Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Motor Insurance Bonus-Malus and Experience Rating Systems

Abstract. Motor insurance bonus-malus systems adjust premiums or rating classes in light of an insured's prior claim experience. They are simultaneously statistical devices for learning about unobserved risk, contracts establishing intertemporal financial incentives, and institutional rules governing the portability and interpretation of insurance histories. This monograph reconstructs their actuarial foundations, transition mechanisms, credibility estimators, claim-reporting incentives, and legal variation. It distinguishes the frequency of accidents from reported claims; identifies why class-transition matrices can be accurately estimated while underlying crash risk remains unidentified; and shows how actuarially predictive history can interact with exposure, volatility, asymmetric information, and distributive fairness. A finite-state Markov illustration, a Gamma–Poisson credibility derivation, and a reporting-threshold illustration are expressly hypothetical and reproducible. Comparative primary-law analysis covers France's regulated coefficient schedule, Italy's universal merit-class arrangements, European Union claims-history portability and statement formatting, and the more heterogeneous risk-based rating context of the United States. The central conclusion is that a bonus-malus system must be assessed across four different objects: the physical accident process, the reporting and settlement process, the premium-transition rule, and the institutional allocation of costs. A stable average bonus, strong statistical discrimination, and reduced observed claims cannot independently establish accident prevention, equitable treatment, or adequate premiums. The dossier offers an integrated critical synthesis of existing actuarial and insurance-economic knowledge without claiming a new theoretical system or original empirical estimation.

Keywords: Bonus-malus; experience rating; motor insurance; automobile insurance; merit rating; Markov chains; credibility theory; Poisson–Gamma model; claim frequency; claim suppression; bonus hunger; insurance incentives; risk classification; fairness; claims-history statements; cross-border insurance.

Introduction

The premium paid for motor insurance is a price for assuming an uncertain and legally specified obligation. An insurer begins with observable characteristics of the driver, vehicle, use, territory, coverage, and insurance period, then may alter the price at renewal according to what happened during earlier periods. A bonus-malus system makes that retrospective adjustment explicit. It rewards favorable recorded experience by a bonus, penalizes specified claims by a malus, or moves a contract between classes having different premiums. Under experience rating more broadly, the insurer estimates an individual or group risk using accumulated experience and a statistical comparison with a broader reference population. The formal and practical relations among these operations are the object of this monograph.

The literature establishes a substantial actuarial foundation. Loimaranta's asymptotic analysis of bonus systems (1972) made finite-state Markov methods central to measuring long-run premiums and the responsiveness of tariffs to claim frequency. Lemaire's book on automobile bonus-malus systems (1995) consolidated design methods and comparisons; Lemaire and Zi's cross-country comparison of 30 systems (1994) evaluated simulated tariff outcomes for systems from 22 countries. That historical comparison describes the systems examined in 1994 and is not evidence of current rules in those jurisdictions. More recent work makes claim-reporting behavior endogenous to the insurance mechanism: Charpentier, David, and Elie's optimal claiming strategies analysis (2017) shows how the claims decision itself changes the transition probabilities.

This study applies a source-traceable analytical method. Mathematical propositions are derived from stated assumptions; stylized calculations are labeled hypothetical. Empirical findings are attributed to their original study populations. Legal provisions are taken from the operative public instruments where available and distinguished from insurance industry practice and actuarial standards. No insurer microdata were independently obtained, so this is a critical synthetic monograph and not an empirical estimate of current national accident or reporting rates. Source research and statutory verification were performed through November 14, 2026.

Three distinctions govern the investigation. An accident is an event involving physical harm or potential damage; a claim is a request under an insurance contract; a premium-relevant claim is an event that the applicable tariff or regulatory schedule actually counts. The absence of a claim may reflect good driving, low exposure, good luck, self-financing of small losses, an excluded event, or a contractual reporting boundary. Finally, a bonus-malus class describes the operation of a specified tariff; it is not by itself an independent measurement of an individual's causal contribution to road safety.

Conceptual, Historical, and Primary-Source Foundations

Definition and domain

A classical discrete bonus-malus contract contains a finite set of rating classes, an initial class, a premium scale indexed by class, and a rule assigning the next class given the present class and specified claim experience over a defined observation interval. Claim-free renewal may lead to a cheaper class; an at-fault or otherwise tariff-countable claim may lead to a more expensive class. Some jurisdictions express this logic through a multiplicative coefficient instead of a numbered class. Lemaire (1995) discusses the construction and optimization of these systems; the terminology and mathematical structure are also explicit in Loimaranta's original article.

The design should identify the insured object and the exposure base. Annual third-party motor liability insurance, vehicle-damage coverage, a fleet risk-sharing agreement, and a voluntary telematics product need not penalize the same events. When an accident is only partially attributable to the insured driver, a statutory or contractual rule may reduce the malus; when a collision is attributable solely to another party, the claim's treatment may differ from a culpable at-fault accident. A record describing all damage claims is therefore not automatically interchangeable with the set of claims that affect renewal premiums.

Experience rating is broader than a named bonus-malus table. It updates an estimate of expected future cost after observing experience. The new estimate may modify a premium continuously, through a coefficient, or by assigning a discrete class. The Dionne–Vanasse regression approach (1989) shows how observable classification and posterior experience can enter the same automobile model. Their subsequent automobile ratemaking study (1992) uses individual classification variables and retrospective experience in a statistically integrated premium framework, reporting a sample of 19,013 drivers. These works establish the importance of combining prior group information and posterior individual evidence rather than treating claims history as an isolated risk index.

