Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Driver and Vehicle Risk Classification Fairness

Abstract. Driver and vehicle risk classification is the allocation of motor-insurance applicants and exposures to risk categories used in underwriting, pricing, coverage design, and portfolio management. Its fairness depends on several analytically distinct properties: measurement validity, the relationship of rating factors to covered loss, the distribution of its consequences, applicable restrictions on classification, the reliability and governance of data, and access to essential insurance protection. This monograph examines these properties across actuarial science, insurance economics, empirical discrimination research, privacy, and comparative law. It differentiates accident risk from insured claims; claim frequency from severity; technical premium from charged price; predictive association from a legally permissible or ethically defensible classification; and group-level disparity from a causal finding about a particular rating factor. Original analytical synthesis connects these distinctions to age, driving experience, mileage, driving record, geography, vehicle characteristics, credit-based scores, and usage-based telematics. The discussion compares United States state-based regulation, including California and Colorado, with European Union equality and data-protection requirements and United Kingdom insurance and consumer rules. Empirical anchors include the Federal Trade Commission's automobile-insurance study, peer-reviewed actuarial fairness research, a study of vehicle telematics, and road-safety evidence from WHO and the International Transport Forum. A reproducible audit specification identifies appropriate outcomes, denominators, counterfactual comparisons, calibration checks, distributional measures, and governance documentation. The central finding is that justifiable motor-insurance classification is an evidence- and jurisdiction-dependent institutional decision in which forecasting quality, lawful differentiation, causal interpretation, accessibility, privacy, and procedural accountability must be assessed together. This work contributes original cross-source interpretation and review architecture. It claims no new actuarial theory, fairness doctrine, or independently estimated population effect.

Keywords: automobile insurance; driver risk; vehicle risk; insurance rating; risk classification; actuarial fairness; indirect discrimination; proxy variables; credit-based insurance scores; telematics; usage-based insurance; statistical discrimination; motor liability; calibration; regulatory governance; GDPR; insurance law; consumer protection; predictive modeling.

Introduction

Automobile insurance converts heterogeneous road-use exposures into contractual protection and a price. The social importance of this transaction arises from the combination of widespread vehicle dependence, mandatory motor liability in many jurisdictions, expensive bodily injury claims, and highly differentiated driving and vehicle circumstances. Classification is therefore both a technical element of loss estimation and an institutional rule determining who can obtain protection, on what terms, and at what expense. The NAIC account of automobile rating identifies driving records, location, age, experience, vehicle type, vehicle use, mileage, credit history, and contract choices among factors encountered in the U.S. market. Their appearance in an industry overview does not establish their acceptability in every state or country.

Road collisions generate human, physical, and economic consequences on a scale larger than the insured claims of any one portfolio. The World Health Organization's 2023 road safety report estimates approximately 1.19 million road traffic deaths annually. This estimate covers a population health outcome across road users, not motor-insurance claim frequency or insured severity. The International Transport Forum's Road Safety Annual Report 2025 presents comparative road safety performance among participating IRTAD countries and documents variations in availability and vintage of country-level data. These sources provide environmental context and important warnings about denominator selection; their published counts cannot substitute for the loss, exposure, coverage, and claims records needed to fit a rating plan.

This study addresses three questions. First, how should predictive driver and vehicle risk classes be constructed, interpreted, and validated? Second, when can an otherwise predictive class generate direct or indirect discrimination, and what evidence can distinguish proxy effects from legitimate risk differences? Third, how do privacy, consumer protection, antidiscrimination law, and insurance supervision alter permissible classification across jurisdictions? The inquiry covers personal motor insurance and relevant methodological extensions to fleets and emerging vehicle technologies. The unit of analysis is stated explicitly wherever it affects inference: person, driver, insured vehicle, policy, trip, vehicle-mile, exposure-year, reported claim, incurred claim, or ultimate loss.

The research method is a critical synthesis of verified institutional publications, legal texts, supervisory materials, and selected original scientific studies. It uses standard frequency–severity modeling, causal reasoning, comparative statutory interpretation, and reproducible audit specifications. Primary legal sources are read as jurisdiction- and date-specific rules. Academic results are attributed to the samples and design actually described by their authors. Mathematical illustrations below are analytical identities or transparently hypothetical examples, not empirical findings from an insurer. No policyholder microdata were obtained, no original statistical coefficients were estimated, and no external peer review is represented as having occurred.

A fundamental discipline governs all subsequent interpretation: road crashes, damage, legal liability, covered occurrence, claim notification, reserve development, and final insurance payment refer to different events. A predictive variable may affect more than one stage. A model trained on paid claims measures a function of insured damage, contractual coverage, deductible, claim reporting, investigation, and settlement practices. An observed association with paid loss is consequently a narrower proposition than a causal claim about unsafe driving.

Conceptual, Historical, and Primary-Source Foundations

The insurance transaction and its units

A rating class partitions insured exposures for prospective assessment. A driver risk factor describes a characteristic or behavior attributed to a driver, such as experience, previous violations, or usage intensity. A vehicle risk factor describes the insured object or its use, including repair cost, safety systems, theft susceptibility, vehicle performance, annual distance, or commercial service. A geographic factor may represent accident density, theft, environmental hazard, repair markets, claim litigation, traffic mix, or socioeconomic conditions. The same variable may combine different mechanisms across coverages.

Third-party bodily injury liability indemnifies legally recognized obligations subject to policy wording and applicable law. Third-party property liability involves another party's property loss. Collision and comprehensive coverages commonly address damage to the insured vehicle under different insured-event definitions. The same driver–vehicle pair can therefore have different estimated relativities for bodily injury liability, physical damage, and theft. A unitary unexplained 'risk score' may conceal these distinct targets.

Let N denote claim count over a specified covered exposure E, and let Y_j denote the insured amount for claim j on a consistent development basis. Define frequency f(x) = E[N | X=x]/E and severity s(x) = E[sum(Y_j) | X=x]/E[N | X=x], where the latter is meaningful only for a positive expected count. The technical expected loss per exposure is then approximately f(x)s(x), with adjustments for loss development, trends, and coverages. The actual premium further includes expenses, capital costs, reinsurance, taxes, profit provisions, regulatory constraints, and commercial decisions. This decomposition is an accounting and actuarial framework; it makes no statement that any one protected or correlated characteristic must enter the rating model.

