RESEARCH DOSSIER
Telematics-Based Motor Insurance Pricing and Risk Scores
Abstract. Telematics-based motor insurance connects observed vehicle use to the estimation, classification, communication, and pricing of insured motor risk. Its actuarial importance lies in the simultaneous measurement of exposure, behavior, and driving context. A trip sensor records events; an insurance contract defines covered losses; a claims system records insured outcomes; and a ratemaking process translates conditional expected cost into a premium. Their combination can improve classification while creating difficult questions about measurement error, participation, behavioral incentives, price stability, data protection, and distributional fairness. This monograph analyzes those connections through primary actuarial research, controlled behavioral studies, supervisory documents, and legal instruments. It separates pay-per-mile, pay-how-you-drive, and dynamic usage-based contracts; constructs the mathematical relationships among mileage, trip composition, event rates, claim frequency, severity, and indicated premium; and explains how calibrated risk scores should be assessed under coverage-specific exposure definitions. An original analytical synthesis shows that a pricing score is best understood as a versioned estimate conditional on a defined observation system, contract, population, and date. It also demonstrates why predictive discrimination, causal safety effects, and equitable access require different evaluation designs. Published empirical results support the informational value of mileage, speeding, driving context, and some harsh-driving indicators, while randomization studies reveal heterogeneous effects of feedback and financial incentives. The legal comparison focuses on United States actuarial and supervisory guidance and European privacy and insurance governance, with explicit attention to the narrower life-and-health provisions of the EU AI Act. The conclusions identify reproducible requirements for tariff construction, prospective validation, consumer explanation, and disciplined model change without proposing a new named theory or reporting invented empirical estimates.
Keywords: Telematics; motor insurance; automobile insurance; usage-based insurance; pay-as-you-drive; pay-how-you-drive; risk scores; actuarial ratemaking; claims frequency; exposure measurement; driver behavior; adverse selection; calibration; predictive fairness; GDPR; insurance governance.
Introduction
Motor insurance rates arise from contracts covering different combinations of bodily injury, property damage, first-party vehicle loss, medical expense, legal liability, and associated claim costs. A record of fast braking has no intrinsic monetary value: its relevance depends on how reliably it predicts a covered event or responds to a behavior-changing intervention. Accordingly, telematics pricing begins with a precise target: the future insured cost of an identifiable coverage over an identified contract period, subject to the actual policy wording, deductibles, limits, location, and applicable law.
The NAIC's telematics overview, describes usage-based insurance as monitoring mileage and driving behavior through on-board devices, connected-car systems, and mobile applications. It identifies road location, time of day, acceleration, cornering, braking, and other signals that may contribute to rating. These measurements allow a finer view of vehicle use than a proposal form or an annual mileage declaration. Finer measurement is valuable when the observation system is sufficiently reliable and the additional distinctions improve decisions about future claims rather than simply multiplying numerical categories.
The essential questions are therefore simultaneously statistical, contractual, and institutional. Which units of exposure explain claim incidence? Which sensor behaviors predict losses after accounting for congestion, route type, vehicle technology, weather, and driver selection? How can an insurer change a premium on the basis of a score while preserving adequate coverage-specific pricing, understandable terms, privacy, and access? Which driver reactions are established by experimental evidence, and which associations only express self-selection or different baseline risk? These questions organize the six substantive chapters.
The method is a critical synthesis of published studies and authoritative instruments available by November 15, 2026. Mathematical examples below are original analytical demonstrations; every monetary amount and hypothetical coefficient in such examples is expressly illustrative and is not an estimated result from an insurer. Jurisdiction-specific obligations are identified as such. Public research often relies on selected portfolios, proprietary telematics records, and short observation windows, so findings are transferred cautiously to other countries, road environments, coverages, and product types. No new dataset, field experiment, industry-wide empirical estimate, or external peer-review result is claimed.
Conceptual, Historical, and Primary-Source Foundations
The insured object, exposure unit, and observed event
A motor policy establishes the relationship between an insured interest, a covered peril, a limit or indemnity basis, and a specified period. The risk observed by a telematics system is typically a driver-vehicle-trip process; the insured liability measured by claims data is a contract-conditioned monetary obligation. A single collision may produce several claims under several coverages, while a telematics near miss may produce no claim. A driver may travel many miles without a reported loss; a parked vehicle may still experience theft, fire, or weather damage. Rate construction must preserve those distinctions.
Three commercial families illuminate the evolution of the field. Pay-as-you-drive or pay-per-mile programs attach price more directly to quantity of use, often with a base component for standing risk. Pay-how-you-drive contracts add observed driving habits to mileage. More dynamically updated arrangements price or communicate recent behavior over short windows while still requiring explicit rules on when a quoted or renewed premium can change. The NAIC's usage-based insurance descriptions document the technologies and measured characteristics without establishing that every program uses identical scoring, premiums, or consumer consequences.
The relevant exposure denominator depends on the peril. Vehicle-years are natural for annual policies and some fixed hazards; distance is often useful for moving-vehicle collision incidence; trips or hours can matter when risk is concentrated at starts, stops, intersections, nighttime travel, or periods of congestion. Each denominator creates a different causal and predictive question. Claim frequency per mile cannot automatically substitute for claim frequency per year, because the annual number of miles is itself a determinant of total exposure. Likewise, two identical distances driven on different routes need not produce identical hazard distributions.