Historical evolution and limits of generalization

Early bonus systems developed in European motor liability markets before the modern predictive-modeling industry. The sequence from conditional premium steps to systematic actuarial evaluation was documented in Loimaranta (1972), in De Pril's comparative definition of efficiency (1978), and later in Lemaire and Zi (1994). The research history matters because certain indexes and comparisons depend on assumptions—Poisson counts, fixed annual exposure, stable risks, identical initial classes—that are transparent in the classical models but can fail when imported mechanically into contemporary portfolios.

International comparison starts from a distinction between a country-specific legal schedule and insurer-specific experience rating. Some systems prescribe transition coefficients or legally recognized merit classes. Others allow companies to establish different no-claims discounts, claim surcharges, underwriting rules, or telematics-linked pricing under local supervision. The United States illustrates the latter diversity: the National Association of Insurance Commissioners' motor insurance rating overview identifies driving record and claims history among numerous rating inputs, while insurance practice and statutory permissible factors are determined at a more granular level. This is not evidence that every US insurer uses a European-style statutory bonus-malus matrix.

A comparative manuscript also must resist temporal substitution. The publication year of a journal article, the year of an EU directive, the application date of an implementing regulation, and the current consolidated national rule refer to different legal or evidentiary periods. The 2021 EU amending directive and the 2024 standardized statement regulation demonstrate why a current claim-history analysis should not reproduce the pre-amendment text of Directive 2009/103/EC as if it stated all current obligations.

Four separate processes inside a bonus-malus record

The first process is exposure: trips, vehicle-kilometers, hours on road, road classes, territorial conditions, and policy duration. The second is the stochastic accident and damage process, which generates events before insurance administration. The third is the reporting, classification, payment, and settlement process, which transforms events into records. The fourth is the tariff process, which maps qualifying records into a future price. A complete causal interpretation requires knowing which process produced an observed change. For example, a reduction in reported collision claims might result from reduced travel, safer driving, a larger deductible, diminished reporting incentives, or changed event definitions. An unchanged reported-claim count after exposure grows may conceal a fall in the rate per unit of driving.

This four-process decomposition is an organizing synthesis of the established evidence, not an independently proposed economic doctrine. Its utility is evidentiary: it allocates each proposition to its observable generating mechanism and prevents a pricing-rule evaluation from being mistaken for a causal road-safety evaluation.

Bonus-Malus Transition Rules and Long-Run Dynamics

Formal representation

Let the finite class set be S={0,1,…,K}S=\{0,1,\ldots,K\}. Write CtC_t for the class applied during policy year tt, NtN_t for the number of tariff-countable claims during the observation window, and T(i,n)T(i,n) for the authorized transition after nn such claims from class ii. The next class is

Ct+1=T(Ct,Nt).C_{t+1}=T(C_t,N_t).

Let bib_i be the premium relativity in class ii, and Bt(x)B_t(x) a base premium determined by approved contemporaneous risk factors xx. A simplified renewal gross premium before expenses, taxes, discounts or other adjustments is Pt=Bt(x)bCtP_t=B_t(x)b_{C_t}. This decomposition identifies an important operational fact: a driver can move to a better bonus class and still pay a higher monetary premium when the underlying base rate changes because of inflation, vehicle substitution or broader experience. The tariff coefficient and the monetary premium must be audited separately.

For a time-homogeneous independent Poisson claim-count model with annual mean λ\lambda,

Pr⁡(Nt=n∣λ)=e−λλn/n!,Pij(λ)=∑n:T(i,n)=je−λλn/n!.\Pr(N_t=n\mid\lambda)=e^{-\lambda}\lambda^n/n!,\qquad P_{ij}(\lambda)=\sum_{n:T(i,n)=j}e^{-\lambda}\lambda^n/n!.

The matrix P(λ)P(\lambda) has nonnegative elements and row sums equal to one. If the finite chain is irreducible and aperiodic, it has a unique stationary distribution π(λ)\pi(\lambda), satisfying πP=π\pi P=\pi and ∑iπi=1\sum_i\pi_i=1. The stationary average relativity is bˉ(λ)=∑iπi(λ)bi\bar b(\lambda)=\sum_i \pi_i(\lambda)b_i. This is the standard line of analysis initiated by Loimaranta's Markov framework and extended in De Pril's efficiency study. Stationarity does not imply a competitive or financial equilibrium, adequacy of rates, or unchanged claims behavior.

A reproducible three-class example

Consider the hypothetical classes 0,1,20,1,2, with premium relativities (0.70,1.00,1.50)(0.70,1.00,1.50). A year with zero tariff-countable claims moves the insured down one class toward zero, and any positive number of such claims moves the insured up one class toward two. Define a=e−λa=e^{-\lambda} and q=1−aq=1-a. The transition matrix is

P=(aq0a0q0aq).P=\begin{pmatrix}a&q&0\\a&0&q\\0&a&q\end{pmatrix}.

The stationary equations give π1=(q/a)π0\pi_1=(q/a)\pi_0 and π2=(q/a)2π0\pi_2=(q/a)^2\pi_0. Therefore, with r=q/a=eλ−1r=q/a=e^{\lambda}-1,

π(λ)=(1,r,r2)1+r+r2,bˉ(λ)=0.70+1.00r+1.50r21+r+r2.\pi(\lambda)=\frac{(1,r,r^2)}{1+r+r^2},\qquad \bar b(\lambda)=\frac{0.70+1.00r+1.50r^2}{1+r+r^2}.

