RESEARCH DOSSIER
Motor Insurance Claim Frequency and Severity Modeling
Abstract. Motor insurance claim costs emerge from the interaction of covered exposure, claim incidence, claim development, conditional payment amounts, legal responsibility, and contractual limits. This monograph develops a comparative and statistical account of frequency and severity modeling for third-party bodily injury, third-party property damage, first-party collision, and comprehensive coverage. It reconstructs the relevant observation units and exposure denominators; derives the conditions under which separate count and severity regressions generate an unbiased expected aggregate loss; and examines Poisson, negative binomial, hurdle, zero-inflated, gamma, lognormal, compound Poisson–gamma, and dependent frequency–severity specifications. It distinguishes claims that are reported, incurred, settled, and paid, and separates prediction for prospective underwriting from retrospective development of outstanding liabilities. The documentary evidence combines original actuarial research, professional standards, insurance supervisory publications, accident-data definitions, and a transparent recalculation of published U.S. paid-claim frequency and severity by coverage in 2019 and 2024. The recalculation finds that paid loss per insured car-year rose under all four presented coverage groups even though the observed frequency declined in three, demonstrating the importance of disaggregating calendar-period change. Statistical validation is considered across calibration of claim counts, claim size distributions, aggregate costs, subgroup composition, heavy tails, temporal drift, telematics selection, and consumer fairness. The synthesis establishes a practical hierarchy: define the insured outcome and its denominator first; fit the probability and magnitude of claims with allowance for dependence; validate expected cost and tail behavior against prospective cohorts; and document the legal and governance conditions under which estimates can be deployed. No original causal estimate, newly proposed mathematical theory, or unobserved portfolio is asserted.
Keywords: Motor insurance; automobile insurance; claim frequency; claim severity; compound distributions; generalized linear models; Poisson regression; negative binomial regression; Tweedie model; dependent frequency and severity; pure premium; bodily injury; collision; telematics; calibration; predictive uncertainty; actuarial pricing.
Introduction
A motor insurance claim is produced by a chain of conditions. A vehicle or driver must be exposed to a covered peril during an operative period; an event or legally cognizable loss must arise; the policy or compulsory regime must respond; the loss must become known and assessed; and its financial consequences must be recognized, paid, or reserved. A statistical record taken at any point in that chain measures a different quantity. Accident frequency describes crashes or injuries; reported claim frequency describes notifications; closed-claim frequency concerns settlement; and paid claim frequency concerns payments. Each may be a valid object of analysis when its observation rule and purpose are explicit.
Classical actuarial ratemaking separates how often claims occur from how costly they are. The generalized linear model formulation of Nelder and Wedderburn (1972), its motor insurance application in Dionne and Vanasse (1989), and the practical treatment by Ohlsson and Johansson (2010) provide a tractable statistical foundation. These methods are useful because exposure, observed heterogeneity, sampling variability, and model uncertainty can be examined separately and then reconciled in a prospective loss estimate.
The central analytical question is the expected cost per unit of covered exposure. An insurer also needs the shape of the aggregate loss distribution, adequate treatment of high-severity cases, coherent measurement across insurance coverages, and controlled translation of technical risk prices into actual charged premiums. The final price additionally involves expenses, capital, reinsurance, taxes, discounts, cross-subsidies, market competition, and applicable law. The underwriting model is therefore one component of an institutional pricing process.
This research uses critical synthesis and worked arithmetic derived from identified public sources. It does not claim access to insurer-level claim microdata or fit a new empirical model. Mathematical identities are established explicitly; empirical numbers are labeled by their original source, coverage, period, and limitations. A single cross-coverage U.S. table is used as a transparent descriptive illustration, not a basis for universal international estimates.
Conceptual, Historical, and Primary-Source Foundations
The insurance observation unit
A policy record is not automatically a homogeneous risk unit. One contract can cover multiple vehicles, drivers, coverages, insured locations, and deductibles; coverage can change during the term. The appropriate unit for annualized private-passenger motor pricing is often a vehicle–coverage–period, optionally split into intervals whenever important covariates, coverage terms, or exposure conditions change. A covered vehicle-year conventionally corresponds to 365 insured days, while commercial fleets may require additional dimensions such as driver, route, vehicle weight, business activity, and insured distance.
Let exposure for record i be e_i in vehicle-years and N_i denote its count of claims under a specified coverage and counting rule. The annualized frequency rate is estimated as the total claim count divided by total exposure. If the record contains 180 insured days, exposure is 180/365 under the simplest time basis. An insured mileage denominator and an insured-time denominator answer different questions. Mileage may be especially relevant to collision frequency, but its use changes the interpretation of the intercept, the effect of stationary hazards, and the relation of exposure to vehicle use.
The ISO/Verisk paid-claim series defines claim frequency per 100 earned car-years, making the denominator sufficiently concrete for descriptive rate comparisons. The source presents paid claims, coverage-specific exclusions, and a deductible convention for physical damage. A claim count divided by active policies at year-end would answer a different question and can be biased by policy cancellations, vehicle additions, and differing coverage terms. An exposure denominator should be reconciled with every claim numerically, temporally, and contractually before estimation.