The distinction between exposure and experience matters. A driver with more reported accidents over ten years may have a lower annualized claim rate than a driver with fewer accidents over one short year. A high-mileage driver may have more aggregate opportunities for collision while also being more experienced, driving on different road types, or traveling fewer miles at high-risk hours. A vehicle used for ride-hailing may carry an entirely different usage exposure from the same make and model used only on weekends. Failing to align the model denominator with the covered exposure creates apparent risk differentials that are artifacts of measurement.

Origins of modern classification and the continuing regulatory compromise

Insurance classification developed to replace undifferentiated premiums with group estimates of heterogeneous expected losses. The historical justification is pooling: risk is transferred collectively while premium relativity reflects information about expected insured costs. Risk classification also introduces distributive consequences. More refined differentiation narrows some cross-subsidies while concentrating costs on particular individuals and groups. The economics of classification is inseparable from the legal boundary between permitted actuarial differentiation and socially unacceptable exclusion.

The California statutory rating hierarchy provides a prominent example of public intervention in that boundary. Insurance Code § 1861.02, introduced through Proposition 103, orders three primary factors for automobile insurance in decreasing importance: driving safety record, annual mileage, and years of driving experience, followed by other commissioner-authorized factors substantially related to loss risk. The accompanying California rating-factor regulation identifies additional optional factors, including vehicle type, use, and various vehicle characteristics. The rule demonstrates that the legal priority of factors can differ from a data scientist's unconstrained ordering by incremental predictive fit. California is one U.S. jurisdiction; its hierarchy is not a universal American rule.

A different method appears in the European Union's Equal Treatment Directive and the Court of Justice's Test-Achats judgment. Directive 2004/113/EC, Article 5, framed equal premiums and benefits irrespective of sex; the Court held that the derogation allowing indefinite sex-based variation was invalid from December 21, 2012. The result illustrates a classification restriction that remains operative even where relevant actuarial data could show group differences. Whether another variable acts as an impermissible proxy requires separate legal and factual analysis; the ruling does not ban all motor-insurance differentiation.

Risk characteristics and normative status

A useful descriptive partition separates four kinds of factors. Direct behavior measures include recent telematics observations of speed and braking, with careful attention to context and measurement. Historical experience measures include prior claims, violations, and license tenure. Physical and exposure measures include mileage, vehicle equipment, repairability, use, and road environment. Indirect socioeconomic indicators include some uses of credit scores, education, occupation, and granular territory. A variable may straddle categories: location is both a road-condition indicator and a potential demographic proxy; credit information may predict a claim outcome without measuring a vehicle's mechanical hazard.

The NAIC credit-score topic paper explains the distinction between credit-based insurance scores and ordinary credit scores, as well as differences among states in their permitted use. It reports market use conditional on state legality, while assigning no universal normative approval. Historical regulation of motor rating thus has no single linear direction: jurisdictions select differing compromises between cost correlation, interpretability, protected characteristics, access, and competition.

The insured event is not the complete road-safety event

An insurer's frequency–severity data are affected by deductibles, claim thresholds, reporting propensity, fraud control, litigation, liability apportionment, repair inflation, bodily injury treatment patterns, and claims-handling behavior. These effects can vary with income, vehicle age, urbanicity, and access to repair facilities. Claim frequency may decrease when deductibles increase without a corresponding reduction in collisions. Recorded severity may rise when repairs become more expensive despite improvements in crash prevention. The model's dependent variable must therefore be named: all accidents, notified incidents, paid losses, or ultimate losses, each under specified contracts and censoring windows.

Comparative national crash evidence should not be merged directly with an insurer's policy file. WHO mortality rates are generally person- or population-oriented; IRTAD indicators may use inhabitants, registered vehicles, vehicle-kilometers, and country-specific reporting practices. Rating may use earned vehicle-years or distance traveled by covered drivers. Alignment requires an explicit mapping between events, populations, periods, currencies, coverage, and claims development. These requirements underlie the validation and fairness tests developed later.

Predictive Risk Classification

Frequency, severity, and actuarial structure

The classical prospective rating exercise estimates a conditional expected covered loss given the information available at the pricing date. For insured unit i, with earned exposure e_i, claim count N_i, and covariate vector x_i, a Poisson generalized linear model can specify

Ni∣xi∼Poisson⁡(eiλi),log⁡λi=β0+xi⊤β.N_i \mid x_i \sim \operatorname{Poisson}(e_i \lambda_i), \qquad \log \lambda_i = \beta_0 + x_i^\top \beta .

The exposure offset log(e_i) prevents low- and high-exposure contracts from being compared as if they represented the same time at risk. Overdispersion, common shocks, heterogeneity, and excess zeros may require negative-binomial or hierarchical alternatives. Claim severity is often modeled conditionally on an occurrence, with Gamma, lognormal, compound, or heavy-tailed models selected with regard to the particular loss layer. For payment per exposure, expected frequency and expected conditional severity should reflect their dependence, not merely the convenience of a factorized expression.

A credibility or partial-pooling component can stabilize estimates for sparsely populated vehicle types, regions, or driver groups. Low-exposure cells are particularly vulnerable to abrupt and unjustified indicated differentials. Cross-validation must respect time, policy linkage, renewal dependence, geographic clustering, and catastrophic loss shocks. Random policy-row partitioning can leak the experience of the same driver or vehicle into both training and validation sets, producing excessively optimistic predictive performance.

The outcome may be all incurred claims, claims above a threshold, bodily injury claims, or non-fault collision claims. Each selection answers a distinct actuarial question. If the claim filing system differentially records events across social groups, fitting a high-capacity model to its labels may faithfully reproduce an administrative recording process rather than an underlying physical hazard. It is necessary to test data-generation validity before comparing fairness metrics computed from those labels.

Driver age and driving experience

Chronological age and years licensed are distinct factors, even though they are often correlated. Experience measures exposure to learning and licensing history, but 'years licensed' can misstate actual distance driven. Age can capture physical capability, life stage, driving patterns, and other correlated circumstances, some of which are affected by technology and mobility design. Their incremental signal should be analyzed conditionally on mileage, vehicle type, driver use, and prior record, with due attention to sample composition and legal restrictions.