These formulas also prove monotonicity of long-run premium relativity for this particular choice of classes and transitions: as λ\lambda increases, probability mass shifts toward the higher classes. The result is conditional on the example's assumptions, not an empirical assertion about any named national bonus-malus system. A system with severe claim penalties, claims-band rules, or slow bonus recovery can create very different financial paths. In a real tariff, a bonus scale should be calibrated to loss costs and commercial loads; no such calibration is attempted in the illustrative numbers.

The distinction between transient and stationary behavior is material. A new insured starting in class one initially pays relativity 1.00 irrespective of the stationary distribution. After one claim-free period the next relativity is 0.70, whereas a positive claim count moves it to 1.50. Long-term average premiums depend additionally on the probability of future claims, renewal survival, switching, and rating-variable changes. Hence a stationary distribution computed for an immortal, unchanged policy cohort cannot be interpreted as an immediately observed market distribution.

Measures of system performance

A mathematically defined performance comparison can examine the stationary premium curve, its sensitivity to λ\lambda, variation in policyholders' payments, speed of bonus recovery after a claim, and the fraction of premium movement due to class versus base rate. Lemaire and Zi (1994) explicitly considered stationary levels, variability and elasticity, along with the incentive to suppress small claims. De Pril (1978) generalized efficiency measures for finite horizons and whole risk groups.

A useful elasticity diagnostic is η(λ)=dlog⁡bˉ(λ)/dlog⁡λ\eta(\lambda)=\mathrm{d}\log\bar b(\lambda)/\mathrm{d}\log\lambda when differentiability and positive values hold. It asks how sensitive long-run price is to a proportional change in claim frequency. It is a descriptive property of the model. It should not be advertised as a uniquely defined universal welfare or fairness score. A stronger malus may increase discrimination between modeled risk frequencies while reducing affordability, increasing strategic underreporting, or causing churn. A steep price response to recorded claims is therefore not sufficient evidence of efficient prevention.

Heterogeneity and persistence

A homogeneous Poisson model makes λ\lambda identical for all members of a given risk cell and stable over time. Actual fleets are mixtures of drivers and exposures. If Λ\Lambda varies across policyholders, then N∣ΛN\mid\Lambda may be Poisson while the marginal count distribution is overdispersed. Correlation between repeated annual claims can then emerge from persistent latent heterogeneity. The Dionne–Vanasse mixed model is an early automobile-specific account of this problem. A stationary class distribution conditional on each λ\lambda cannot be replaced by the distribution obtained from a single mean E[Λ]E[\Lambda]: in general E[π(Λ)]≠π(E[Λ])E[\pi(\Lambda)]\neq\pi(E[\Lambda]) because the transition and invariant-distribution mappings are nonlinear.

Claims generated by a change in mileage, by driver entry or exit, by an exceptional catastrophe, or by coverage redesign may also break time homogeneity. A credible evaluation should specify the exposure unit, calendar interval, claims development, and survival or renewal process before using a Markov-chain result to support public or commercial decisions.

Credibility, Experience Rating, and Claim Suppression

Individual risk learning under a Gamma–Poisson model

The conceptual role of experience rating is to combine information available before a contract period with information generated by the insured's own claims. A simple Bayesian credibility calculation makes the tradeoff explicit. Let ete_t be a contract's exposure during period tt, expressed in consistent policy-year equivalents, and assume the reported count Nt∣λ∼Poisson⁡(etλ)N_t\mid\lambda\sim\operatorname{Poisson}(e_t\lambda). Suppose the constant individual intensity λ\lambda has a Gamma prior with shape α\alpha and rate β\beta. Across mm independent periods conditional on λ\lambda, set E=∑tetE=\sum_t e_t and n=∑tNtn=\sum_t N_t. Then

λ∣N1,…,Nm∼Gamma⁡(α+n,β+E)\lambda\mid N_1,\dots,N_m \sim \operatorname{Gamma}(\alpha+n,\beta+E)

and

E[λ∣N1,…,Nm]=α+nβ+E=(1−Eβ+E)αβ+Eβ+EnE.E[\lambda\mid N_1,\dots,N_m]=\frac{\alpha+n}{\beta+E} =\left(1-\frac{E}{\beta+E}\right)\frac{\alpha}{\beta} +\frac{E}{\beta+E}\frac{n}{E}.

The credibility weight Z=E/(β+E)Z=E/(\beta+E) increases with the exposure actually observed, not merely with calendar age. This is one reason that two contracts carrying the same number of claim-free renewals may offer different statistical evidence when one vehicle traveled little and the other traveled extensively. The Actuarial Standards Board's Credibility Procedures (ASOP No. 25, revised 2013, effective 2014) gives professional guidance about selecting and applying credibility methods in the practice areas to which it applies. It is a US actuarial practice standard, not a universally binding insurance law or a mandatory formula for a statutory bonus-malus tariff.

As an explicitly invented calculation, take α=2\alpha=2, β=10\beta=10, a prior mean of 0.20 countable claims per exposure-year, E=5E=5 full exposure-years, and one observed claim. The posterior mean is (2+1)/(10+5)=0.20(2+1)/(10+5)=0.20, exactly the prior mean in this special numerical configuration. Five claim-free exposure-years instead give 2/15≈0.13332/15\approx0.1333. With three claims over the same exposure, the posterior estimate is 5/15≈0.33335/15\approx0.3333. These figures are model outputs, not an estimated population accident frequency. A credibility premium proportional to each posterior mean would require separate expected-severity assumptions, permissible classification rules, expenses, taxes, and rate adequacy considerations.