Events, claim units, persons, and coverage
One crash may generate a third-party bodily injury claim, multiple third-party property damage claims, a first-party collision claim, medical benefits, legal expense, and payments under separate policies. An injury claimant is not necessarily equivalent to an accident, claim file, payment transaction, or policyholder. The counting rule must specify whether one liability occurrence creates one claim, one affected person creates one claim, or one coverage account creates one claim. It must also prevent duplicate transactions from being treated as new events.
The NAIC auto coverage guide distinguishes liability protection from physical damage coverages. Bodily injury liability responds to legally relevant injury to others under policy terms; property damage liability concerns loss imposed on a third party; collision addresses insured vehicle damage; and comprehensive generally addresses specified noncollision perils. Deductibles, limits, subrogation rights, no-fault structures, and compulsory insurance conditions change both the recorded claims and their monetary amounts.
The EU Motor Insurance Directive establishes an insurance framework for civil liability arising from motor vehicle use, and the 2021 amending directive changes parts of that regime. Actual claim eligibility and compensation administration depend on applicable transposed law and national institutions. A severity model calibrated on one jurisdiction's bodily injury payments cannot be exported mechanically into a jurisdiction with different liability allocation, injury benefit structures, or settlement practices.
The separation of road injury and insured claim datasets
The World Health Organization's 2023 global road safety report concerns population road trauma and mortality. The U.S. Crash Report Sampling System estimates police-reported crashes, while the Fatality Analysis Reporting System addresses fatal U.S. traffic crashes. These are essential external indicators of road risk, yet their denominators, inclusion thresholds, observation mechanisms, and time bases differ from insurer paid and incurred claim records.
A change in road fatalities may affect the distribution of severe injuries without translating proportionally into automobile insurer claim counts. Coverage uptake, liability law, legal time limits, claims propensity, repair costs, compensation norms, case management, and insurer market shares intervene. Matching public crash counts to insurance claim costs therefore requires a defined population and a justified linkage; simply dividing insurance paid claims by police accidents produces an unstable statistic.
From classical ratemaking to statistical learning
The core transition in automobile ratemaking is from aggregate tariff cells to explanatory regressions that use several risk factors while accounting for uneven exposure. Dionne and Vanasse (1989) extended automobile claim count regression using a negative binomial specification and considered individual experience rating. The CAS insurance rating monograph provides contemporary practical treatment of GLM construction, diagnostics, and selection. In the newer literature, Wüthrich and Merz (2023) connect conventional actuarial models with broader statistical learning and predictive evaluation.
This development enlarges the available function classes without changing the fundamental obligation to define the target. A high-capacity prediction model cannot repair inconsistent coverage definitions, missing inflation indices, unknown claim maturity, omitted exposure, or labels that inadvertently contain information obtained after the intended pricing decision. The value of more predictors must be tested against temporal stability, interpretability, deployment permissions, and marginal improvement in cost prediction.
Claims Counting Models
Poisson regression and the exposure offset
For a homogeneous interval, let N_i conditional on covariates x_i follow a Poisson distribution with expectation and variance equal to lambda_i. A typical motor frequency GLM is
The offset log e_i has a coefficient fixed at one, so expected counts scale with exposure time while annual claim rate depends on x_i. This scaling is an empirical assumption about the selected exposure metric. If very low-mileage vehicles have material stationary peril exposure, a mileage-only offset may be unsuitable for comprehensive insurance. If trip length alters the probability of an accident differently from insured time, a fixed unit elasticity in mileage should be tested.
The Poisson model gives the zero-claim probability exp(−lambda_i) and a complete count likelihood. An empirical model can compare observed and predicted counts within exposure bands, risk cells, calendar periods, and coverage categories. A single sample-wide average frequency may be correct even when every important subgroup has systematic errors. Independent validation data should preserve the original exposure records, not normalize every observation to a synthetic one-year contract before the count likelihood is evaluated.
Heterogeneity, negative binomial models, and credibility
Observed counts can be overdispersed relative to a Poisson model because of unobserved driver characteristics, recurring exposure conditions, persistent vehicle use, heterogeneous locations, and shared accidents. A common negative binomial mean–variance specification is
A Poisson–gamma mixture provides one interpretation of this variance formula. Neither a large estimated dispersion parameter nor improved fit, by itself, identifies the causal source of heterogeneity. The negative binomial motor model of Dionne and Vanasse is an early specific example of incorporating regression and individual experience in automobile rating.
An individual historical claim count can update a prediction through a credibility or random-effects structure, but the update requires a defensible relationship between past and future risk. Changes in drivers, insurer, vehicle, policy terms, geographic exposure, and claims reporting undermine unconditional stationarity. Claim-history information also requires a defined lookback, claims maturity, and legal permission for its use. A bonus–malus mechanism should be evaluated as a predictor and as a behavioral incentive, with attention to the possibility that smaller claims are suppressed by the incentive to preserve discounts.
