solely on supplier assurances in consequential insurance decisions. The governing entity must understand what it authorizes and retain enough information to assess risk and respond to challenge. A proprietary model used as an outsourced rating factor remains part of the insurer's decision system for those purposes.

Monitoring and change control

Monitoring should cover data drift, label drift, missingness, subgroup calibration, aggregate adequacy, relative premium changes, complaints, correction rates, override trends and unexpected distribution shifts. Relevant baselines include both the training population and previous deployed decisions. Drift should be examined at the feature, score and outcome levels. Because insurance claims develop over time, early signals may rely on exposure and initial-report indicators, with later reassessment using mature claims; the timing of final outcome observation must be documented.

A retraining event should not silently reset fairness review. Substituting variables, changing target definition, adjusting threshold or modifying downstream premium overlays can change treatment even if aggregate prediction accuracy improves. Governance must require a change description, reason, comparative validation and explicit authorization. The Canadian E-23 model lifecycle guidance details many of these controls, though its revised requirements take effect on May 1, 2027, rather than already applying as of this dossier's review date.

An insurer can organize controls according to the NIST AI Risk Management Framework as a voluntary cross-sector governance reference, while still relying on binding sector- and jurisdiction-specific rules for actual compliance. Governance frameworks help allocate ownership and evidence; they do not resolve the substantive question of whether a particular rating factor is allowed in a particular country and line.

Production explanations and document integrity

The operational record of a model decision should preserve a reproducible point-in-time reconstruction: input data or legally maintained references to them, feature transformations, executable model/version identity, rating tables, overrides, thresholds and final customer communication. Storage and access need privacy and security protections, including lawful retention and deletion. The scientific reason for reconstruction is direct: without an accurate record of what was used, a local explanation generated months later against an updated algorithm may be an explanation of a different decision.

At the time of release of an insurance model, the review packet should include a plain-language description of the tested scope, statistical performance, fairness diagnostics with uncertainty, explanation methods, known constraints, jurisdiction-by-product applicability, reviewer names and conditions for revalidation. That packet is an institutional control for the insurer. This research monograph supplies general methodology and a comparative evidence synthesis; it is not the validation report of any specific operational insurance system.

Integrated Research Synthesis

Explanation, fairness, and justification form different evidentiary questions

The research establishes three questions that must be answered sequentially during an actuarial model review. The descriptive question asks what the deployed model and downstream decision function did. The statistical question asks whether predictions and treatment satisfy specified, empirically tested performance or fairness criteria for a defined population. The justificatory question asks whether the inputs, purpose, distributional consequences and remedies conform to the governing insurance arrangement and law. The questions are connected by evidence, but no single question can substitute for another. A SHAP plot does not legalize a tariff factor; a balanced selection rate does not establish accurate risk prediction; and a legal rule does not eliminate the need to test data quality.

This is an analytical interpretation of established research, not a newly instituted theoretical system. Its practical significance is that each question has appropriate methods, experts, and records. Explanation fidelity belongs to technical validation. Expected loss calibration and residual uncertainty belong to actuarial review. Protected-class rules and proportionality belong to relevant legal and supervisory interpretation. Actual correction and remedies belong to the insurer's decision process and applicable complaints institutions. The roles may overlap, but the evidentiary standards should remain distinguishable.

What can be concluded from a group gap

A group difference in average premium can result from true differences in the legally relevant risk distribution, prohibited or unjustified proxies, expense allocations, commercial demand targeting, historical dataset selection, unequal coverage terms, errors in group measurement, or a combination of these. The observed gap proves that a measured disparity exists under the specified comparison; it does not by itself identify which explanation is correct. A proper analysis decomposes technical prediction, decision overlays and population composition while preserving uncertainty. It then tests alternative permissible models and records the effect on solvency-relevant aggregate adequacy and customer outcomes.

Conversely, absence of a detected aggregate gap is insufficient to exclude disparate treatment. Subgroup overcharges and undercharges can offset in the mean; high-risk and low-risk portions of the same group can be affected differently; data-inference errors can hide an effect; and denied applicants can be excluded from observed-policyholder comparisons. The correct audit response is therefore neither automatic acceptance nor automatic condemnation, but disciplined investigation of what has been measured, whom it represents, and what decision it affects.

Accuracy and distributive design

Predictive improvement can make underwriting more efficient, reduce some cross-subsidies and create opportunities for lower premiums among particular groups. It can also fragment risk pools, increase the cost of covering high-risk people, and alter insurance availability. Whether such effects are desirable depends on product purpose and the governing allocation framework. A fair-value regime, compulsory solidarity arrangement and individually risk-rated optional insurance market can choose different permissible distributions. The actuary's distinctive contribution is to calculate and communicate how alternatives affect expected claims, premium adequacy, uncertainty, cross-subsidy and access, while respecting the operative legal and contractual rules.

The literature on antidiscrimination pricing models and proxy-versus-demographic fairness shows why a fairness constraint must be selected on conceptual grounds before it is optimized computationally. If the selected metric is the wrong description of the relevant injury, improving its score can mislead decision makers. The evaluation should therefore report the institution's definition of the fairness objective and the consequences for predictive error, premium distribution, and access.