If a priori covariates explain part of λ\lambda, it is typically more coherent to estimate a conditional baseline λ(x)\lambda(x) and a residual individual effect than to pool incomparable risks and let the bonus-malus classes absorb all differences. Dionne and Vanasse (1992) show how a priori and a posteriori factors can be integrated in automobile ratemaking. A single posterior risk estimate should also be treated carefully when the covered exposure, driver, vehicle, or territory has changed since the history was recorded.

The observed count is selected information

Claim reporting is an economic decision partly influenced by the tariff. Define AtA_t as the number of accidents meeting a stated physical-event definition and RtR_t as the number reported as claims. In a simplified independent-thinning model, let At∼Poisson⁡(etλA)A_t\sim\operatorname{Poisson}(e_t\lambda_A), and suppose each accident generates a claim with constant probability qq. Then

Rt∼Poisson⁡(etλAq).R_t\sim\operatorname{Poisson}(e_t\lambda_A q).

This elementary result supplies an identifiability warning. Without independent observation of crashes or a defensible restriction on qq, the claim database identifies the product λAq\lambda_Aq, not the physical accident frequency λA\lambda_A and reporting propensity qq separately. Two populations with the same reported rate may differ substantially in crash exposure and reporting behavior. Adding risk classes does not solve that statistical problem by itself.

The simple thinning equation is deliberately narrow. In actual bonus-malus systems, reporting probability may depend on damage size, fault, deductible, current class, the number of earlier claims in the year, expected future premium increases, contractual reporting obligations, and the remaining time to renewal. The Boucher–Denuit–Guillén panel study (2009) models the distinction between accidents and claims using zero-inflated Poisson approaches and a numerical illustration drawing on reported claims from a Spanish insurer. Its estimated implications depend on the assumed censoring mechanism. It cannot be generalized into a known universal fraction of unreported accidents.

The claim count and the damage count can also differ without deliberate suppression. Multiple victims or coverage heads may arise from one crash; one event may generate several transactions; a nil payment may still be notifiable under a policy; an unrecoverable third-party loss may have a different premium effect from an insured-at-fault payment. Data dictionaries must explain whether observations count accidents, claim files, fault-qualified claims, paid claims, settled claims, or claims reported during a defined development window.

Bonus hunger and rational claiming

The term bonus hunger describes a policyholder's incentive to pay small losses directly or otherwise avoid a claim that would damage a valuable future premium position. It is a long-studied effect, not a newly discovered behavioral mechanism. Charpentier, David, and Elie (2017) develop optimal reporting thresholds in a Markov setting and show that reporting decisions alter the effective transition matrix. Their analysis is explicitly about financial choices under the modeled rules and does not repeal a driver's duty to notify an insurer, a police authority, or an injured third party where such a duty applies.

To state the decision generically, suppose a covered incident would reimburse I(L,d)I(L,d) on damage amount LL after the applicable deductible dd, while reporting changes the expected discounted future premium cost by ΔV\Delta V. In a highly simplified wealth-neutral, legally permitted decision environment, the insured prefers the claim when I(L,d)>ΔVI(L,d)>\Delta V. Equality makes the insured indifferent, assuming identical administrative and liquidity costs. The quantity ΔV\Delta V must reflect future stochastic paths, renewal probabilities, time discounting, changing base rates, class recovery and any claim-free protection, rather than an infinite sequence of fixed premium differences.

Consider the invented three-class system with current class zero, next-year base premium $1,200, and the 0.70 versus 1.00 next-year class relativities. The immediate next-year malus difference is $360. At a 5% one-year discount rate in this illustrative arithmetic, its present value is 360/1.05≈$342.86360/1.05\approx\$342.86. With a $250 deductible, a $450 covered loss produces a simplified reimbursement of $200, below that one-year difference; a $650 covered loss produces $400, above it. This is an example of the arithmetic only. Earlier or later changes in classes, future accidents, taxes, claims handling, and legal reporting duties can change the actual preferred or required action. In particular, mandatory liability reporting and duties toward injured parties cannot be displaced by this private premium comparison.

The full sequential decision problem can be written as a discounted dynamic program. Let V(i,t)V(i,t) denote expected discounted future private cost at class ii and time or renewal state tt, conditional on the model's information. For each permissible reporting action aa, an insured minimizes a sum of immediate out-of-pocket incident cost and δE[V(T(i,Na),t+1)]\delta E[V(T(i,N^a),t+1)]. The action changes the count NaN^a recognized by the rating rule. The resulting optimization depends on the admissible action set, which is restricted by insurance law and contract terms. Treating every claim as suppressible would make an actuarial model less realistic and an accompanying consumer interpretation unsound.

Feedback into experience rating

A tariff calibrated to historical reported claims may face feedback when a schedule is changed. Increasing the malus increases expected future cost of reporting, which can lower the frequency of small reported claims even if underlying crashes remain constant. A pricing model fitted after a reform might therefore appear to show an improvement in low-cost loss frequency that partly reflects policyholder behavior. Similarly, a new deductible changes self-financing decisions and the observable mix of damage severities.

The combination of the claim-suppression panel model and the optimal claiming framework supports a methodological conclusion: estimating bonus-malus effectiveness from claims administration alone is generally a joint study of underlying risk, coverage and reporting behavior. Separation requires additional data, carefully specified exclusion assumptions, or a credible experiment. Credibility weight attached to observed claims can otherwise be precisely computed yet attached to an incompletely measured phenomenon.