Excess zeros and the hurdle distinction
Zero counts are natural when claim events are uncommon. An ordinary Poisson distribution with a low expected count can already explain many policies without claims. Lambert's zero-inflated regression adds a separate probability mass at zero alongside a Poisson count mechanism; the work's original application was manufacturing defects, so applicability to motor claims requires independent justification. The zero-inflated Poisson probability of zero is pi_i+(1−pi_i)exp(−lambda_i).
A hurdle model separately estimates whether any claim occurs and, among claim-positive records, the positive count distribution. The statistical meanings differ: a zero-inflated model allows zero counts from both components, whereas a hurdle model puts all zeros in the incidence component. Repeated observations on the same insured unit or a stable external predictor may help identify structural zero mechanisms. A bare count of zeros cannot establish that a class of insured vehicles has genuinely no possibility of claim.
For motor insurance, deductible-driven nonreporting, active but unobserved policy coverage, and data that exclude zero-payment settlements are alternatives to structural nonrisk. Their effects can be evaluated by contrasting reported claim files with paid claim datasets. The production choice between Poisson, negative binomial, hurdle, and zero-inflated models should follow predictive out-of-sample accuracy and coherent interpretation, not selection of the most complicated distribution.
Temporal clustering and correlated policies
Conditional independence between policy records may fail if several vehicles share a household, employer, neighborhood, storm, or road network. A weather event can simultaneously increase comprehensive counts; a multi-vehicle crash can generate several liability payments; a repair backlog can shift notifications across months. Clustered standard errors account for some estimation uncertainty but do not automatically correct biased mean predictions arising from omitted common variables.
Useful designs include policyholder or fleet random effects, geographic and period effects, hierarchical partial pooling, and event-level aggregation. The aggregation decision should reflect the question: a catastrophe exposure model needs accumulation across risks, whereas individual tariff estimation needs future expected cost per coverage unit. Independence assumptions must be recorded both within frequency likelihoods and between distinct insured units in portfolio loss simulations.
Bodily Injury and Physical Damage Severity
Defining the claim amount
Severity is generally the conditional amount per claim within an identified claim population. A paid claim amount can include or exclude allocated loss adjustment expenses; an incurred amount can include case reserves; a final ultimate amount estimates payment after development and reopening. The model must identify gross versus net of reinsurance, before versus after deductibles and salvage, claimant versus accident aggregation, and treatment of expenses. Changes in these conventions alter severity even when the economic damage from incidents has not changed.
Consider an underlying physical damage amount G, deductible d, and coverage limit L. A simplified covered insurer payment, without salvage and recoveries, is min(max(G−d,0),L). If records exclude claims that do not exceed d, the observed conditional paid severity refers to the truncated subset G>d. The resulting observed paid-claim frequency and average positive payment are both functions of d. A deductible change can lower recorded frequency and increase the observed average gross loss among the remaining claimants while reducing expected insurer payment.
Exposure must also be specific to the coverage. Some insured vehicles carry liability but decline collision, and comprehensive may be purchased under different conditions. Cross-coverage comparisons therefore cannot use the same exposure count without verifying coverage participation. The NAIC's coverage descriptions are useful for explaining contract categories; actual payment calculations require applicable contracts and law.
Positive severity families
A gamma GLM with logarithmic link represents a positive mean and variance that scales as the square of the mean, conditional on the chosen gamma parameterization. A lognormal model is suitable for multiplicative variability, but its conditional mean must be computed from both its log-location and log-variance, not simply exponentiated from a fitted log-severity predictor. Heavy-tailed alternatives, mixtures, generalized Pareto excesses above a justified threshold, and finite-limit censoring may be appropriate when sufficiently many high-value claims exist.
Bodily injury losses frequently involve sparse high-value claims, especially where future care, wage loss, pain and suffering, court costs, or long-run disability are material. Large individual payments can control aggregate severity despite representing a tiny fraction of claims. Property damage and collision severities more directly reflect parts, labor, vehicle values, repairability, salvage, and claims administration, though litigation and total losses can also generate pronounced tails.
The GDV's 2026 replacement-parts comparison provides an example of a repair-cost pressure documented for a selected set of replacement parts in Germany. It does not estimate total German motor claim severity, and the parts sampled are not the full insured vehicle portfolio. Repair inflation, vehicle age, and model mix must be separated empirically; the published comparison illustrates a plausible mechanism rather than a causal estimate for a given insurer.
Frequency times severity: exact scope of the identity
Let N be a nonnegative claim count and Y_j the payment for the jth claim. Aggregate loss is T=sum from j=1 to N of Y_j, with T=0 when N=0. If the individual conditional severity expectation remains constant at m(x) across claim counts and claim units given the modeled risk state x, then
This is an expected-value identity under specified conditional assumptions, not proof that realized claim counts and severities are statistically independent. More generally, write m(n,x)=E[T/N\mid N=n,x] for positive n. Then
The naive multiplication E[N|x] E[Y|positive claim,x] needs care because an ordinary claim-size sample weights policy records with many claims more heavily. When frequency and average payment are dependent, their separate unconditional predictions may give incorrect aggregate cost. Garrido, Genest, and Schulz (2016) explicitly model average severity conditional on claim count in a Canadian automobile dataset, offering a direct way to represent this dependence.