Young drivers can differ in exposure duration, vehicle access, and nighttime driving. Older drivers can differ in distance and road choices. Within-category dispersion is substantial. The correct empirical question is whether the candidate factor improves prospective estimation of a defined claim outcome after adjustment for measured exposure and alternatives, and whether its use is permissible. An aggregate accident proportion by age cannot by itself supply a rating relativity.

Experience is also affected by migration, foreign-license recognition, suspended-license intervals, periods without a vehicle, and incomplete insurer histories. Administrative years licensed are a convenient proxy for active driving experience. Where licensing data are weak, model risk may attach to the data system rather than the driver. Sensitivity tests should compare license tenure with observed driving distance and distinguish a novice with high annual exposure from a long-licensed person with little current driving.

Mileage and vehicle use

Distance is an intuitive opportunity measure, but accident risk per mile varies by road, weather, time, traffic, driver behavior, and journey purpose. The response may depend nonlinearly on mileage, and the risk rate may change with accumulated practice. A model of annual count with mileage as a scalar predictor is different from a model with log(miles) as an offset. The first permits non-proportional aggregate frequency; the second assumes proportional exposure while allowing other variables to shape per-mile intensity. The appropriate specification is determined empirically and by the relevant covered peril.

The California system explicitly places annual mileage among its principal factors. For pay-as-you-drive products, actual mileage can be measured periodically, while traditional products often use an applicant's estimate or a rating class. Data quality, odometer verification, policyholder incentives, and the treatment of midterm changes influence both accuracy and perceived fairness. For commercial use, a vehicle's operating hours, route classes, payload, or passenger carrying service can matter more than distance alone.

The UK insurance industry's telematics explanation identifies distance, road type, speed, braking, and driving time among commonly monitored inputs. The parameters of a telematics contract—including any curfew, time-of-day rule, fee, and midterm adjustment—alter the insured's effective exposure and should enter the analysis of fairness. A scoring system that applies a penalty to all nighttime journeys may disadvantage shift workers even where their actual night-driving behavior is cautious. That distributional observation calls for contextual testing; it does not establish an automatic legal violation.

Driving safety record and claims history

Prior at-fault accidents and driving violations may be powerful descriptive indicators, but they are recorded through legal and administrative systems with differing enforcement intensity. A violation count depends on policing, traffic volume, legal definitions, the time since licensing, and opportunities for detection. A claim record depends additionally on insured coverage, reporting incentives, and claims management. These variables should not be collapsed into a vague history score without assessing their constituent mechanisms.

The time window matters: a single event ten years ago may have a different prospective relationship than multiple recent violations. Experience-rated penalties can be moderated by decay, recency, claim type, fault determination, or accident forgiveness. The different treatment of not-at-fault claims, comprehensive losses, and bodily injury claimants should be explicit. Otherwise, the classification can penalize an individual for a harm that did not arise from that individual's driving decisions.

Data quality checks should include de-duplication, consistent fault coding, missing-state records, claims transferred across companies, and vehicle-driver attribution. For household policies, individual drivers may share vehicles; household claim history is a coarse measure of each named driver's behavior. The admissibility of historical information can differ from its predictive usefulness and must be evaluated against jurisdiction-specific rules.

Vehicle characteristics and technology

Vehicle type can affect expected loss through crash compatibility, repair parts, labor cost, theft exposure, utilization, and safety performance. Advanced driver-assistance systems may change crash incidence while making repairs more complex and expensive. Electrical propulsion, battery configuration, driver-assistance software, and connected-vehicle platforms may change both probability and cost. Insurers should distinguish manufacturer model designation from the installed equipment and from actual software or sensor condition.

Predictive models should be evaluated for survivorship: recently introduced vehicles have short claim histories, unusually concentrated buyers, and uncertain repair networks. Limited claims experience may justify partial pooling, external engineering information, or a provisional rate accompanied by planned monitoring. Inference from one vehicle technology, country, or repair-cost period cannot be automatically transferred to another. Model error for new vehicles may be economically important even where average portfolio fit looks good.

Usage-based data may distinguish driver behavior from vehicle capability, but sensor readings are not causally self-explanatory. Harsh braking may indicate aggressive driving, emergency avoidance of another road user, a steep road gradient, or imperfect device calibration. Geographic speed limits and traffic context may be missing or outdated. A classification that rewards a low braking count without contextual validation can misidentify defensive driving as dangerous behavior.

Credit-based insurance scores and risk correlation

The Federal Trade Commission's 2007 automobile-insurance investigation found that credit-based insurance scores predicted claims in the data studied, while their distribution differed across racial and ethnic groups. The FTC stated that scores derived a comparatively small part of predictive power from their correlations with race and ethnicity and reported difficulty identifying an alternative score that both retained comparable predictive performance and reduced group differences. These are findings from an identified historical U.S. dataset and modeling exercise. They do not establish that the score measures driving skill, that causal mechanisms were identified, or that the fairness question has been permanently settled.

The correct scientific interpretation separates three propositions: a score may predict paid losses; it may correlate with group membership; and its use may or may not meet a particular legal and normative standard. The first two are empirical questions with estimation uncertainty; the third requires a jurisdiction-specific legal test and an articulated fairness criterion. A credit-based factor cannot be defended solely by a claim that its regression coefficient is statistically significant. Nor can it be rejected solely by a marginal difference between groups without examining the applicable standard, causal structure, alternatives, and impacts.

Historical insurer data can amplify past market selection. If premiums influence which applicants purchase insurance, the observed insured sample is conditioned on prior prices, availability, and consumer responses. A new score may appear well calibrated on the selected sample yet perform poorly for previously excluded drivers. Off-policy evaluation and sample-reweighting methods can help, but require stated overlap and selection assumptions.

Telematics and granularity

Baecke and Bocca's 2017 telematics selection study compares predictive approaches using sensor-generated vehicle data and finds value in constructed telematics information for the studied risk-selection problem. Its contribution is evidence that additional behavioral measurement can improve some predictions; the study does not establish general fairness, legal admissibility, or universal performance for all vehicle classes and years.