Fairness, Incentives, and Institutional Design

Actuarial differentiation and its moral relevance

Claim history provides information about future costs when accident propensity, exposure or behavior is persistent. A risk-based tariff may therefore charge different amounts to insureds with otherwise similar recorded characteristics. This operation has an actuarial rationale under appropriate calibration. The ASOP No. 12 risk-classification standard articulates professional considerations when building or evaluating classification systems in its applicable practice settings. Statistical predictive association, legal permissibility, causal responsibility and distributive justification remain different questions, each requiring a separate criterion.

A crash in one year may result from weather, the driving of another road user, construction hazards or random variation. Repeated claim-free years can combine safer driving with fortunate circumstances. Even if claim frequency is predictive, a tariff based on realized claims imposes part of the temporal volatility of random loss experience on the policyholder. The insurer can smooth risk through pooling; a class-transition schedule deliberately transfers some of that uncertainty back into future individual premiums. Whether that allocation is fair depends on the purpose of the scheme, the extent of controllable risk, consumer expectations, legal rights, vulnerability and the affordability of required liability cover.

There are at least four separable evaluation questions. Predictive fairness asks whether a factor is relevant to expected insured loss conditional on exposure and other variables. Procedural fairness asks whether the insured can discover how a claim enters the system, challenge an incorrect record, and carry verified history to another insurer. Distributional fairness asks who bears larger price swings, particularly among low-income or low-mileage insureds and those with limited alternative transport. Behavioral fairness asks whether penalties reward reasonably controllable prevention or simply penalize involuntary events. These are analytical evaluation dimensions and make no claim to establish a novel moral doctrine.

Why accident reduction and reported-claim reduction diverge

Bonus-malus systems may encourage precaution because future premiums become contingent on experience. Their incentive effect, however, cannot be measured by a decline in reported claims alone. As Charpentier, David, and Elie (2017) show, the scheme itself creates a strategic claiming margin. Thus the empirical target matters: fewer reported property-damage claims could mean less physical damage or more self-insurance. Road safety evaluation should preferentially combine claims with independently recorded incidents, vehicle exposure and injury severity where lawful access exists, and should account for changing recording conventions.

Another complication is asymmetric learning. Cohen's automobile insurance information study (2005) reports evidence consistent with information asymmetries, selective switching and imperfect disclosure of past claims in its study setting. It does not show that all switching drivers misstate claims history, nor does it establish a uniform effect of any particular national class schedule. A new insurer observing a standardized history may reduce some informational disadvantages while retaining uncertainty about unobserved accidents, driving intensity, changes in vehicles and policyholder behavior.

Fairness assessment should consequently be conditional on both data generation and regulation. In many systems, the policyholder's claim record is an important rating input, yet the actuarial gain from adding that input can be smaller or larger depending on how much of the same information is already captured by mileage, territory, age, driving record, telematics or other permitted variables. The question is empirical: does retrospective information improve out-of-sample calibration for comparable exposure after accounting for how the history was produced?

Affordability, risk pooling and incentive strength

Strong malus incentives can improve alignment between future premiums and modeled loss frequency, but may also make an already compulsory motor liability product unaffordable for some insureds following accidents. Severe penalties can increase lapse and insurer shopping, concentrating observed higher risks within particular portfolios and changing the mix of persons for whom the long-run class distribution is computed. A class scale that works well for a closed cohort can therefore perform poorly in an open market with entry, exit and regulated portability.

An insurer's average relativity is meaningful only relative to an adequate base rate. If nearly every experienced driver receives a deep nominal discount while the entry premium fails to reflect the portfolio's expected claims, the resulting scheme may erode premium adequacy. Conversely, a class scale yielding a high average loading may be financially sustainable but place a large fraction of cost on a relatively small group of recent claimants. Tariff choice cannot be judged from class elasticity or the number of discount levels alone. Lemaire's optimal-system design treatments address competing objectives and alternative loss functions within actuarial models; they do not establish a single internationally applicable objective function.

Protected no-claims discounts and claim forgiveness complicate the effective incentives. If a protection endorsement preserves the displayed bonus class after an event, the underlying base premium may nevertheless change at renewal under a permissible tariff. An advertised protected discount is therefore not equivalent to a guarantee of an unchanged monetary renewal premium. The full product description, including its coverage and tariff definitions, is needed before comparing protected and unprotected policy options.

Legal status, information rights and contestability

Legal rules determine which events can produce a malus, whether coefficients must be portable, and what a consumer may contest. Record accuracy is foundational because an incorrectly attributed at-fault claim has two potential consequences: immediate claim handling effects and future rating effects. Reason codes, fault determinations, effective policy dates, insured vehicle identity, and insurer transfer records must remain traceable. A sound audit procedure should distinguish (i) an incorrect recorded event, (ii) a correct event with incorrect tariff classification, and (iii) a correct tariff step applied to a changed base premium.

International regulation does not impose a single globally shared claim-record taxonomy. The EU's amended motor insurance framework develops a common set of rights regarding claims-history statements and comparable treatment across Member States. Individual bonus-malus systems nevertheless remain national or contractual arrangements. This institutional separation explains why record portability can be harmonized while the value of a given record in pricing remains different across companies or jurisdictions.