A worked dependence example
Suppose that one policy-year has probabilities 0.75, 0.20, and 0.05 of zero, one, and two claims, respectively. Conditional average cost per claim is $2,000 when N=1 and $4,000 when N=2. Expected frequency is 0.30 claims per policy-year. Expected aggregate cost is 0.20×$2,000+0.05×2×$4,000=$800. The average claim cost across expected claims is $800/0.30, or approximately $2,666.67.
If one instead uses an unweighted mean of the two conditional severity strata ($3,000) and multiplies by the frequency 0.30, the estimate is $900; the discrepancy comes from the weighting convention and conditional dependence, not from rounding. A separate model that correctly predicts mean severity for a randomly selected claim would recover $800 through 0.30×$2,666.67. The example demonstrates why explicit sampling and conditioning definitions precede every frequency–severity product.
Direct aggregate and compound distributions
A compound Poisson model specifies a count-generating process and positive individual claim amounts, from which aggregate loss can be simulated or analytically summarized. The compound Poisson–gamma Tweedie treatment of Jørgensen and de Souza (1994) permits an exact mass at zero and a continuous positive component under the familiar Tweedie power-variance range 1<p<2. In a Tweedie regression, the mean can be modeled directly, often with a log link, while exposure handling and dispersion have to agree with the observation unit.
A Tweedie likelihood is attractive for aggregate policy-level cost because zero and positive claim costs are modeled together. It need not be best for every objective: the analyst may want a separate estimate of claim count, case-level severity, conditional high-payment probability, or ultimate loss development, each of which a single aggregate mean does not identify. Simulations and tail quantiles are also sensitive to the assumed count and size dependence. Shi, Feng, and Ivantsova (2015) examined three-part and dependent specifications using a U.S. motor insurance application, reinforcing the practical value of testing dependence against simpler alternatives.
Monetary development, inflation, and ultimate cost
Claim amounts from successive accident years are not immediately comparable. A claim incurred in a recent quarter can remain unsettled; older claims have had longer to report, litigate, and develop. A severity model trained on paid amounts can understate ultimate cost for recent periods if their claims are immature. At minimum, the dataset needs loss occurrence date, report date, valuation date, payment date, claim-open status, transaction type, coverage, and the rule used to select the valuation snapshot.
Inflation is multidimensional. Repair labor, medical services, compensation awards, rental replacement, supply-chain constraints, and wage expectations can change at different rates. Changes in case mix and claim handling can produce apparent severity increases even after a general consumer-price deflator is applied. A prospective loss estimate may need separate indices by coverage and loss component, along with defensible assumptions for trend after the experience period.
Reserve methods and prospective pricing are related but answer different questions. The IFRS 17 accounting framework concerns recognition and measurement of insurance contracts and their liabilities; its accounting objective does not supply a frequency–severity tariff formula. Ultimate claim cost, prospective pure premium, and the insurance-contract liability should each have an explicitly specified evaluation date and scope.
Predictive Uncertainty and Calibration
Prediction targets and scoring rules
For each insured risk and time interval, the model might predict a count distribution, the probability of any reported claim, a positive severity distribution, expected aggregate covered payment, or a particular high-quantile loss. These targets are related but different. Count likelihood, frequency deviance, calibration of predicted zero probabilities, severity log likelihood or deviance, mean absolute error, and aggregate-loss errors each address aspects of fit; no one metric determines suitability for every insurance task.
A training comparison should include an exposure-adjusted intercept model, a structured GLM, a negative binomial or hurdle competitor where justified, a separate severity model, a Tweedie aggregate-cost model, and more flexible competitors if the sample supports them. Model choice should be prospective: estimate on earlier data and test on subsequent cohorts and on policies that would actually have qualified for a quote. Random splitting of repeated policies from the same insured across both train and test sets can produce overoptimistic results through leakage.
The actuarial learning reference of Wüthrich and Merz provides a useful methodological foundation for evaluating actuarial learning rather than equating fitting accuracy with prospective utility. Performance should be reported with uncertainty intervals, model versions, and exact data cohort definitions.
Calibration of incidence and amounts
For frequency, divide insured records into bins based on predicted annual rates and compare the total observed counts to total predicted counts using each record's actual exposure. Also inspect temporal curves, zero-claim proportions, multiple-claim rates, and within-group gradients. For severity, compare observed positive payments against predicted conditional amounts by coverage, injury type, settlement maturity, and monetary band, with uncertainty-aware diagnostics in sparse upper tails.
A model predicting correct aggregate loss cost can still have errors in the two components that offset in the total. That behavior matters for claims staffing, reinsurance, high-severity reserves, and pricing transparency. Conversely, a high Gini coefficient or ranking statistic can coexist with systematic overall underestimation. Premium adequacy requires calibrated monetary predictions, not ranking alone. The exact estimator used for calibration should be consistent with the claim count and amount sampling design.
Aggregate calibration can be expressed as sum of observed loss divided by sum of predicted loss in each disjoint test cohort. This ratio should be checked across coverages, risk tiers, regions, and time, with explicit reporting of low-exposure and low-claim uncertainty. A favorable point estimate on one held-out calendar year may fail when repair-price inflation, driving conditions, or legal settlement practices change.