From rating classes to behavioral data
Conventional motor ratemaking relies on historical policy and claims records, claims experience, vehicle characteristics, insured use, territorial variables, and approved classification factors. Telematics adds repeated within-policy observations. In a Belgian sample of young drivers, Verbelen, Antonio, and Claeskens (2018) used the predictive information in telematics data to investigate how driver behavior and vehicle use may contribute to insurance pricing. Their source population matters: a telematics enrollment program aimed at young drivers cannot by itself support a universal parameter for all vehicle ages and jurisdictions.
Ayuso, Guillén, and Nielsen (2019) developed mileage and driver-behavior ratemaking in a frequency-model setting, explaining how telematics could supplement an insurer's previously established pricing knowledge. That transition is institutionally important. An insurer usually has many more policy-years of ordinary experience than telematics-equipped years of fully matched sensor and loss observations. Discarding the former may increase estimation variance, especially for rare and expensive bodily injury outcomes.
Guillén, Nielsen, Ayuso, and Pérez-Marín (2019) examined telematics for automobile insurance rates using risk exposure and driving habits in a claim-count context with substantial zero observations. Zero claims can result from little driving, safer conditions, a low underlying risk rate, or differences in claiming; no single interpretation follows automatically from a zero count. Their findings regarding mileage and behavior show why a pure distance rule is a useful starting point but an incomplete account of observed variation.
Subsequent research has shortened the observation interval. Guillén, Pérez-Marín, and Nielsen (2024) studied weekly motor pricing with telematics using 19,214 drivers and approximately 181.4 million kilometers from a Spanish insurance program; the authors report 922 at-fault accidents in the described dataset and analyze weekly exposure and context. Their evidence demonstrates the informational potential of recent behavior, while the authors' data-access limitation prevents independent reproduction of the full estimation from openly accessible individual records. Yanez, Guillén, and Nielsen (2025) subsequently investigated a weekly telematics bonus-malus approach, connecting observed signals with short-horizon adjustments and credibility. These studies concern different scoring constructions and do not establish that real-time tariff changes are legally available under every policy.
Actuarial standards and the status of regulation
For United States actuarial practice, the Actuarial Standards Board's ASOP No. 12 provides professional guidance on risk classification, including considerations relevant to selecting and evaluating classes. The Board's ASOP No. 53 addresses estimates of prospective property/casualty costs. These standards govern relevant actuarial services under their stated applicability and are not universal statutes dictating an identical telematics tariff in every US state. State-specific insurance rules, filings, contract terms, privacy law, and permissible factors must be examined separately before implementing a price.
A foundational distinction is between risk classification and tariff permission. A variable can exhibit predictive information in a historical study while its use is legally restricted, insufficiently supported in a local filing, excessively opaque to consumers, or unstable across technologies. Conversely, a permissible factor might have modest incremental information and require credibility adjustment. Scientific appraisal and legal authorization are cumulative requirements of an operational product.
Limits of what historical evidence proves
The historical record supports three developments: distance-sensitive exposure measurement, integration of driving behavior and context with conventional actuarial models, and experiments in shorter premium or feedback intervals. It does not support a claim that telematics universally lowers total claims, identifies a driver's inherent moral quality, eliminates claim uncertainty, or automatically produces a fair market. Any statement about reduction in accident rates must identify an appropriate counterfactual; any statement about lower premiums must identify the rate structure, expenses, market competition, selection, and redistribution across insureds.
Telematics Features and Scoring
The sensor-to-loss measurement chain
A credible telematics score starts with a documented chain from raw measurement to contracted outcome. The raw stream may include timestamp, position, speed, acceleration vectors, gyroscope readings, device status, trip start and end, and smartphone interaction. Processing identifies trips, associates them with the insured vehicle or eligible driver, matches roads and speed limits where permitted, detects events, and constructs exposure summaries. The resulting features enter a model using historical policy characteristics and a claims ledger. Each link can generate error: GPS drift, outdated speed limits, phone-passenger misclassification, device dropout, firmware changes, trip omission, claim reporting delay, or uncertain identity of the actual driver.
A typical risk-feature inventory includes distance by road class, distance by time band, miles above a defined posted speed-limit threshold, severe acceleration and braking counts, harsh cornering, active handheld phone exposure, and coverage-specific vehicle information. A rate such as 12 hard-braking events per 1,000 miles is interpretable only with its event threshold, filtering logic, sensor version, road context, and observation window. A single hard brake on a crowded road can be prudent collision avoidance. Rates should therefore be interpreted statistically across exposure and context rather than as proof that an isolated maneuver was blameworthy.
Let M_i denote validated distance for policy i, H_{ik} the count or duration of event type k, and C_i the relevant trip-context vector. A simple feature is Z_{ik}=1,000 H_{ik}/M_i for M_i>0. For M_i=0, Z is undefined; assigning zero would incorrectly equate absence of exposure with demonstrated safe driving. Appropriate implementations use an explicit no-travel state, a minimum-exposure credibility rule, and a separately modeled standing component. Different modes of missingness require different treatment: an unobserved trip is not the same as a documented trip of zero miles, and an app turned off during travel is not equivalent to a week with no travel.
Counts, exposure, and cost
For an appropriate moving-vehicle peril, a familiar actuarial frequency specification is
N_i | X_i,T_i,M_i ~ Poisson(μ_i), log μ_i = log M_i + α + β'X_i + γ'T_i,
where N_i is a defined count of eligible claims, X_i contains conventional rating variables, T_i is a telematics feature vector, and M_i is consistent validated mileage over the same outcome period. When mileage has nonlinear relationships with loss frequency, the imposed unit coefficient on log mileage requires testing; a flexible model can include functions of distance or separate baseline time and driving exposure. A negative-binomial alternative may allow overdispersion, while zero-inflated formulations require their own identifiable interpretation. Guillén and colleagues' zero-inflated count analysis shows why structural and ordinary zero outcomes deserve explicit consideration.