Granular data create a tradeoff among temporal adaptation, robustness, privacy, and exclusion. An insurer may replace imprecise demographic proxies with observed driving patterns and gain closer measurement of exposure. But a highly granular system can produce surveillance costs, device-selection bias, data vendor dependence, and indirect discrimination through work schedules or mobility constraints. A score trained on app users may not generalize to drivers without compatible phones, stable connectivity, or willingness to share continuous location data.

A trustworthy telematics approach records sensor provenance and time stamps, missingness, device class, calibration, map matching, driver attribution, and environmental context. The retention and deletion rules should fit the data's continuing actuarial purpose. In addition, the insurer should specify which measured events alter premiums and provide a meaningful way to challenge a false inference.

Technical price and charged premium

An indicated pure premium for class g is conceptually E[L|g], where L is covered loss on the specified exposure basis. If a multiplicative expense loading a and fixed expense b apply, an illustrative calculation may write P_g = E[L|g]/(1-a) + b, subject to the insurer's actual pricing framework. This expression is a simplified accounting example, not an asserted universal rate-making formula. Taxes, reinsurance, capital, profit and state constraints may require different treatment.

Commercial pricing can deviate from technical expected cost because of loyalty discounts, new-business campaigns, competition, price optimization, expense allocation, renewal rules, and taxes. Fairness evaluation must therefore inspect both model output and the price actually presented to the consumer. Excellent loss calibration can coexist with an unfair distribution of discounts or with inaccessible compulsory insurance.

Proxy Discrimination and Fairness

Direct treatment, indirect effects, and multiple fairness questions

Direct discrimination in insurance classification can occur when a prohibited protected characteristic itself determines a disallowed difference in treatment. Indirect discrimination arises under applicable law when an apparently neutral rule produces a prohibited disadvantage, subject to the jurisdiction's test of justification, necessity, proportionality, or statutory exceptions. In the statistical literature, proxy discrimination often describes a model learning information about a protected variable from correlated nonprotected variables. These concepts overlap but should not be treated as exact equivalents: one is a legal conclusion under a particular regime; another is a modeling diagnosis.

Lindholm, Richman, Tsanakas, and Wüthrich's Discrimination-Free Insurance Pricing provides mathematical treatment of direct and indirect effects in insurance pricing and constructs a pricing approach designed to avoid learning protected attributes through other features. The authors' 2022 published article is an important framework for distinguishing the model's information structure from the insurer's declaration that protected variables have simply been removed.

Removal of a protected column is generally insufficient to establish independence from that attribute. If postal code and occupation jointly reveal group membership, a flexible model may reconstruct much of the missing information. A fairness audit can examine changes in predicted premium, acceptance probability, actual charged price, and errors across groups; compare direct and residual predictive paths; and test alternative admissible variables.

There are several possible questions, each with a different answer: Does the model predict the specified loss? Does it depend on prohibited information? Does it create disproportionate burden? Is the burden justified under applicable law? Does a consumer understand and contest its treatment? Is coverage obtainable at a reasonable price? These questions require separate outcome definitions and cannot be settled through one composite fairness number.

Fairness metrics and incompatible estimands

Statistical parity examines whether the distribution of a decision, such as acceptance or price band, is equal across groups. Conditional calibration examines whether average observed losses align with predicted losses within each relevant group. Equalized odds and related classification criteria compare conditional error rates for a binary outcome. Individual fairness asks whether suitably similar individuals receive similar treatment under a chosen similarity measure. Counterfactual fairness evaluates what changes under specified interventions in a causal model.

These measures can disagree even in idealized settings. The 2022 discussion of insurance fairness criteria shows tensions between common group-fairness axioms and actuarial pricing, using an analytical example without proxy discrimination. Its argument cautions against importing classification fairness metrics into continuous-premium decisions without specifying the actual object and loss function. A claims frequency model with unequal base loss distributions may be calibrated within each group and still produce unequal average indicated prices.

For continuous premiums, one audit estimand is a conditional premium ratio R(z) = E[P | A=a, Z=z] / E[P | A=b, Z=z], where A denotes a protected group and Z denotes prespecified legitimate comparison characteristics. The ratio measures association conditional on Z; it is not automatically a causal or legal discrimination measure. Another is the difference in loss-calibration errors across groups. The relevant comparison must be fixed before model evaluation to avoid selectively reporting favorable fairness definitions.

A multidimensional audit should inspect error calibration, offered premiums, coverage eligibility, nonrenewal, denied claims, and customer disputes. It should also compare people exposed to the market with people actually insured. An insurer's excellent calibration among renewed policyholders may obscure the affordability and availability effects on those who left after large increases.

Causal pathways and counterfactual evaluation

Suppose territory T correlates with race A, driving environment D, claim costs C, and premium P. A useful causal analysis asks whether the T–C relationship reflects road density, local repair cost, socioeconomic circumstances, or administrative claims practices. Without identifying which variables mediate legitimate risk and which reproduce inequitable treatment, a residualized coefficient or feature-importance score cannot decide the issue.

A conceptual directed structure might contain A → residence T; T → road environment D; D → physical accidents; T → repair prices C; and T → insurer class P. Historical disadvantage may also affect vehicle type, driving schedule, and prior insurance availability. An intervention on territory is therefore not equivalent to changing only a mailing address in the rating software. Causal fairness claims demand a carefully defined intervention and justification for holding other variables constant.

Counterfactual tests can nevertheless reveal instability: construct plausible matched records differing in one nonprotected factor, compute premium shifts, examine which changes are legally material, and identify cases needing review. Such tests must avoid synthetic combinations of attributes that are impossible in the covered population. If no common support exists between comparison groups, a model-based disparity estimate may largely reflect extrapolation.

Economic distribution, compulsory protection, and access

Risk-based differentials affect the premium-to-income ratio, the fraction of prospective drivers declining coverage, and insured-pool composition. Mandatory liability coverage increases the consequences of exclusion because legal driving may depend on affordable participation. A high indicated premium for a young, low-income, or geographically disadvantaged driver can be actuarially associated with losses and simultaneously cause a public access concern. These observations describe two measurable dimensions of the same arrangement.