Parameter, process, and model uncertainty
Parameter uncertainty arises from finite training samples; process uncertainty remains even if the conditional distribution were known; model uncertainty arises from plausible distributional or covariate specifications; and future regime uncertainty arises from changes to the operating environment. These sources should not be collapsed into a single standard error without an articulated stochastic construction.
For Poisson counts with mean lambda, the conditional process variance equals lambda; a negative binomial model adds overdispersion. For compound losses with independent identically distributed severities independent of N, expected aggregate loss is E[N]E[Y], while variance is E[N]Var(Y)+Var(N)(E[Y])². The variance formula changes when claim counts and severity are dependent, and the tail changes materially under common-event clustering. A conditional forecast interval cannot be presented as covering unmodeled structural shifts merely because it includes parameter uncertainty.
Bootstrap resampling needs the correct cluster: policyholders, fleets, geography, calendar periods, and shared events may be distinct units. An apparently large table with millions of claim transactions need not imply millions of independent risk observations. Shared severe events and highly concentrated bodily injury outcomes can sharply reduce effective tail sample size.
Telematics, mileage, and selection
Telematics supplies additional potential risk factors: insured distance, trip timing, acceleration, braking, road class, and device-use patterns. Verbelen, Antonio, and Claeskens (2018) studied a Belgian young-driver telematics program. That evidence is especially useful for testing additional predictive information relative to conventional rating variables, but its selected driver population and particular contract design limit extrapolation to the entire motor insurance market.
Selection into telematics may correlate with cautious driving, price sensitivity, driver age, technology access, or expected future mileage. Participation, device attrition, and missing trip data can alter observed predictors. A validation study should separate the value of mileage as an improved exposure denominator from additional driving-behavior features, and it should retain a nontelematics benchmark population with carefully stated comparability.
Interventions create a further identification challenge. A premium linked to driving behavior may alter the behavior that generated the original data. As the tariff is deployed, the model becomes part of the environment being predicted. Such feedback is relevant even when the model's original observational discrimination is excellent. Experimental or quasi-experimental designs would be required to infer a causal reduction in losses attributable to feedback or discounts.
Governance of risk classification and model use
The U.S. ASOP 12 risk classification standard concerns actuarial considerations in classifying risks in its applicable professional scope. ASOP 53 on prospective costs and ASOP 56 on modeling respectively concern estimating future property/casualty costs and the governance of actuarial model development, use, and evaluation. They do not substitute for applicable state law, a filed rating plan, or jurisdiction-specific consumer protections.
EIOPA's differential pricing supervisory statement distinguishes risk-related pricing from certain harmful renewal practices that take advantage of customer inertia. Its 2025 opinion on AI governance discusses model control and fair treatment in insurance activities. In the United States, the NAIC's adopted AI model bulletin is a model regulatory document and must be distinguished from a rule actually adopted and applied by a particular state. These documents establish separate professional, supervisory, and legal levels of analysis.
A governed motor pricing model has documented data provenance, variables, coverage scope, training horizon, access permissions, uncertainty, material limitations, owner, validation results, change-control procedure, and monitoring triggers. Classification fairness calls for examining legally protected attributes, proxies, differential error rates, affordability consequences, and the distinction between genuine expected loss and commercial propensity to renew. Predictive correlation supplies neither a legal permission nor a normative justification for every rating factor.
Comparative Evidence, Data Uncertainty, and Validation
Observed U.S. paid-claim patterns, 2019 and 2024
The Insurance Information Institute's ISO/Verisk table reports private passenger automobile paid claim frequency per 100 earned car-years and mean paid severity for liability bodily injury, liability property damage, collision, and comprehensive coverages. The 2019 and 2024 entries below reproduce those printed inputs, then compute paid cost per car-year mechanically as frequency divided by 100 multiplied by severity. Currency is nominal U.S. dollars and has not been adjusted for inflation. This is an analysis of the table's published cells, not an original insurer microdata estimate.
| Coverage | Frequency 2019 | Severity 2019 | Paid cost per car-year 2019 | Frequency 2024 | Severity 2024 | Paid cost per car-year 2024 | Change in implied cost |
|---|---|---|---|---|---|---|---|
| Bodily injury liability | 0.93 | $19,151 | $178.10 | 0.80 | $28,278 | $226.22 | +27.0% |
| Property damage liability | 3.14 | $4,326 | $135.84 | 2.50 | $6,770 | $169.25 | +24.6% |
| Collision | 5.00 | $3,828 | $191.40 | 4.16 | $5,489 | $228.34 | +19.3% |
| Comprehensive | 2.99 | $1,796 | $53.70 | 3.95 | $2,306 | $91.09 | +69.6% |
These observations have important boundaries. The liability bodily injury series excludes Massachusetts and most no-fault states; property damage liability excludes Massachusetts, Michigan, and New Jersey; the physical damage series excludes Massachusetts, Michigan, and Puerto Rico, and uses collision coverage with a $500 deductible. Comprehensive excludes wind and water losses but includes glass. The table is an aggregation of paid claim experience for each listed coverage and geographic sample. Adding the coverage-specific costs as if they described a single common population would be unjustified. The figure is an implied loss-cost identity for the stated data, not a quoted premium or a measure of underwriting profitability.