The pure premium for coverage c is the conditional expectation of aggregate insured loss and allocable loss-adjustment expense, not a score percentile. Under a simplified frequency-severity decomposition,
Π_{ic}^{pure} = E[N_{ic}|F_i] × E[S_{ic}|N_{ic}>0,F_i],
provided the relevant conditioning and claim definition are consistent. More generally, it is E[Σ_{j=1}^{N_{ic}} S_{icj}|F_i]; dependence, multiple covers, policy conditions, inflation, limits, deductibles, and costs can invalidate a naive product of marginal sample averages. Telemetry may predict frequency more readily than rare injury severity. Damage repair cost, medical inflation, liability jurisdiction, and repair-network effects also demand independent modeling. Usman and colleagues (2024) investigate claim prediction and premium pricing with telematics and regularized Poisson regression, showing a route from behavior variables to loss prediction without establishing a universal fixed tariff coefficient.
The actuarial price indication then includes acquisition and administrative expenses, cost of capital, target margin, taxes or levies where applicable, and adjustments required by filing and product constraints. A predictive score should not conceal these transformations. If two drivers have equal modeled pure loss but different final prices, the model documentation must explain the legitimate non-loss components or business rules behind the difference.
Normalized risk scores and identification
A useful scalar score may summarize estimated relative covered loss:
r_i = exp(f(X_i,T_i)) / E_{ref}[exp(f(X,T))],
where the denominator is the exposure-weighted mean prediction for a declared reference portfolio and model edition. This normalization assigns mean score 1 to the reference portfolio but does not make score 1 a universal or time-invariant absolute risk. A consumer-facing score of 0–100 can be generated by a monotone transform of r_i, with clearly published orientation, ties, percentile population, update rules, and consequences. Percentile ranks change when the reference distribution changes even if an individual risk estimate is stable.
At least four distinct objects should be documented: a trip-event score, a frequency multiplier, an expected-loss estimate by coverage, and the final tariff or discount. Combining them into one number obscures causal and actuarial meaning. The same driver could have moderate collision risk per mile, high annual mileage, relatively low noncollision standing risk, and substantial third-party bodily injury limits. One common score can communicate an approximate behavior profile, but coverage-specific expected costs remain necessary for pricing.
The time dimension
Weekly scoring raises an important temporal distinction. Features recorded during week t may forecast claims in week t+1, provide contemporaneous safety feedback, or determine an adjustment to a future renewal premium. These operations have different targets and leakage risks. A score computed after a reported accident might incorporate post-event behavior, claim handling, or device changes. Training a prospective tariff on such information inflates apparent predictive accuracy. Guillen, Pérez-Marín, and Nielsen (2024) report results using lagged weekly driving information, which supports the feasibility of more temporally disciplined comparisons within their dataset.
Production systems should define timestamp boundaries, event ingestion delay, late-arriving telemetry, allowed backfills, claim occurrence versus notification dates, and a freeze date for rating decisions. An insurer may update safety feedback frequently while changing contractual premiums only on authorized rating dates. When a policy provides rolling changes, notice, rate caps, and consumer access to supporting data become essential elements of the pricing design.
An explicitly illustrative tariff calculation
Consider a hypothetical policy with a fixed annual pure-loss component of $80 covering standing exposure, a moving-risk pure-loss cost of $0.018 per validated mile, 6,000 annual miles, and a normalized behavioral multiplier of 1.25 for the variable component. Its modeled pure loss is 80 + 6,000 × 0.018 × 1.25 = \$215. Adding an illustrative 20% multiplicative loading and $50 fixed expense produces (215 × 1.20) + 50 = \$308. With the same exposure and a variable-risk multiplier of 1.00, the analogous price would be (80 + 108) × 1.20 + 50 = \$275.60. The $32.40 price difference is a constructed illustration of one mapping from score to tariff. None of these numbers comes from a fitted insurer study, and the assumed loadings have no market or regulatory status.
This calculation demonstrates why describing a score change as a percentage premium change can mislead. A 25% increase in the variable-risk multiplier produces only an approximately 11.8% change from $275.60 to $308 when a standing component and fixed expenses are retained. Different deductible, coverage, loading, and mileage structures generate different elasticities. Tariff documents should display these relationships explicitly, especially where a large score swing generates a smaller or larger final customer change.
Credibility, shrinkage, and measurement reliability
A newly enrolled motorist with 15 valid miles and three detected harsh events has little stable exposure for estimating an annual habit. The scoring system can shrink a short-window estimate toward an appropriately defined prior or portfolio expectation. This is consistent with established actuarial credibility reasoning and with the dynamic bonus-malus research. Reliability depends on exposure quantity, event detection error, dependence between trips, and whether recent change is itself predictive. Credibility is therefore a statistical property of information, not merely a fixed time-on-program threshold.
A defensible implementation should retain a transparent fallback for device outages and unusual use, limit pricing shocks from extremely small denominators, maintain separate vendor-specific calibration, and audit changes in detection sensitivity. The score may be continuous, but the evidence supporting it is finite, selected, and temporally situated.