Risk pooling always entails some transfers between realized winners and losers; ex ante price differences also redistribute expected costs among risk groups. The appropriate benchmark cannot be inferred from predictive loss alone. The insurer's solvency, legal rating obligations, affordability programs, and market competition matter. A comparison between completely uniform pricing and fully individualized risk pricing leaves out intermediate practices such as group bands, mandated factors, caps on relativities, assigned-risk pools, and public subsidy.

The distributional audit should ask whether a policy change improves loss prediction, how it shifts premium levels, how many households approach affordability thresholds, how risk selection changes portfolio experience, and which feasible alternative rating designs meet legal and financial constraints. Averages obscure tail risks: a reform may reduce overall error yet substantially increase prices for a small, already vulnerable class.

Why accuracy alone cannot establish fairness

A predictor can be highly accurate because it measures a legitimate insured peril; because it captures a prohibited characteristic; because it learns past underwriting exclusions; or because its outcome encodes differential claims processing. Accuracy metrics do not separate these explanations. Equally, observed disparate impact may be driven by true insured-risk differences, selection, outcome recording, or unlawful proxies, and its interpretation requires additional evidence.

The actuarial research by Lindholm and coauthors provides mathematical tools, but its proposed discrimination-free adjustment does not displace statutory rules and supervisory judgment. Practical implementation requires a stated protected-attribute definition, data access controls, counterfactual assumptions, stable loss definitions, model drift monitoring, and legal counsel to determine which factors are admissible. The proper conclusion of a fairness audit is claim-specific and evidential: what was measured, among whom, when, using which counterfactual, and under which rule.

Privacy and Regulatory Constraints

United States: state insurance supervision and federal consumer reports

The United States does not impose a single national table of permitted motor-insurance rating factors. State legislation, insurance codes, filed-rate review, market-conduct oversight, consumer disclosure, and court interpretations differ. The NAIC automobile-insurance overview is a useful comparative starting point, while enforceable requirements must be identified in the law and regulatory decisions of the relevant state.

California's § 1861.02 and related rate regulations require the factor hierarchy described above. They matter because they prioritize driving safety record, annual mileage, and experience and constrain optional factors. The official regulatory objective is narrower than an abstract command to maximize machine-learning predictive power. A data-driven model that places another variable first may require restructuring before its use in an approved California class plan, irrespective of its fit in an unregulated benchmark.

Colorado offers a governance-oriented example. Colorado Revised Statutes § 10-3-1104.9 addresses unfair discrimination involving external consumer data, algorithms, and predictive models; the Colorado statutory definition includes correlation with specified protected classifications and disproportionately negative outcomes exceeding reasonable correlation with the underlying insurance practice. The Division's August 2025 adoption notice states that amended Regulation 10-1-1 applies its governance and risk-management framework to private passenger automobile insurers, individual life insurers, and health benefit plan insurers using external consumer data and derived models, with an October 15, 2025 effective date. The statute and the regulation should be read together; the notice proves adoption and scope but does not replace the full operative regulatory text or any later individual examination.

When credit reports enter automobile rating, federal requirements also become relevant. The Fair Credit Reporting Act regulates permissible use and consumer protections; the U.S. Code's adverse-action provisions address notification when specified decisions are based on consumer-report information. The precise notice duty depends on facts and statutory definitions. State restrictions may independently limit the role of credit even where federal reporting requirements are met.

The NAIC's 2023 AI model bulletin announcement records a model supervisory instrument emphasizing governance, fairness, applicable insurance law, and regulator access to documentation. A model bulletin's adoption by NAIC does not automatically make every word enforceable in every U.S. jurisdiction; adoption and binding implementation must be assessed state by state. The NAIC's AI regulatory topic update also describes ongoing third-party-model oversight work and a state pilot of a systems-evaluation tool in 2026. Pilot activity must be reported as a pilot, not as a universally applicable national regulation.

European Union: equality law and insured classification

Directive 2004/113/EC addresses equal treatment between women and men in access to goods and services, including insurance. The Test-Achats ruling invalidated Article 5(2)'s indefinite actuarial derogation effective December 21, 2012. This example demonstrates that sex-based premium differences may be impermissible even where sex predicts expected losses. The effect operates through EU equality law and national implementation, rather than a universal statistical theorem about group averages.

The ruling does not determine all questions about age, territory, experience, or vehicle type. Nor does it provide an automatic judicial finding that a correlated neutral characteristic constitutes impermissible indirect discrimination. The directive's equal-treatment provisions set the relevant legal frame; national enactments, case law, and the factual mechanisms of a specific rating factor require independent review.

European Union: personal data and algorithmic decision-making

Personalized motor pricing commonly processes driver identity, claims history, location, journeys, vehicle identifiers, and sometimes highly granular behavioral records. The General Data Protection Regulation requires an applicable lawful basis, transparent and purpose-limited processing, data minimization, accuracy, security, and accountability. Articles 13–15 address information and access rights; Article 22 sets requirements and protections for certain decisions based solely on automated processing that produce legal or similarly significant effects. Application depends on decision architecture, legal basis, exceptions, and national or EU rules. It would be inaccurate to assert that every algorithm-assisted quote is automatically prohibited by Article 22.

The risk of inference is wider than the fields collected. Continual journeys can reveal work sites, religious attendance, health visits, household relationships, and behavioral routines. A raw GPS trace may become especially sensitive through linkage even where the immediate rating model uses only mileage and time of day. Data minimization therefore concerns the collection and retention of underlying observations, not merely the number of final model covariates. An insurer must also distinguish the data it needs to issue a rate from data maintained for fraud investigation, claims settlement, or scientific study.

The EU AI Act's Annex III explicitly lists certain AI risk-assessment and pricing systems in life and health insurance among its high-risk use cases. That entry does not categorically designate all motor-insurance pricing systems as high-risk solely because they are insurance systems. Other AI Act obligations and classifications may apply according to the specific system and its function; a motor-rating model should be assessed on that basis rather than borrowing a life-and-health category.

EIOPA's insurance AI governance report recommends proportional governance, fairness, transparency, data quality, robustness, and human oversight. It is a governance reference, not a standalone substitute for the GDPR, equality law, insurance directives, or national binding provisions. The distinctive European regulatory lesson is that lawfulness of data processing and lawfulness of the pricing distinction are separate questions requiring separate review.