Exact multiplicative decomposition
The change in implied pure paid loss cost factors multiplicatively into frequency and severity movements, because loss cost is defined by their product in these aggregate tabular data. In bodily injury liability, frequency fell by approximately 14.0% between the two selected years, while paid severity rose by approximately 47.7%, giving a product increase of approximately 27.0%. Property damage liability frequency fell by 20.4%, severity rose by 56.5%, and implied cost rose by 24.6%. Collision frequency fell by 16.8%, severity rose by 43.4%, and implied cost rose by 19.3%. Comprehensive frequency rose by 32.1%, severity rose by 28.4%, and implied cost rose by 69.6%.
For any two comparable periods, let f_1, f_2 be published per-exposure frequencies and s_1, s_2 their corresponding mean positive severities. Then the ratio of implied costs satisfies
The log change is log(c_2/c_1)=log(f_2/f_1)+log(s_2/s_1), an exact decomposition when all inputs are positive. The decomposition contains no causal identification. Changes in claim types, insured vehicle mix, driving conditions, settlement time, policy deductibles, and source data definitions can affect both factors. The aggregate records do not let us conclude that vehicles became safer or repairs alone caused the cost changes.
Independent supervisory series and noncomparability
In its February 2026 auto database release, the NAIC reported the 2023 national combined average premium per issued vehicle at $1,438 and the 2022 national incurred average collision claim at $7,191, compared with $6,113 for 2021. These observations describe different years, denominators, and measurement types from the ISO/Verisk paid claim table. The $7,191 incurred claim amount should not be substituted for the $5,377 paid-claim severity shown for collision in the ISO series for 2022, because claim timing, expenses, sources, and population definitions differ.
This reconciliation illustrates why external verification must go beyond matching a monetary label. An individual model should specify whether its severity target is paid, case-incurred, reported, ultimate, or indemnity-only; whether expenses are included; and whether the exposure is earned car-years, issued vehicles, or another unit. Joining incompatible administrative datasets can create differences larger than the improvements attributed to advanced statistical methods.
Cross-jurisdictional validity
Road injury compensation, insurer obligations, compulsory liability limits, coverage availability, claims litigation, parts supply, repair economics, and driver populations differ internationally. A single global frequency intercept has no stable meaning unless the numerator and denominator are harmonized. Legal covariates should be dated to accident occurrence and contractual cover, rather than attached only to the current residence or the insurer's head office.
The EU's motor insurance framework and subsequent legal amendments are concrete sources for jurisdictional differences in mandatory liability coverage and victim protection. Their existence supports legal stratification in the data architecture; it does not establish a common mean injury severity across the EU. National transposition and case-law applications require additional country-specific research before operational insurance rates are estimated.
A scientifically coherent comparison can proceed in three stages. First, harmonize exposures, payment currency, claim unit, and legal definitions. Second, estimate coverage-specific local patterns and their time dynamics. Third, investigate whether hierarchical models can share statistical strength across comparable markets without forcing identical liability or settlement processes. International pooling of unharmonized claims should not be mistaken for replication.
Data leakage, reporting selection, and reproducibility
Claim data arrive through reporting, adjudication, and payment systems. The existence of a claim file can be correlated with policy terms, insurer practices, and legal access. A dataset restricted to positive paid claims systematically excludes no-payment settlements and unknown unreported losses. The severity distribution of a random paid claim is therefore not the distribution of every accident's economic cost. For an ultimate cost prediction, recent immature observations require a development method or a sufficiently aged validation cohort.
Practical leakage occurs when the model uses postloss repair estimates, final legal outcome, eventual number of claimants, claim file closure indicators, or later telematics availability as if they had been known at underwriting. This may improve cross-validated fit while making the model inoperable. All predictors must have an information-availability timestamp and an intended decision point. Training and test sets should be created at the policy inception or renewal date appropriate to the use case.
A reproducibility record should contain the source extract identifier, anonymization procedure, cohort and coverage filters, claim transaction aggregation logic, exposure calculation, reporting lag, inflation assumptions, contract terms, outcome definitions, model formulas, hyperparameter search, random seeds, train–test split, evaluation metrics, and a cryptographic fingerprint for the final research files. These requirements allow an independent reviewer to distinguish a forecast that was tested from a retrospective narrative about model plausibility.
Worked statistical checks and limits
Consider a homogeneous portfolio with 10,000 earned vehicle-years and 400 paid collision claims. The observed frequency is 0.04 claims per vehicle-year, or 4 claims per 100 vehicle-years. If total covered paid collision loss is $2,000,000, mean paid severity is $5,000 and implied paid cost is $200 per vehicle-year. This is arithmetic on a declared synthetic example; no real insurer or study is represented.
If the portfolio contains many small claims and a few very large ones, bootstrap uncertainty around the mean severity requires resampling at a suitably independent policyholder or event level. Resampling payments within the same claim as if each were an independent loss artificially increases the apparent sample. If the portfolio grows to 20,000 earned years under otherwise identical conditions, a simplistic Poisson scaling would double expected claim count, but there is no automatic reason for uncertainty about future price trends, litigation, coverage mix, or severe-event clustering to halve.