Driver Feedback and Selection
Pricing is an intervention as well as a prediction
Telematics insurance can alter the behavior used to calculate its own risk classification. If speeding raises next year's premium, some insureds may reduce speeding; some may drive less; others may stop transmitting data, change the phone used for tracking, or leave the program. A pricing equation estimated on drivers before an incentive takes effect need not remain invariant after its introduction. These responses create a feedback loop in exposure, selection, and measured event rates.
The observable premium reduction of a self-selected program member cannot establish that the program prevented crashes. People who anticipate favorable scores may enroll preferentially, while privacy-sensitive or higher-mileage customers may decline. A reduction in reported claim rate among participants may be caused by different policyholder composition, altered actual driving, changes in claim submission, or a combination. The insurer should distinguish the estimand for a predictive model—future insured cost conditional on observed features—from the causal estimand for an incentive intervention—how outcomes would differ under a specified alternative program.
Selection into observation
Let A_i indicate enrollment in a telematics program and let Y_i denote future loss. A dataset restricted to A_i=1 estimates E[Y_i|X_i,T_i,A_i=1], while the insurer's intended full-market target may require E[Y_i|X_i] for all eligible policyholders. The difference matters when selection depends on expected driving risk, privacy preferences, device access, or socioeconomic circumstances. Peiris, Jeong, Kim, and Lee (2024) directly study integration of traditional and telematics data, emphasizing scarcity of telematics observations and favorable selection. Their analysis underscores that the distribution of observed sensor data is itself a substantive part of the prediction problem.
A valid reweighting exercise requires overlap: comparable nonparticipants must have some observable counterparts among participants after conditioning on measured covariates. If all low-connectivity drivers or a particular vehicle class are absent from the telematics sample, weighting cannot reconstruct their missing behavioral measurements without additional assumptions or new data. A prudent strategy preserves a conventional actuarial baseline, records acceptance and dropout, and evaluates score effects on the exact population to which the price will apply.
Observed dropout warrants separate analysis. Some customers leave after a poor score; others lose eligibility after buying a different vehicle; some experience device malfunction or a change in employment route. Administrative events should be coded and dated. Censoring models can address some systematic attrition, but their assumptions must be declared and challenged by sensitivity analysis.
Experimental evidence of behavior change
Randomized interventions offer more direct evidence about feedback and incentives. Stevenson and colleagues (2021) conducted a telematics feedback randomized trial with a 28-week observation framework and 174 drivers completing at least one trip. The trial found no statistically significant improvement for its primary individual driving-behavior outcomes compared with controls; a secondary composite score improved for the feedback-plus-incentive group. It demonstrates that apparent theoretical appeal does not guarantee detectable change in every outcome or sample.
Delgado and colleagues (2024) studied feedback and financial incentives among enrolled US usage-based auto insurance customers. In a randomized trial with 2,020 completing participants, combinations of comparison feedback and money reduced handheld phone use relative to control; feedback alone and a standard delayed incentive alone did not show corresponding significant reductions. The outcomes were smartphone-use measures, not a randomized estimate of collision claims reduction. This distinction limits the permitted inference: behavioral exposure changed within the reported experimental context, whereas the ultimate insured claims effect remains a separate empirical question.
A later national experiment by Ebert and colleagues, published in 2026, studied simulated usage-based insurance incentives with 1,449 recruited drivers after a six-week baseline and randomized intervention assignments. The study reported improvements in its composite safety score and several measured behaviors across incentive-and-feedback conditions, with no significant superiority among the three active intervention types after multiple-comparison adjustment. Its simulated insurance design and selected participant recruitment differ from actual binding renewal tariffs, so the interpretation should remain attached to that intervention.
Together, the experiments support qualified causal claims about measured behaviors under defined incentive designs. They also reveal that the design of the comparison, reward timing, financial salience, and outcomes changes measured effects. They do not justify treating every braking score as a proven crash-prevention mechanism.
Feedback architecture and possible behavioral substitution
Feedback should identify understandable actions under the driver's reasonable control: avoiding handheld phone use, maintaining appropriate speeds, or reducing avoidable abrupt maneuvers. A response dashboard should provide context and uncertainty, including opportunities for correction when roadworks, emergency braking, wrong speed limits, or vehicle passenger status affect the score. Behavioral changes may also substitute across margins: a driver might reduce recorded hard braking by following traffic less closely, which is beneficial, or by hesitating to brake sharply when necessary, which is dangerous. An event-rate target should therefore be validated against meaningful safety or loss outcomes rather than rewarded unconditionally.
The safest product logic is to separate immediate coaching from delayed actuarial pricing. Feedback may be more frequent and provisional, while financial consequences should depend on adequate exposure, complete capture, contractually specified windows, and verified risk relationships. Communication must never imply that a consumer should avoid necessary emergency actions to protect a numerical score.
Coverage decisions and market segmentation
Telematics may influence consumer choice of deductible, optional coverage, or carrier in addition to driving. Soleymanian, Weinberg, and Zhu (online 2024; issue 2025) examine coverage choices with sensor data in a sample of 135,540 customers, a reminder that participation affects product purchasing and portfolio composition. Those outcomes are economically relevant without proving that the mere possession of a device causes all observed policy changes. An insurer should model whether favorable-score discounts attract more low-risk insureds, whether surcharges drive high-risk insureds out, and how resulting coverage selection changes the loss distribution.