Comparative Evidence, Data Uncertainty, and Validation

France: a statutory reduction–increase coefficient

France provides a useful example of a regulated multiplier rather than merely an insurer-adopted voluntary no-claims discount. The operative annex to Article A121-1 of the French Insurance Code, version applicable since July 24, 2025, begins the coefficient at 1.00. Under the general schedule, a claim-free annual period reduces the coefficient by 5%, subject to a lower limit of 0.50; a tariff-countable claim increases the existing coefficient by 25%, with an upper limit of 3.50. The same primary source provides differing percentages for certain professional vehicle uses, reductions for partial responsibility, a two-consecutive-claim-free-years recovery rule limiting the resulting coefficient to 1.00, and specific exclusions for non-attributable events. The text also includes special protection for the first claim after at least three years at the 0.50 coefficient. These details materially alter mechanical comparisons with generic bonus-malus examples.

The French clause uses rounding down to two decimal places at defined steps. It identifies the reference premium and the coefficient as distinct quantities and specifies information provided on renewal notices. Its current claims-history provisions refer to the EU standard form, introduced by Regulation 2024/1855. Reconstructing French coefficients from administrative data therefore requires the precise observation years, the event's responsible share, vehicle use, contract discontinuities, class history and the version of the regulation in force. An analyst who simply multiplies a year-end coefficient by 0.95 or 1.25 without the statutory exceptions can misstate the legally applicable coefficient.

As an arithmetic illustration of the basic French 5% rule, an initial coefficient of 1.00 would become 0.95 after one qualifying claim-free period, then ⌊100(0.95×0.95)⌋/100=0.90\lfloor 100(0.95\times0.95)\rfloor/100=0.90 after a second under the specified rounding approach. This is a stylized application of the general rule; the exact operative claim period and contract-specific eligibility remain legally relevant. A coefficient of 0.90 does not itself mean a premium 10% below last year's invoice, since the contemporaneous reference premium can change.

Italy: merit classes and portability

Italy employs bonus-malus class concepts in its compulsory motor liability market. The Italian supervisor IVASS describes the universal merit class and claims-history mechanism, including movement in insurance classes depending on recorded claims and continuity of claims experience when the policyholder changes company. The regulator's Measure No. 72 of 2018 sets criteria and evolution rules for the universal conversion merit class. A class number is part of the institution's rating information; it does not alone specify an insurer's final monetary price.

Italy's system illustrates the need to distinguish a common reference classification from insurer-specific pricing. Identical class histories can be quoted differently across insurers because base tariffs, covered risks, ancillary coverage, expenses, other permitted factors and commercial pricing differ. International researchers must also check the unit of merit-class ownership, vehicle substitution and any special rules applicable to household or transferred histories rather than assigning a stable universal class solely from an accident-count series.

The IVASS consumer explanation supports the institutional point that qualifying malus events depend on responsibility rules; a shared-fault record and a majority-responsibility event may be treated differently under the relevant provisions. The study here reports those distinctions at the level confirmed by the named regulator's materials. A full Italian legal implementation audit for a specific policy would additionally require the dated contract wording, operative secondary legislation, and any amendments affecting that insured's class.

European Union: standardizing evidence without a single tariff

Under the amended Motor Insurance Directive 2009/103/EC, Article 16 concerns statements of third-party liability claims involving insured vehicles over at least the preceding five years, or the absence of such claims. The insurer or designated provider must supply the statement within the stipulated period. Directive (EU) 2021/2118 adds equal-treatment and transparency requirements concerning the treatment of claims-history statements from another Member State. The 2021 amending text expressly preserves Member State freedom to establish national bonus-malus rules. Consequently, an EU-wide claims-history document should not be described as an EU-wide class-price tariff.

Commission Implementing Regulation (EU) 2024/1855, applying from July 24, 2025, establishes a common statement template and filling instructions. It addresses the issuer, policyholder, vehicles, contracts, accident and claim information, and additional national details. The form is normally provided electronically, with a paper version free of charge on request. Common document fields facilitate record transfer and comparison; they do not standardize accident risk itself, required deductibles, base premiums or discount percentages.

These three legal levels—directive as amended, implementing regulation, and applicable national and contractual rules—must be reviewed together for cross-border pricing statements. A paper relying only on the original 2009 version of Article 16 would omit later obligations. Conversely, a paper interpreting the EU common statement template as preempting national bonus-malus design would overstate the harmonization.

United States: heterogeneous experience-sensitive rating

The US context is more heterogeneous. The NAIC's automobile insurance topic overview recognizes claims history, driving history, mileage and other attributes as variables that insurers may consider under the applicable regulatory regime. The NAIC consumer explanation of rating distinguishes underwriting decisions from the setting of rates and emphasizes the role of expected claims frequency. These are authoritative summaries of insurance practice rather than a substitute for each state's operative statutes, rate filings, claims-surcharge plans and consumer-protection requirements.

A US motor product can respond to recent at-fault accidents without adopting a finite-state statutory European coefficient. Some tariffs use accident surcharge factors, eligibility tiers, claim-free discounts or other rating factors. Policy forms and regulatory filings determine which claims qualify and how far back experience is considered. The United States should accordingly be used as a comparison of heterogeneous experience-sensitive pricing and state-level governance, not presented as though one universal American bonus-malus formula exists. The applicable US professional context includes the ASOP No. 12 classification guidance and ASOP No. 25 credibility guidance, both of which have clearly defined scope and do not replace state insurance law.