The example also demonstrates a distinction between a model's insured loss-cost target and market premium. Taxes, distribution costs, expenses, risk loading, catastrophe provision, capital, reinsurance, and regulated price structures affect the charged rate. A difference between the $200 expected paid cost and a hypothetical higher premium is not in itself evidence of model miscalibration.
Integrated Results and Open Research Questions
Integrated scientific findings
The literature and source reconciliation yield six conclusions. First, frequency and severity form a coherent accounting decomposition only after policy exposure, claim unit, coverage, and payment state have been specified. Second, Poisson and gamma-type GLMs remain valuable baselines because they make assumptions and exposure corrections observable; model elaboration needs prospective evidence of added predictive value. Third, the expected aggregate cost depends on the joint distribution of counts and claim sizes, so conditional dependence and claim-weighting must be tested. Fourth, extreme bodily injury cases, contractual payment ceilings, deductibles, and claim development can govern loss adequacy more strongly than gains in average-frequency ranking. Fifth, a source-defined increase in paid severity can dominate a decline in paid frequency, as the documented U.S. aggregate arithmetic illustrates. Sixth, regulatory and professional standards impose additional constraints on model deployment, separate from prediction accuracy.
These findings support an ordering of analysis grounded in existing actuarial science. Construct a complete exposure and claims data definition; retain separate coverage-specific claim outcomes; use interpretable frequency–severity benchmarks; test overdispersion, zero processes, and dependence; compare against direct aggregate cost models; validate subsequent accident or underwriting periods; and inspect discriminatory variables, legal permissions, inflation, and pricing governance. The ordering is an analytical synthesis of the reviewed literature and standards, not an independent doctrine or newly claimed theory.
Where empirical research remains necessary
Insurer microdata could evaluate whether a conditional count effect on severity survives adjustment for driver history, jurisdiction, vehicle technology, economic inflation, settlement delay, and claim-limit profiles. A multi-country study could examine whether hierarchical sharing improves prediction while respecting noncomparable injury compensation regimes. A telematics experiment could estimate the contribution of true driving change apart from selection into the program. These projects require lawful data access, prospective study design, and enough severe claims for uncertainty estimation.
Other priorities include calibration of large injury settlements after legislative changes, deductible-induced reporting selection, common-event dependence in comprehensive coverage, interaction of repair inflation with vehicle automation and electric drivetrains, and tests of model fairness when commercially attractive variables are weakly related to risk. Observational studies should report identification assumptions and interpret predictive improvements as predictive unless a defensible causal design is actually implemented.
Limits of the present dossier
The empirical table is a published U.S. industry aggregate with coverage-specific exclusions and no microdata, variance estimates, or causal controls. Its 2019-to-2024 comparisons describe nominal observed paid-claim experience; they are not extrapolated into a 2026 market forecast. Publisher abstracts and institutional pages support some historical methodological and regulatory claims; they do not represent a full independent reanalysis of every primary dataset. Contract law varies by jurisdiction, and several referenced actuarial standards have U.S.-specific professional scope. All derived numerical results are independently reproducible from the visible source table; the document introduces neither a fitted predictor nor an invented empirical performance statistic.
Conclusion
Motor insurance frequency and severity modeling is a problem of statistical definition, contractual interpretation, and prospective inference. Claim counts acquire meaning through their covered exposure and event-counting convention. Conditional severity acquires meaning through coverage, claim maturity, deductibles, limits, liability law, settlement practice, and expense treatment. Expected aggregate cost links these objects through their joint distribution. The familiar product of annual claim rate and mean claim amount remains useful when its sampling, conditioning, and dependence assumptions are satisfied; where those assumptions fail, joint, conditional, hurdle, compound, or direct aggregate models allow more faithful estimation.
The comparative evidence illustrates an economically significant pattern: declining observed paid claim frequency can coexist with rising per-vehicle paid loss costs when average payment severity increases enough. In the cited 2019 and 2024 U.S. paid-claim series, that pattern is visible in bodily injury liability, property damage liability, and collision, while comprehensive shows increases in both recorded frequency and severity. These are source-bounded descriptive findings. They do not identify underlying causes or support direct international rate transfers.
A scientifically accountable motor pricing program therefore tests calibration of frequency, conditional severity, aggregate expected loss, and the high-cost tail against future exposure cohorts. It traces each estimate to its legal and data definitions, evaluates drift and clustered claims, documents data availability at the underwriting decision, and distinguishes risk classification from customer-specific commercial price optimization. The result is a reproducible framework for interpreting established actuarial methods under real contractual and supervisory conditions, with explicit boundaries for future empirical investigation.