From a welfare standpoint, the effects should be judged along several axes: risk prevention, accuracy of cost allocation, affordability, privacy cost, consumer comprehension, administrative expense, and the ability of the insurance pool to continue covering diverse risks. A program that predicts loss well may still have a costly measurement architecture or a damaging effect on coverage access. This judgment depends on empirically documented outcomes and the jurisdiction's insurance policy.
Rating Accuracy and Fairness
Competing meanings of accurate pricing
There are at least four relevant accuracy questions. A model may discriminate between higher- and lower-expected claim frequency, be calibrated to actual loss levels, remain stable under new roads or devices, and estimate the correct prospective cost after coverage and claims development. Rank ordering alone is not rate adequacy. A score may improve an AUC-like discrimination measure for an accident indicator while systematically underestimating annual liability costs for a subgroup. Conversely, a model could be well calibrated in aggregate because over- and underpricing cancel, while performing poorly for rare high-cost claims.
A meaningful comparison maintains an identical evaluation population, claim definition, exposure period, development horizon, and training cutoff. Compare a conventional baseline, a mileage-augmented baseline, a behavioral-augmented model, and a context-sensitive model. Assess held-out log likelihood or Poisson deviance for counts, proper scoring for probabilities, error in total predicted covered cost, calibration by score quantile, stability by cohort, and realized-to-expected cost by important strata. The regularized telematics pricing study by Usman and colleagues (2024) shows one structured approach to feature selection and frequency modeling; any gains claimed for a local tariff still require an independent prospective test.
Prediction versus value of a feature
An event rate may correlate with claims because it measures unsafe driving, because it identifies dense urban traffic, because it reflects camera or sensor sensitivity, or because it indirectly captures geography and working hours. Marginal association is insufficient to determine whether removing or penalizing that variable improves safety or fairness. In insurance practice, predictive relevance can be assessed conditional on legally permissible baseline factors, road context, and exposure; decision relevance additionally includes whether the price signal is actionable, stable, and proportionate.
For example, harsh braking may rise on congested urban routes. If the model applies a universal braking surcharge without conditioning on congestion or route geometry, the tariff can penalize drivers for lawful and prudent reactions to external hazards. Conversely, omitting strong context can leave the insurer exposed to genuinely different crash frequencies across roads. A scientific audit should therefore separate underlying hazard exposure from behaviors within that context, then assess whether the consumer can reasonably understand and influence the priced dimension.
A formal model-comparison example
Suppose a held-out set has two exposure groups, each with 10,000 verified miles. A hypothetical conventional model forecasts 0.20 and 0.20 claims; a telematics model forecasts 0.10 and 0.30; the observed counts are 0 and 1. The telematics model assigns more expected claims to the group with the observed event, but these two outcomes are far too sparse to establish calibration or fairness, and an observed zero is consistent with many underlying intensities. This example illustrates why rank accuracy, aggregate adequacy, and interval uncertainty must be evaluated over materially larger and longer exposure bases. It is not an observed study or a model-validation result.
For a larger portfolio, one may compute, within score band g, O_g/E_g = (Σ_i observed losses_i)/(Σ_i expected losses_i) over a consistent period. Ratios near one are useful only when the counts, exposure, loss development, uncertainty, and aggregation are reported. Confidence intervals, bootstrap uncertainty, claim clustering, inflation adjustment, and rare-tail risk matter. For pricing, the same result should be displayed separately by coverage rather than treating a bodily injury shock and a minor vehicle repair as interchangeable data points.
Multiple dimensions of fairness
Actuarial fairness is often used to describe consistency between expected cost and the premium allocation after allowable loadings. Procedural fairness concerns explainability, correction, appeal, consent or alternative lawful grounds for data use, and the consumer's ability to participate without opaque disadvantage. Distributional fairness concerns where the benefits and costs fall across income, age, place, disability, work schedule, phone access, and vehicle age. A specific statutory concept of unfair discrimination may impose different obligations across jurisdictions. These concepts can interact, and an evaluation should identify the exact claim being tested.
Risk factors tied to shift work or night driving may increase predictive power while concentrating higher rates on workers with constrained schedules. Location and smartphone-derived behavior can act as proxies for characteristics unrelated to voluntary risk taking. A device can also measure the same physical event differently across cars or operating systems. Consequently, matching price to historical claim cost is not the sole required question when an insurer argues that the system is equitable.
The ASOP 12 risk classification guidance provides an actuarial frame for evaluating classification. It does not settle every legal or ethical question. A local fairness examination should document the purpose of each variable, relation to covered losses, statutory restrictions, measured error, subgroup calibration, missing-data behavior, alternative factors, and consequences of limiting or excluding a variable. Where legal restrictions prevent using a sensitive characteristic in pricing, suitable governance must distinguish lawful auditing from impermissible use in the decision itself.
Privacy and automated decisions
European personal-data processing is governed, where applicable, by the General Data Protection Regulation. Its rules on lawful processing, transparency, purpose limitation, minimization, data-subject rights, and security affect the collection and deployment of identifiable vehicle-location or behavior data. Article 22 addresses solely automated decisions producing legal or similarly significant effects, subject to its detailed conditions and exceptions; it should not be reduced to a blanket prohibition on every automated insurance quote. The European Data Protection Board's connected-vehicle processing guidelines, adopted in March 2021, explain data protection issues specific to mobility and connected vehicles.