Comparative matrix

Dimension France Italy European Union cross-border layer United States
Principal mechanism in examined sources Statutory reduction–increase coefficient Merit classes and universal conversion class Standardized claims-history evidence and nondiscrimination in its treatment Insurer and jurisdiction-specific rating and claims-history factors
Primary source Insurance Code Annex A121-1 IVASS guidance and Measure 72/2018 Amended Directive 2009/103/EC and Regulation 2024/1855 NAIC rating summaries; individual state law required for binding rules
Price determined uniformly from claim history? Formula for coefficient, alongside reference rate and exceptions Universal merit class is not a universal final price No EU-wide bonus-malus price formula No national uniform class-price matrix
Main comparison risk Missing fault, rounding or statutory exception Confusing class portability with tariff equality Treating a common history statement as a harmonized premium Conflating different state and insurer rating systems

This matrix compares the specific verified sources cited in the accompanying prose. It does not purport to enumerate every permitted motor rating product, every later national amendment or all vehicle subcategories.

A reproducible validation protocol

A rigorous bonus-malus empirical study first defines the observational units: policyholder, driver, household, vehicle, contract, exposure-year and accident. Their relationships may be many-to-many over time. The investigator should then establish the cohort entry and exit dates, policy renewal cycles, reporting lags, claim closure windows, at-fault definitions and whether the available record is insurer-specific or cross-insurer. Reserves, paid loss, gross incurred loss and recovered amounts are monetary variables with different development dynamics; they should not be substituted for each other when estimating claim frequency.

Second, the researcher should reconstruct the rating rule from the contract or law effective during each period. The minimum reproducible tariff record includes initial class, transition function, cap and floor, premium relativity, permitted exclusions, fault treatment, renewal date, reference premium, optional claim protection, and vehicle or driver-transfer rules. When a schedule changes during the sample, the study should segment the periods, not fit one timeless matrix.

Third, model specification should address heterogeneous annual exposure and nonrandom selection into renewal. Conditional Poisson, negative binomial and hierarchical frequency models may be compared using development and holdout periods. The Dionne–Vanasse methodology supplies a foundational combined a priori/a posteriori model; the Boucher–Denuit–Guillén study supplies an explicit strategy for investigating claims censored from accidents. Any treatment of zero inflation must explain whether extra zero observations represent true low-risk behavior, different exposure, deliberate nonreporting, or missing records. Mixture models can fit the same count distribution with very different substantive interpretations.

Fourth, evaluate several outcomes rather than only mean tariff relativity: actual premium paid, expected covered loss, physical crash incidence per exposure, injury severity, reported-claim count, insurance take-up, switching, disputes about fault, and distribution of price changes. Confidence intervals and sampling variation are indispensable when comparing small risk strata. A high apparent prediction gain from the recent-claims count may disappear under temporal validation when the claim process, regulation or insured population changes.

Fifth, the causal effect of bonus-malus penalties requires a credible comparison of counterfactual driving and reporting behavior. A before–after decline in claim submissions does not identify prevention. Plausible research designs include a rule change affecting otherwise comparable contracts, an exogenous eligibility threshold, or linked accident and claims data, with explicit tests of concurrent changes in coverage, enforcement, exposure and selection. The identification assumptions must be independently argued; this monograph does not claim that such data or an experiment have been conducted.

Finally, validation should quantify legal and operational error. Sample records should be traced from original event through the fault code and claim record to the class transition and renewal invoice. Cross-border studies should verify that the record's document fields correspond to the current EU form and that national application dates have been respected. Only after these checks can a published claim about system efficiency be interpreted at the intended actuarial, behavioral or legal level.

Integrated Results and Open Research Questions

What the source synthesis establishes

The mathematical, economic and legal materials support six integrated findings. First, experience rating learns from insured outcomes. A discrete bonus-malus system makes that learning operational through a class and transition rule, whereas a credibility estimate may update expected risk continuously. This difference permits the same statistical evidence to generate different renewal premiums under otherwise rational tariff choices. A legal system may impose a particular schedule for reasons extending beyond actuarial forecast accuracy.

Second, the relation between true accidents and insurance claims is endogenous to the tariff. If only claims enter a rating model, stronger malus can alter both observed risk information and consumer incentives to provide it. In the elementary thinning formulation, observed frequency is a product of physical frequency and reporting probability; without additional assumptions or data these components are not separately identified. This is a structural limitation of the observation process, independent of the computational accuracy of a fitted algorithm.

Third, the equilibrium shape of a bonus-malus portfolio depends on exposure, risk heterogeneity, entry, exit and policyholder strategy. A stationary distribution conditional on a fixed Poisson intensity is a useful mathematical benchmark. It ceases to describe observed market composition without further assumptions about behavior, claims reporting, contracts and insurer competition. The older Markovian performance literature and the later strategic reporting analysis are complementary precisely because they concern different data-generating mechanisms.

Fourth, the legal comparators reveal multiple dimensions of standardization. The French coefficient schedule prescribes particular tariff movements and exceptions. Italy's merit-class arrangements establish a portable institutional risk-history element. EU law standardizes claim-history statements and addresses their cross-border treatment while retaining distinct national tariffs. US regulation encompasses a heterogeneous field of experience-sensitive rating as summarized in the NAIC's automobile insurance guidance. No country-level label, standing alone, reveals the complete premium algorithm.

Fifth, the fairness of a malus cannot be inferred from whether the insured's claims record predicts future losses. A premium increase can reflect persistent risk, exposure, bad luck, responsibility for harm or selection into a portfolio. A defensible appraisal must test statistical calibration, legal compliance, individual contestability and the allocation of unavoidable risk. Financial incentives should be evaluated against actual incident outcomes and consumer welfare, rather than treated as a substitute for these outcomes.