References
Nelder, John A., and Robert W. M. Wedderburn. 1972. “Generalized Linear Models.” Journal of the Royal Statistical Society, Series A 135 (3): 370–384. https://doi.org/10.2307/2344614
Dionne, Georges, and Charles Vanasse. 1989. “A Generalization of Automobile Insurance Rating Models: The Negative Binomial Distribution with a Regression Component.” ASTIN Bulletin 19 (2): 199–212. https://doi.org/10.2143/AST.19.2.2014909
Jørgensen, Bent, and Marta C. Paes de Souza. 1994. “Fitting Tweedie's Compound Poisson Model to Insurance Claims Data.” Scandinavian Actuarial Journal 1994 (1): 69–93. https://doi.org/10.1080/03461238.1994.10413930
Ohlsson, Esbjörn, and Björn Johansson. 2010. “Non-Life Insurance Pricing with Generalized Linear Models.” Berlin: Springer. https://doi.org/10.1007/978-3-642-10791-7
Garrido, José, Christian Genest, and Juliana Schulz. 2016. “Generalized Linear Models for Dependent Frequency and Severity of Insurance Claims.” Insurance: Mathematics and Economics 70: 205–215. https://doi.org/10.1016/j.insmatheco.2016.06.006
Shi, Peng, Xiaoping Feng, and Anastasia Ivantsova. 2015. “Dependent Frequency–Severity Modeling of Insurance Claims.” Insurance: Mathematics and Economics 64: 417–428. https://doi.org/10.1016/j.insmatheco.2015.07.006
Verbelen, Roel, Katrien Antonio, and Gerda Claeskens. 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
Wüthrich, Mario V., and Michael Merz. 2023. “Statistical Foundations of Actuarial Learning and Its Applications.” Cham: Springer. https://doi.org/10.1007/978-3-031-12409-9
Goldburd, Mark, Anand Khare, Dan Tevet, and Dmitriy Guller. 2025. “Generalized Linear Models for Insurance Rating.” Casualty Actuarial Society Monograph No. 5, updated online edition. https://www.casact.org/monograph/cas-monograph-no-5
Actuarial Standards Board. 2011. “ASOP No. 12: Risk Classification (for All Practice Areas).” Revised May 2011, U.S. professional standard. https://www.actuarialstandardsboard.org/asops/risk-classification-practice-areas/
Actuarial Standards Board. 2017. “ASOP No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention.” Effective August 1, 2018. https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Actuarial Standards Board. 2019. “ASOP No. 56: Modeling.” Effective October 1, 2020. https://www.actuarialstandardsboard.org/asops/modeling-3/
National Association of Insurance Commissioners. 2026. “NAIC Releases 2022/2023 Auto Insurance Database Report.” News release, February 13, 2026. https://content.naic.org/article/naic-releases-20222023-auto-insurance-database-report
Insurance Information Institute, with data from ISO/Verisk. 2026. “Facts and Statistics: Auto Insurance.” Private Passenger Auto Insurance Losses, 2015–2024, online table. https://www-legacy.iii.org/fact-statistic/facts-statistics-auto-insurance
National Highway Traffic Safety Administration. n.d.. “Crash Report Sampling System (CRSS).” U.S. Department of Transportation program description. https://www.nhtsa.gov/crash-data-systems/crash-report-sampling-system
National Highway Traffic Safety Administration. n.d.. “Fatality Analysis Reporting System (FARS).” U.S. Department of Transportation program description. https://www.nhtsa.gov/crash-data-systems/fatality-analysis-reporting-system
World Health Organization. 2023. “Global Status Report on Road Safety 2023.” Geneva: WHO. https://www.who.int/publications/i/item/9789240086517
German Insurance Association (GDV). 2026. “Ersatzteilpreisanstieg deutlich über Inflationsrate.” News release, October 12, 2026. https://www.gdv.de/gdv/medien/medieninformationen/ersatzteilpreisanstieg-deutlich-ueber-inflationsrate-216276
European Insurance and Occupational Pensions Authority. 2023. “Supervisory Statement on Differential Pricing Practices in Non-Life Insurance Lines of Business.” March 16, 2023. https://www.eiopa.europa.eu/publications/supervisory-statement-differential-pricing-practices-non-life-insurance-lines-business_en
European Insurance and Occupational Pensions Authority. 2025. “Opinion on Artificial Intelligence Governance and Risk Management.” August 6, 2025. https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en
National Association of Insurance Commissioners. 2023. “Model Bulletin on the Use of Artificial Intelligence Systems by Insurers.” Adopted December 4, 2023. https://content.naic.org/insurance-topics/artificial-intelligence
European Parliament and Council. 2009. “Directive 2009/103/EC Relating to Insurance Against Civil Liability in Respect of the Use of Motor Vehicles.” September 16, 2009. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32009L0103
European Parliament and Council. 2021. “Directive (EU) 2021/2118 Amending Directive 2009/103/EC.” November 24, 2021. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32021L2118
IFRS Foundation. 2023. “IFRS 17 Insurance Contracts.” Effective annual reporting periods beginning on or after January 1, 2023. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
Lambert, Diane. 1992. “Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing.” Technometrics 34 (1): 1–14. https://doi.org/10.2307/1269547
National Association of Insurance Commissioners. n.d.. “What You Should Know About Auto Insurance Coverage.” NAIC consumer coverage overview. https://content.naic.org/article/what-you-should-know-about-auto-insurance-coverage