The EU AI Act identifies certain systems used for risk assessment and pricing of natural persons in life and health insurance as high-risk applications under Annex III; it does not automatically classify every motor telematics tariff as that particular Annex III insurance use case. The consolidated timetable was affected by Regulation (EU) 2026/1744, including postponement of the Annex III high-risk provisions to December 2, 2027. Other AI Act obligations and generally applicable laws must be assessed according to actual system functionality, applicable dates, and jurisdiction; exemption from one enumerated insurance category is not exemption from privacy, consumer, discrimination, or sectoral obligations.
EIOPA's August 2025 Opinion on AI governance interprets existing insurance-sector governance expectations, emphasizing risk-proportionate controls, fairness, data governance, explainability, recordkeeping, oversight, security, and robustness. EIOPA states that the Opinion itself creates no new requirements. Its guidance is a relevant supervisory interpretation rather than a replacement for legal analysis of binding instruments and national implementation.
The price-stability and solidarity problem
More granular predictions can produce larger dispersion of prices as drivers are sorted into narrower categories. Better within-portfolio calibration might therefore coexist with reduced affordability for individuals in hazardous environments they cannot change. EIOPA's discussion of insurance mutualization and personalization expresses the institutional tension: individualized accuracy can improve allocation while fragmenting the risk pool. A sophisticated appraisal should measure the resulting movement in expected claim costs, take-up, lapse, uninsured driving exposure, and cross-subsidy, with separate attention to compulsory motor liability protections.
The analytical conclusion is specific. A risk score is neither a complete premium nor a sufficient certificate of fairness. The demonstrated scientific contribution of telematics lies in improved observation and conditional prediction in certain contexts; operational acceptability requires the additional examination of legal permission, data rights, model drift, and measurable distributional consequences.
Comparative Evidence, Data Uncertainty, and Validation
The evidence hierarchy
This monograph distinguishes four broad forms of evidence. Official legal instruments establish rules within their jurisdictions and dates. Institutional insurance and actuarial publications establish supervisory expectations or professional methods within their stated scope. Peer-reviewed observational studies establish model performance and associations in their particular portfolios. Randomized interventions provide evidence on specified causal behavioral outcomes for their recruited participants and intervention periods. None of these classes alone answers every insurance-pricing question.
The work of Verbelen, Antonio, and Claeskens (2018), Ayuso, Guillén, and Nielsen (2019), Guillén and colleagues (2019, 2021, 2024), Peiris and colleagues (2024), Usman and colleagues (2024), and Yanez and colleagues (2025) spans conventional count models, behavior variables, sample integration, weekly indicators, and dynamic scoring. Their studies collectively support a research program rather than a single universally validated tariff. The published near-miss telematics paper, for example, connects events such as braking, acceleration, and phone use to a pricing illustration; an event is a potential hazard signal and cannot be treated as a settled insured claim.
Dataset construction and observation units
An auditable dataset must reconcile the policy ledger, vehicle or driver identity, coverage terms, raw telematics, feature production, incident records, claims transactions, and time alignment. The principal unit should be declared: vehicle-week, driver-month, policy-year, or trip. Different units change clustering, exposure weighting, and interpretation. A weekly panel cannot be analyzed as independent observations merely because each row has a different date; repeated behavior and shared latent risk induce dependence.
A minimum research record should describe enrollment eligibility, telematics participation rate, active device-days, trip-detection rules, sensor revisions, mileage validation, event thresholds, claims occurrence and report dates, claim development, earned exposure, cancellation, and coverage exclusions. For each field, researchers should know whether it existed at the prediction date. The Spanish weekly data study describes its time period and recording interruption, making explicit the possibility that missing weeks reflect system failure rather than low risk. A tariff should treat that event differently from a driver's decision not to drive.
Location privacy also influences research reproducibility. Coordinates may be sensitive and difficult to release, while aggregation may remove road-context information central to the score. Synthetic data can test software and documentation, but does not independently validate fitted effects. A credible scientific record states exactly which data are unavailable and which numeric conclusions can be rederived from public tables.
Cross-validation must follow time and entity
Randomly splitting vehicle-weeks can place early weeks from a driver into training and nearby weeks from the same driver into testing. This measures short-term interpolation more than prediction for unseen policyholders or a future season. At least three validation views are useful: a strictly later calendar-period holdout; an entity-separated holdout for new customers; and a transport test on a distinct road, device, or geographic population if available. All feature engineering should be refitted within training folds, and thresholds chosen without inspecting the final holdout.
The insurer should evaluate claim frequency with appropriate exposure offsets; paid or incurred claim severity with development adjustments; and pure premium with separately reported tail sensitivity. A model trained on settled claims only may underrepresent severe recent injury claims that settle slowly. Validation should be repeated under alternative claim definitions and different evaluation horizons. Observed accident risk and the insured cost of the accident are separate targets even when they share a data record.
Missingness, selection, and temporal stability
Missing data may reflect sensor outages, user nonparticipation, device incompatibility, battery constraints, damaged hardware, no travel, and extraordinary events. Missing-at-random approaches require meaningful assumptions about what conditioning data explain the missingness. When customers can disable a program precisely during risky trips, observed event rates may understate risk. No statistical imputation method can fully resolve data that are absent in a behavior-dependent way without a credible identification strategy, external comparison information, or sensitivity bounds.
A regulator or researcher should therefore ask how premiums change under alternative plausible missing-data scenarios: counting only verified distance, replacing unverifiable features with neutral credibility-weighted values, reclassifying partial observations, and excluding highly incomplete periods. The goal is to quantify decision sensitivity, not manufacture precision. The data integration analysis of Peiris et al. is pertinent precisely because observing telematics on a selected subset affects inference for the total book.