United Kingdom: motor pricing and insurance equality

The Financial Conduct Authority's ICOBS 6B rules prohibit specified motor and home insurance renewal prices from exceeding the equivalent new-business price, subject to defined comparison methods and scope. This is an intervention into charged-price practices rather than a prohibition on all risk-factor variation. The FCA's implementation policy statement sets January 2022 implementation of key pricing remedies and distinguishes them from other earlier systems-and-controls requirements. The rule makes clear that fairness review includes the commercial layer after actuarial risk estimation.

The Equality Act 2010, Schedule 3 contains insurance-related exceptions and later amendments. Current consolidated text matters: historical enacted text includes provisions that were subsequently omitted, including a former insurance exception in paragraph 22 concerning sex, amended in December 2012. Any assessment of a British driver's present rights must use the current operative version, relevant exceptions, and current regulator interpretation. Legal effect cannot be established from a historical enacted version alone.

The UK's Financial Ombudsman telematics guidance documents real complaint categories involving inaccurate driving scores, unexplained pricing consequences, and disputed cancellation. It describes evidence a reviewer may request, such as data logs, policy information, and insurer explanations. This supervisory-adjacent practice emphasizes that a model's accuracy and explanatory quality become concrete issues when a household challenges a consequential decision.

Telematics data and consent boundaries

A telematics product may have multiple data actors: insurer, policyholder, named driver, mobile application provider, vehicle manufacturer, device vendor, cloud processor, and claims administrator. Contractual responsibility and privacy-law roles must be mapped rather than inferred from who owns the device. A policyholder's acceptance of terms does not automatically prove that every named driver has received adequate information or that all secondary processing has a valid legal basis.

The Association of British Insurers' telematics guide highlights data-sharing disclosures, access, driving-behavior interpretation, and variation among product terms. Continuous location data can support exposure measurement but can also affect the privacy of other travelers. A proportionate design should explain monitored measures, collection frequency, retention, third-party transfer, cancellation consequences, dispute procedures, and whether an opt-out has meaningful alternatives.

The technical implementation of privacy protection requires encryption, access segmentation, deletion practices, identity minimization, trace retention limits, contractual controls over vendors, and review of model inversion or re-identification risks. A privacy-preserving aggregate driving score may reduce exposure of a raw trip history, but the model still needs evidence that aggregation preserves relevant predictive calibration and fairness monitoring. More aggregation is not automatically better if it suppresses evidence needed to correct biased treatment.

Cross-jurisdictional comparison

Jurisdiction or institution Characteristic focus Operative or institutional source Implication for risk classification
California, United States Statutorily ordered auto rating factors Insurance Code § 1861.02 and rating regulation Driving record, mileage and experience receive defined priority; permitted optional factors remain controlled
Colorado, United States External consumer data and model governance Statute § 10-3-1104.9 and amended Regulation 10-1-1 Unequal results and model governance are examined through state-specific definitions and oversight
United States federal level Consumer-report information Fair Credit Reporting Act Permissible purpose, accuracy and notice requirements are distinct from state insurance rate rules
European Union Sex equality in insurance Directive 2004/113 and Test-Achats Sex-based premium variation is restricted regardless of model predictiveness
European Union Personal-data processing and certain automated decisions GDPR Data lawfulness, purpose, transparency and rights attach to individualized processing
European Union AI classification for specified functions AI Act Annex III Express life/health insurance classification cannot simply be generalized to all motor rating
United Kingdom Commercial renewal pricing FCA ICOBS 6B Renewal and equivalent new-business prices are separately constrained
United Kingdom Insurance equality and dispute redress Equality Act and Financial Ombudsman Current exceptions, factual data quality and consumer redress affect lawful outcomes

A comparative table is a navigation aid, not legal advice. Texts differ in legal status: enacted statutes, consolidated regulations, judicial holdings, model instruments, and nonbinding guidance must retain their identity in any country-specific implementation.

Comparative Evidence, Data Uncertainty, and Validation

Evidence design and source hierarchy

A valid synthesis distinguishes the role of each evidence family. WHO and ITF are official road-safety surveillance, primarily informative about public outcomes and exposure context. The FTC's 2007 work is an empirical analysis of historical American automobile-insurance and credit-score information. Peer-reviewed actuarial research gives formal fairness definitions and model proposals. The Baecke–Bocca telematics study examines data-informed predictive selection. National statutory materials establish legal constraints; insurer or association guidance describes practices rather than authoritative compliance findings.

These materials answer complementary questions. They cannot be pooled into a single effect estimate of discrimination because they have different populations, policy instruments, years, outcome definitions, and measurement processes. No estimate of a universal 'fairness gap' can be derived from their combined existence. An evidence table should preserve the original estimand and the institutional competence of its source.

Evidence family Proper estimand or question Principal limitation for a motor-rating fairness audit
WHO global road safety Road traffic injury mortality and policy context Population crash mortality does not identify an insured claim cost function
ITF/IRTAD safety series National traffic-safety indicators Country coverage, vintage and denominator differences limit tariff transfer
FTC 2007 study Credit score prediction and group distributions in the studied policies Historical market, selection, model design and protected-group inference
Telematics selection study Incremental predictive value of observed vehicle behavior in a study setting Sensor adoption and claims-label transportability; fairness not directly tested
Discrimination-free pricing research Mathematical properties of an adjustment to pricing Model assumptions, protected data and law-specific admissibility
Statutes and supervisory texts Binding rules or official governance expectations Jurisdiction, effective date, legal status and coverage-specific interpretation
Industry and ombudsman guidance Product practices, disputes and information quality Descriptive and complaint-selection effects; no population incidence estimate

An auditable comparison population

A reproducible motor-rating evaluation should start with policy and exposure definitions. Include the effective policy dates, rating jurisdiction, coverage, deductible, vehicle class, drivers, earned exposure, reported and ultimate claim development, monetary trend, and exact rating version. Identify how historical policies changed across tariff revisions and whether rate capping or renewal discounts altered the premium from the model indication.

For each candidate characteristic, record its provenance and measurement: declared annual mileage versus odometer, license issue date versus active driving, address versus vehicle garaging, device-recorded speed versus posted-limit lookup, or insurance score versus consumer credit record. Record missingness both overall and by group. Missing sensor observations may reflect technology failure, deliberate disengagement, low connectivity, or changes in product participation. Treating every missing value as an intermediate risk band without analysis can be discriminatory and predictively unsound.