Contestability is a test of institutional meaning

A model is meaningfully explainable to an affected person when the reason can be connected to the person's actual decision, checked against the underlying record, and addressed through a defined route of correction or challenge. This proposition follows from the difference between a model function and an institutionally authorized action. It can be assessed by sample case reconstruction: can the institution identify the historical inputs, reproduce the output, explain an override, locate the proper rule, and correct a documented error? A technically sophisticated feature-attribution method may help this work, but the ability to reproduce and remedy a decision is the decisive operational test.

The cross-jurisdiction comparison supports a wider conclusion: institutional accountability depends on preserving an auditable link between predictive evidence and the authorized decision. New York specifies consumer disclosures and vendor responsibility; EU law establishes category-specific protections and staged AI duties; the FCA examines price and value outcomes; OSFI develops enterprise model control. These are not identical regulatory programs, yet each demonstrates why the operational record is part of the fairness question.

Research Method, Evidence Limits, and Reproducibility

This dossier uses analytical and comparative literature review, primary-source legal reading, statistical reasoning and explicitly hypothetical mathematical examples. Its insurance-specific core is drawn from actuarial fairness scholarship, model explanation research, professional actuarial standards and regulator materials. The legal review uses published original instruments or official supervisor descriptions, with an as-of date of November 14, 2026. The method preserves differences among journal results, professional standards, model bulletins, supervisory guidance and binding legislation. It does not derive jurisdiction-wide legal claims from a single regulator's website or treat a research group's experimental fairness definition as a universal norm.

The principal explanatory-method references are the original LIME publication, the SHAP framework, the dependent-feature correction research, and the counterfactual explanation analysis. The core actuarial fairness comparisons use Lindholm and colleagues (2024) and Xin and Huang (2024). Statistical claims in this dossier are limited to consequences of the expressly given formulas and the qualitative or formal results attributed to those works; no original field measurements or proprietary actuarial observations are asserted.

The legal review deliberately preserves effective dates. The July 27, 2026 consolidation of the EU AI Act is used for its prospective high-risk provisions, and OSFI's E-23 revision is labeled with its May 2027 effective date. The FCA's July 2026 update is identified as the current version reviewed. Model guidance from the NAIC is differentiated from adoption by individual states. These distinctions guard against converting a future regulatory requirement into an alleged current obligation or a nonbinding model into binding law.

Primary limitations arise from the literature-review design: there is no insurer data to test disparate impact in a real portfolio, no third-party vendor code, no individual claimant case record, no independent external peer review, and no prospective observation of how hypothetical fairness interventions alter selection or solvency. Consequently, the dossier's numerical demonstrations are deductive illustrations, not empirical estimates or normative proof of one optimal price schedule. The mathematical definitions are intentionally simplified to isolate relationships; heavy-tailed loss, exposure heterogeneity, reserve development and competing institutional duties require context-specific extensions in live actuarial work.

Each externally linked research source is described in the accompanying private EVIDENCE.json ledger, including its bibliographic identity, source location, checked claim and verification scope. The reader-facing bibliography preserves full names and titles, while internal evidence identifiers remain outside the scientific narrative. Contextual hyperlinked phrases are deliberately short and point to the works or official documents supporting their surrounding statements. The final scientific interpretation is an independent synthesis of those materials and claims no new theory, causal discovery, tested intervention efficacy, or scientific priority.

Conclusion

Explainability in actuarial models is the capacity to reconstruct, describe and critically test how model inputs, assumptions and transformations affect a specified prediction and its actual insurance use. Fairness is an explicitly chosen property of treatment, error distribution, predictive calibration, access, affordability or review, assessed for a defined population under a governing institutional and legal rule. A valid actuarial analysis connects the two through data provenance, prospective validation, appropriate group comparisons, examination of proxy channels, documented decision authority and a workable method of correction.

The reviewed literature shows that direct exclusion of a protected variable leaves possible proxy effects; that group disparity and proxy discrimination are mathematically distinct; that multiple desirable fairness criteria can conflict; and that feature-attribution methods depend on reference data, feature dependence and local approximation. The constructed examples demonstrate why portfolio-level calibration may coexist with subgroup overprediction, why error distribution differs from aggregate mean accuracy, and why equality of quoted premiums changes the allocation of resources under explicit pooling assumptions. None of these results supplies an automatic legal or ethical verdict. Their value is to identify the question that must be answered and the evidence capable of answering it.

Comparative insurance regulation adds enforceable and supervisory substance: permitted use depends on product and jurisdiction; insurer responsibility extends through vendor systems and downstream decision rules; effective dates must be observed; and review must permit reconstruction of an actual case. The most defensible institutional practice is therefore a complete, versioned evidentiary chain connecting legitimate actuarial purpose, reliable data, model validation, selected fairness criteria, authorized pricing or underwriting use, understandable decision reasons and meaningful correction. This conclusion is a grounded analytical synthesis of the literature and regulatory record reviewed in this monograph.

References