Score stability also depends on new vehicle driver-assistance systems, changing road networks, camera or phone software updates, repair inflation, and consumer responses to published rating rules. A threshold for harsh braking calibrated on one vehicle model can drift when braking systems change. Documented recalibration, model versioning, back-testing, and controlled implementation are therefore necessary to keep a pricing system consistent over time.
Comparative jurisdictional assessment
US programs operate under state-specific insurance regulation and privacy regimes, with actuarial standards informing professional work. EU-market programs additionally encounter the GDPR where applicable and EU and national rules governing insurance distribution, consumer rights, and specific automated systems. Other markets require independent local research; neither a US industry summary nor European guidance can establish legal validity in Japan, Australia, Brazil, India, or every other motor market.
A comparative matrix should distinguish at least the following columns: allowed data collection and legal basis; permitted rating variables and rate-filing evidence; policy notice and adjustment frequency; access to alternative nontelematics products; retention and deletion rights; automated decision safeguards; dispute resolution; and treatment of low-connectivity customers. The same mathematical model may need materially different permissible inputs, explanations, and implementation controls across jurisdictions.
Scholarly limits and reproducibility
No raw study-level telematics records were accessed for this monograph, no proprietary insurer data were modeled, and no independent replication of published numerical coefficients was conducted. The reported sizes of selected studies and the cited experimental results are attributed to their publications. The original calculations are algebraic demonstrations, not new estimates. The bibliography identifies directly reviewed public pages, abstracts, or publisher records; where full proprietary datasets remain closed, the exact boundary of reproducibility is recorded. These limitations are substantive conditions on the scientific inferences and are carried into the evidence ledger.
Integrated Results and Open Research Questions
Result: a score has four necessary reference conditions
The first integrated result is definitional and operational. A telematics risk score acquires actuarial meaning only when the insured coverage, observation window, measurement apparatus, and reference population are stated. Change any one of these and the interpretation may change. A speeding feature measured during one week can be associated with future own-damage frequency, while a second model may estimate liability claims over a year. A change in device sensitivity can shift event counts without any change in actual driving. A percentile rank can shift when portfolio composition changes. The same numeral is therefore insufficient documentation for an insurance risk estimate.
This follows from the mathematical structure of conditional expectation and the observed dependency of empirical results on sample and measurement. It is an analytical synthesis of existing actuarial and statistical reasoning, not a newly named theory.
Result: tariff construction requires a coverage-based bridge
The second integrated result concerns the path from scoring to money. Features yield frequency or severity predictions only through a statistical model; predicted insured costs pass through contractual terms, exposure definitions, claim development, and permissible expenses; and the tariff then passes through law, approved classification, and operational rules. The ASOP 53 prospective cost framework and published telematics ratemaking research support keeping these stages explicit. A score percentile should never be presented as an independent percentage of premium without stating the conversion rule.
Result: causal safety effects require intervention evidence
The third result is that pricing accuracy and behavior modification are separate questions. Randomized studies identify effects of feedback-and-incentive designs on selected measured behaviors under specified protocols. The 2024 phone-use trial and 2026 national field experiment establish more than a correlation for their actual interventions, while neither supplies a general causal estimate of the effect of all telematics premiums on insured claim severity or market-wide accident totals. Observational results are indispensable for prediction but must be interpreted with self-selection and behavioral adaptation in mind.
Result: fairness is a portfolio and procedural outcome
The fourth result concerns the distribution of insurance protection. The consequences of a risk score depend on where the price signal falls, whether the factor is controllable, the fairness and accuracy of measurement, transparency, the availability of alternative policies, and the behavior of consumers and insurers. A rate can be calibrated within an enrolled sample and still create unfavorable outcomes for customers excluded by device availability or priced according to poorly contextualized work schedules. The combination of connected-vehicle data guidance and EIOPA governance expectations reinforces the need for controls beyond optimizing a single predictive metric.
Questions suitable for new empirical research
The most important next investigations require data or experiments rather than additional labels. Which independent contribution do braking and acceleration indicators make to coverage-specific future cost after conditioning on traffic density, weather, speed, mileage, vehicle systems, and driver adaptation? How do reported claim patterns change when a customer knows that a particular variable affects renewal price? Can partial randomization of discount or feedback timing identify changes in collisions and insured losses over several years, while retaining the protection of the policyholder? How much of model improvement survives new vehicle and device generations? Which mechanisms for appeal, score freezing, and alternative coverage mitigate inequitable exclusion without materially undermining premium adequacy?
A rigorous empirical program would preregister relevant outcomes, protect privacy, preserve exposure definitions, and publish sufficiently detailed model specifications and anonymized aggregate diagnostics for independent assessment. Multi-jurisdictional work should report legal and contract differences instead of interpreting them as nuisance variation.
Practical scientific conclusions for model governance
The following analytical implications arise from the evidence reviewed. Keep claim frequency, severity, mileage, behavior, and context in distinguishable components; publish score and tariff version identifiers; validate on future periods and customer cohorts; explicitly model unobserved mileage and selection; demonstrate stable calibration by coverage and subgroup; use behavioral trials when claiming safety improvements; audit privacy and legally permissible variables at the point of deployment; and disclose how score revisions affect premiums and consumer rights. These implications are applications of established statistical, actuarial, and regulatory work, with no invented proprietary data or unsupported declaration of a new doctrinal system.