Select validation cohorts by time and underwriting decision point. A model intended for new business should be evaluated on appropriately out-of-time new applicants; one designed for renewal should use renewal decisions and known censoring. For long-tail injury claims, compare evaluation periods on a consistent ultimate-loss basis or explicitly acknowledge the immaturity of recent years. Normalize losses for inflation, repair costs, changes in coverage and reinsurance only with transparent methods.

Mathematical audit specification

For a group g and validation period t, define earned exposure E_gt, observed claims N_gt, and the model's summed expected claims M_gt. An aggregate count-calibration ratio is C_gt = N_gt / M_gt where M_gt > 0. Its interpretation should use interval estimates and preferably case-level calibration curves; a ratio near one can mask major errors among subgroups. For a paid-loss model, compare actual developed losses and predicted losses on the same accounting basis.

For model feature sets X_0 and X_1, the change in out-of-time proper scoring loss is Δ = Loss(X_0) − Loss(X_1). Its sign indicates incremental predictive value on the chosen evaluation sample; it does not resolve legal admissibility. Feature ablation should be supplemented by conditional permutation, stability tests, and analysis of correlation among factors. Highly correlated age and experience measures can substitute for one another, making an individual coefficient or importance ranking unreliable as a fairness explanation.

For the pricing distribution, measure conditional differences D_g(z) = E[P|A=g,Z=z] − E[P|A=reference,Z=z], as well as premium-to-income ratios, quote acceptance, binding, cancellation, nonrenewal, and coverage gaps where records permit. Specify which Z variables are legitimate comparison factors; changing that set changes the question. Provide confidence intervals or uncertainty bands where sample size allows, and disclose suppressed low-count cells. Auditing a disparity without a denominator or uncertainty assessment is an incomplete statistical statement.

A simple hypothetical illustration clarifies the distinction. Suppose two equal-sized groups each have the same observed annual claim frequency of 0.05 under identical coverage, but one group is charged $1,200 on average and the other $1,500. The $300 difference requires examination of expenses, severity, policy terms, and permissible covariates. It does not automatically prove discrimination because equal claim frequency alone does not establish equal expected covered loss. Conversely, if both groups have the same fully adjusted expected loss and all permissible expenses, a persistent price difference may indicate commercial optimization, omitted legitimate factors, or prohibited treatment. The figures are invented solely for analytical exposition and are not derived from market data.

Model selection, nonstationarity, and data bias

Risk models must withstand out-of-time testing and population drift. Driver behavior, traffic patterns, vehicle systems, claims inflation, litigation, weather, urban development, and regulatory factors change. A factor that predicted 2015 claims may perform differently in 2026. An insurer should monitor cohort performance, calibration by risk band and group, quote conversion and drift in external data suppliers. Retraining can alter disparate outcomes even if the model's name and feature list stay constant.

The 2007 FTC findings are historically specific. They cannot be treated as a current direct measurement of a 2026 insurer's credit-scoring system. The 2022 Lindholm et al. fairness research is a formal method, not an empirical national audit. The 2017 telematics study cannot decide whether a particular insurer's app fairly interprets harsh braking or day/night patterns. The 2023 WHO report and the ITF report dated February 18, 2026 cannot be cited as evidence that current insured motor rates are actuarially adequate. These scope constraints are substantive scientific findings, not editorial trivia.

Model evaluation should test sensitivity to economic conditions, inflation assumptions, claim development, alternative exposure denominators, unobserved selection, missing protected attributes, and the decision to exclude certain features. Where protected-group labels cannot lawfully be used in ordinary pricing, governance may require a separately controlled and legally supported audit dataset. The permissible audit design is not identical to the permissible production model.

Source-level and record-level verification

Source verification has two dimensions. Bibliographic verification asks whether the document exists, has the claimed authors, version and date, and supports the attributed proposition. Inferential verification asks whether the empirical design supports the strength of the manuscript's conclusion. A legally accurate sentence about Colorado cannot be generalized to the United States as a whole; a study demonstrating correlation cannot be rewritten as a causal demonstration; an advisory EIOPA report cannot be presented as binding European law.

The present source set prioritizes regulator publications, statutory texts, court judgment, institutional safety datasets, and publisher records of peer-reviewed research. For proprietary empirical work, accessible abstracts and institutional records are distinguished from examination of full underlying datasets. An original insurer cohort was unavailable, so no claims are made about fitted coefficients, exact effect sizes, protected-group outcomes, or external validation for a specified carrier.

Reproducible review sequence

An applied audit should proceed through the following connected operations: define covered loss and insured exposure; reconstruct contracts and rate version; check data permissions and retention; establish the out-of-time cohort; verify label consistency and loss development; fit and calibrate baseline and candidate models; examine factor increments and stability; identify potential protected and proxy pathways; evaluate distributional and availability outcomes; compare feasible legal alternatives; test third-party model controls; review consumer explanations and disputes; archive the decision with sources and versioned evidence. Each operation has its own acceptance criterion. A model may be approved for prediction while remaining under review for legality or consumer-impact concerns.

External audit, when commissioned, must be able to reproduce the decision without needing undisclosed claims about a model vendor. That may require confidential regulatory access rather than public release of identifiable driver data. Confidentiality and reproducibility are compatible when access, method, version, and change history are adequately governed. The NIST AI Risk Management Framework offers a voluntary, cross-sector vocabulary for governance, mapping, measurement, and management; it does not replace specific insurance and privacy obligations.

Integrated Results and Open Research Questions

Findings across the three analytical dimensions

The reviewed evidence supports an integrated assessment of motor-insurance classification.

First, the classification target must be a clearly defined insured outcome. Physical accident probability, claim frequency, ultimate covered severity, and charged premium are not interchangeable. The causal and normative evaluation changes when a factor predicts claim reporting or repair prices rather than driving hazard.

Second, driver age, experience, mileage, recorded history, vehicle characteristics, territory, and telematics contain overlapping information. The incremental value of any factor depends on the available alternatives, coverage, cohort, year, and measurement method. A variable's prominence in a market list is not evidence of universal legitimacy or transportable predictive power.