Sixth, source provenance is a substantive dimension of motor rating research. A transition percentage quoted without the vehicle category, fault category, statutory or contractual version, effective date, observation window and rounding rule is not a replicable tariff rule. Likewise, a reported reduction in claims without the exposure denominator and claims-development convention is not a replicable safety result. Scientific transparency depends on these definitions as directly as on numerical estimation.

Sensitivity and partial identification

The elementary credibility estimator illustrates sensitivity to unobserved accident reporting. Take the five-exposure-year hypothetical record of one reported claim and Gamma prior (α,β)=(2,10)(\alpha,\beta)=(2,10). With full reporting, the posterior mean for a reported event intensity is 0.20. If, for illustration alone, a known and independent reporting probability were q=0.5q=0.5, the correct conditional likelihood for the underlying physical accident intensity would use effective accident-information exposure qE=2.5qE=2.5, giving posterior mean (2+1)/(10+2.5)=0.24(2+1)/(10+2.5)=0.24. The estimates differ despite observing exactly the same claim. The altered inference requires assuming qq known and reporting independently of claim size, fault and class; those assumptions generally require external validation. If qq is unknown, any numerical estimate of the accident intensity inherited from claims alone is conditional on the chosen reporting model.

A second sensitivity arises from exposure. One claim in five full policy-years implies an observed annual rate of 0.20 in the raw ratio. One claim in five contracts with only two total exposure-year equivalents implies 0.50. Calling both a five-year history conceals a 2.5-fold difference in the observed event rate. More precise kilometer-based observation is not automatically a complete solution because risky road conditions and trip types can covary with kilometers. It nevertheless makes the exposure assumption explicit and testable.

A third sensitivity is monetary rather than probabilistic. Suppose the reference premium grows by 12% between renewals while the insured's class relativity falls from 1.00 to 0.95. The new premium before other changes is 1.12×0.95=1.0641.12\times0.95=1.064 times the old premium, a 6.4% increase despite a 5% bonus. This example exposes the error in interpreting a class improvement as a guaranteed invoice reduction. It follows from the decomposition of a renewal rate into baseline and experience multipliers, not from any claim about actual national premium inflation.

Reproducibility and authorship of the synthesis

This monograph's original contribution is an integrated analytical reading of established actuarial scholarship and operative regulatory evidence; the four-process mapping, comparative audit criteria and illustrative calculations are presentation and interpretive syntheses. They are not advanced as novel statistical theorems, new regulatory principles or empirical discoveries. Every calculation is hypothetical, with assumptions stated so that it can be recalculated independently. The comparison intentionally maintains the distinction between actual source-observed facts, formal model implications, institution-specific rules and proposed research tests.

The numerical Markov example can be reproduced by constructing the stated three-by-three transition matrix for any positive λ\lambda, solving πP=π\pi P=\pi, and evaluating bˉ=πb\bar b=\pi b. The Gamma–Poisson example can be reproduced from the posterior parameters. The no-claims recovery and French coefficient examples can be recomputed using the legislated rounding convention. Reproducible calculations are valuable because they expose mistakes in assumptions without requiring a proprietary insurer database; they are not substitutes for estimating the actual parameters of a live portfolio.

Questions suitable for future empirical research

A useful prospective study would join independently observed collisions to insurer claim data while recording mileage, vehicle use, fault, damage severity and the operative tariff version. It could ask how much of a post-reform fall in claims represents prevented accidents rather than suppressed small claims. Another study could examine whether standardized cross-border claim-history statements change insurer-switching patterns and the distribution of premiums conditional on comparable driving exposure, using careful design around the July 2025 application date. A third could compare the welfare consequences of a severe malus and an exposure-adjusted moderate malus, including claim disclosure, insurance continuation, consumer financial volatility and injuries, under an explicitly stated objective.

Such studies require administrative access, appropriate privacy safeguards, credible counterfactual identification, careful legal versioning and an evaluation of missing physical incidents. The present evidence justifies these questions; it does not answer them with invented data. Scientific priority for a new theory or a new estimate is not claimed.

Conclusion

Motor insurance bonus-malus and experience rating are mechanisms for translating past insured experience into prospective contractual prices. Their actuarial power derives from the predictive relationship between past and future risks, tempered by stochastic variability and the credibility of individual observations. Their behavioral consequences arise because the tariff can influence both driving precautions and the decision to submit a claim. Their institutional structure determines which events count, how records travel across insurers, and how errors or differences in fault attribution affect the renewal premium.

The principal methodological result is the need to connect four separately measured systems: accident exposure and physical events, claim reporting and settlement, premium transitions, and legal–institutional rules. A researcher who analyzes only one system risks attributing its change to a different mechanism. Long-run premium sensitivity cannot be used as a single measure of safety, fairness or actuarial adequacy. A decline in insurer claim frequency cannot establish a decline in accident frequency without evidence about exposure and reporting. The transferability of a claims-history statement does not imply transferability of an identical monetary price.

The international legal comparison establishes a spectrum of design rather than one world formula. France's prescribed coefficient system, Italy's universal merit-class framework, the European Union's common claims-history statement and the United States' diverse rating arrangements each organize retrospective insurance information in a different way. Each can be evaluated using shared actuarial mathematics, provided the event definitions, time windows, source dates and legal scope remain explicit.

A scientifically defensible motor bonus-malus evaluation therefore combines a transparent transition rule, conditional risk estimation, an account of claim suppression, exposure-sensitive validation, and a legally grounded assessment of consumer treatment. This integrated standard of inquiry makes empirical claims testable and institutional comparisons intelligible. It defines the final conclusion of the research without asserting a new authorial doctrine or an unverified claim of scientific novelty.

References