Conclusion
Telematics-based motor insurance pricing is the application of measured exposure and behavior to the conditional estimation of insured motor losses and the construction of legally permissible premiums. The literature documents material information in distance traveled, speed-related signals, road context, and some near-miss indicators. It also documents the difficulties of selected participation, zero-heavy claims, short observation periods, overlapping and missing records, and translating a risk signal into an actuarially adequate and understandable tariff.
The central scientific conclusion is that a risk score must be evaluated as a complete, dated information product within an insurance contract and a defined decision. It estimates a conditional relationship in a specified observed population; a consumer premium additionally reflects coverage, loads, legal constraints, and institutional choices. The distinction has practical consequences for validation, explanation, and fairness. Feedback and financial incentives can change some measured behaviors, as the randomized studies show, yet actual claims prevention and total social benefits require their own evaluation. Greater individualization can improve risk matching while changing the distribution and accessibility of insurance protection.
Future progress depends on prospective, coverage-specific data; transparent observation and versioning; experimentally credible assessment of behavioral change; and regulatory governance that reconciles actuarial evidence with customer rights. On the evidence examined, these are substantiated research and implementation directions. No original empirical study, universal premium formula, novel named theory, or independent external peer-review decision is claimed.
References
Actuarial Standards Board. (2005; effective 2006; revised presentation 2011). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). Official ASOP 12.
Actuarial Standards Board. (2017; effective 2018). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. Official ASOP 53.
Ayuso, M., Guillén, M., & Nielsen, J. P. (2019). Improving automobile insurance ratemaking using telematics: Incorporating mileage and driver behaviour data. Transportation, 46, 735–752. https://doi.org/10.1007/s11116-018-9890-7.
Delgado, M. K., Ebert, J. P., Xiong, R. A., et al. (2024). Feedback and financial incentives for reducing cell phone use while driving: A randomized clinical trial. JAMA Network Open, 7(7), e2420218. https://doi.org/10.1001/jamanetworkopen.2024.20218.
Ebert, J. P., McDonald, C. C., Xiong, R. A., et al. (2026). A national randomized field experiment testing the impact of simulated usage-based insurance programs on driver safety. Accident Analysis & Prevention, 225, 108331. https://doi.org/10.1016/j.aap.2025.108331.
European Data Protection Board. (2021, March 9). Guidelines 01/2020 on processing personal data in the context of connected vehicles and mobility related applications (final version). EDPB connected-vehicle guidelines.
European Insurance and Occupational Pensions Authority. (2025, August 6). Opinion on Artificial Intelligence governance and risk management. EIOPA Opinion.
European Insurance and Occupational Pensions Authority. (2025, September 16). Supervision of AI: Finding the right balance. EIOPA supervisory discussion.
European Parliament and Council. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union. GDPR official text.
European Parliament and Council. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. Official Journal of the European Union. EU AI Act.
European Parliament and Council. (2026). Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence. Official Journal of the European Union, 2026/1744. AI Act amending regulation.
Guillén, M., Nielsen, J. P., Ayuso, M., & Pérez-Marín, A. M. (2019). The use of telematics devices to improve automobile insurance rates. Risk Analysis, 39(3), 662–672. https://doi.org/10.1111/risa.13172.
Guillén, M., Nielsen, J. P., & Pérez-Marín, A. M. (2021). Near-miss telematics in motor insurance. Journal of Risk and Insurance, 88(3), 569–589. https://doi.org/10.1111/jori.12340.
Guillén, M., Pérez-Marín, A. M., & Nielsen, J. P. (2024). Pricing weekly motor insurance drivers' with behavioral and contextual telematics data. Heliyon, 10(16), e36501. https://doi.org/10.1016/j.heliyon.2024.e36501.
National Association of Insurance Commissioners. (n.d.; accessed November 15, 2026). Telematics. NAIC insurance topic.
Peiris, H., Jeong, H., Kim, J.-K., & Lee, H. (2024). Integration of traditional and telematics data for efficient insurance claims prediction. ASTIN Bulletin, 54, 263–279. https://doi.org/10.1017/asb.2024.6.
Soleymanian, M., Weinberg, C. B., & Zhu, T. (2025; first published online 2024). Insurtech, sensor data, and changes in customers' coverage choices: Evidence from usage-based automobile insurance. Journal of Risk and Insurance, 92(1), 227–256. https://doi.org/10.1111/jori.12490.
Stevenson, M., Harris, A., Wijnands, J. S., & Mortimer, D. (2021). The effect of telematic based feedback and financial incentives on driving behaviour: A randomised trial. Accident Analysis & Prevention, 159, 106278. https://doi.org/10.1016/j.aap.2021.106278.
Usman, F., Chan, J. S. K., Makov, U. E., Wang, Y., & Dong, A. X. D. (2024). Claim prediction and premium pricing for telematics auto insurance data using Poisson regression with Lasso regularisation. Risks, 12(9), 137. https://doi.org/10.3390/risks12090137.
Verbelen, R., Antonio, K., & Claeskens, G. (2018). Unravelling the predictive power of telematics data in car insurance pricing. Journal of the Royal Statistical Society: Series C (Applied Statistics), 67(5), 1275–1304. https://doi.org/10.1111/rssc.12283.
Yanez, J. S., Guillén, M., & Nielsen, J. P. (2025). Weekly dynamic motor insurance ratemaking with a telematics signals bonus-malus score. ASTIN Bulletin, 55(1), 1–28. https://doi.org/10.1017/asb.2024.30.