Third, the FTC's historical evidence establishes that a credit-based score can predict claims and display demographic disparities simultaneously. A fair interpretation preserves both findings and the age and geography of the evidence. Prediction and demographic distribution answer different questions.

Fourth, proxy risk cannot be measured by deleting a protected variable from the design matrix. Correlated features may convey similar information. Formal actuarial research provides alternative pricing constructions, but the applicable law and policy objective determine what counts as an unacceptable pathway.

Fifth, a fairness test should distinguish loss calibration, pricing disparity, offer availability, insured participation, and procedural redress. A single group fairness score suppresses economically and legally important dimensions. In particular, a well-calibrated model can still produce significant social access consequences.

Sixth, indirect discrimination requires an identified rule, affected group, comparison, causal account, and legal standard. Disparities alone do not prove a specific prohibited mechanism; predictive justification alone cannot immunize a disallowed classification. Both shortcuts fail on their own terms.

Seventh, telematics may reduce dependence on coarse proxies through direct observation of driving behavior, but adds device selection, context errors, intimate location data, and opaque scoring mechanisms. Its scientific value depends on whether an appropriately controlled study actually demonstrates both risk estimation and accountable treatment.

Eighth, the legal comparisons reveal multiple regulatory architectures. California orders rating factors; Colorado addresses external consumer data and algorithm governance; U.S. federal law governs aspects of consumer reports; EU equality law restricts sex-based insurance differentiation; the GDPR regulates personal-data use; and the UK FCA constrains specified renewal pricing practices. These architectures cannot be replaced by one universal checklist of legally acceptable variables.

Ninth, statutory, model, and guidance documents occupy different levels of authority. The NAIC model AI bulletin supplies a widely relevant governance reference but requires jurisdictional adoption assessment; EIOPA and NIST are important nonbinding governance sources; actual legal prohibitions and permissions arise from applicable law and decisions.

Tenth, scientific validation must account for model selection and population selection. A rate affects whether a consumer purchases and retains coverage, thereby changing the future labeled data. Longitudinal fairness monitoring should examine market exit as well as claims among the insured.

Eleventh, model governance and consumer rights belong in the substantive design. An insurer should be able to explain why a driver's price changed, how contested data were checked, and how erroneous scores or vendor records were corrected. Effective challenge procedures contribute evidence about measurement failures.

Twelfth, this synthesis supports a practical separation between scientific assurance, editorial assurance, legal implementation, and public deployment. A manuscript about classification can meet scholarly expectations through traceable sources and transparent uncertainty without claiming to have conducted a population-level causal experiment.

Priority questions for subsequent research

Several questions remain important for empirical work, without requiring the creation of a new theoretical framework. How much of a territory factor's incremental predictive signal reflects traffic environment, repair markets, claim handling, and socioeconomic composition? When device data replace broad demographic features, do outcome disparities decrease after accounting for driver adoption and missingness? How stable are fairness and calibration results across inflation shocks, vehicle technologies, and policy cohorts? Which affordability effects arise after the market reacts to a change in indicated premiums? How much of a premium difference derives from loss cost, distribution expense, pricing optimization, or a regulatory constraint?

Other testable questions concern measurement itself. Does vehicle-mile exposure better predict observed insured claims than annual policy-years once distance is independently verified? How do estimates change when at-fault incidents are separated from not-at-fault and theft claims? Which telematics variables are robust to device class and road context? What is the minimum identifiable data needed to validate a score while minimizing raw-location collection? Can a regulator reproduce a third-party model's effects under secure access conditions? Which legal exemptions or consumer protections apply to the exact product and version being evaluated?

These questions can be answered only with relevant policy, quote, exposure, claims, and protected-class audit data assembled under a valid legal basis. No single observational study can resolve them for every jurisdiction and insurance coverage. The appropriate next empirical unit is a prespecified, jurisdiction-specific design with a recorded target outcome, comparisons, sample inclusion rules, and decision-relevant uncertainty.

Authorial analytical contribution

This monograph's original contribution lies in the structured reconciliation of actuarial loss measurement, the semantics of rating factors, fairness estimands, legal admissibility, and consumer redress into a single source-based analysis of driver and vehicle classification. It identifies common category errors in transferring crash mortality to insured loss rates, interpreting score prediction as causation, treating removal of protected variables as conclusive fairness, equating model indication with charged price, and extending one jurisdiction's legal rule to another. These interpretations are original synthesis and editorial reasoning grounded in cited literature and public primary sources. They do not institute a named theory, doctrine, mathematical model, or new legal principle.

Conclusion

Driver and vehicle risk classification is an evidence-dependent decision system operating within a contract, a market, and a legal jurisdiction. Its predictive content is generated by exposure, behavior, history, vehicle characteristics, claim definition, and model design. Its fairness is evaluated through the way those inputs are collected, the kind of insured loss they actually forecast, the distribution of underwriting and price outcomes, the presence of prohibited direct or proxy pathways, and the availability of transparent and effective review.

The strongest scientific conclusion is methodological and substantive. Risk classification requires a specified insured outcome and a sound exposure denominator; pricing fairness requires a clear comparison population, auditable statistics, legally applicable distinctions, and attention to access; data protection requires a justified purpose and proportionate processing. A model can perform well on one dimension and require correction on another. None of these dimensions can be inferred from model accuracy alone.

Comparative law confirms the necessity of precise attribution. California's ordered rating factors, Colorado's governance rules, the FTC's credit-score evidence, EU sex-equality obligations, the GDPR, the AI Act's expressly listed insurance category, and the UK renewal-pricing framework each answer distinct institutional questions. Historical empirical findings and legal texts must retain their original date, scope, and status. Market conditions and regulatory implementation warrant renewed study before operational decisions.

A mature research and supervisory practice therefore combines prospective prediction, loss development, counterfactual scrutiny, distributional analysis, data governance, interpretability, consumer remedies, and longitudinal monitoring. The result is a more defensible account of what motor-insurance classifications measure, whom they affect, and when their use can be supported by evidence and law. This research is an internally reviewed scientific synthesis for a reserved canonical route. Its public publication and production verification remain separate future acts.

References