Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Loss Frequency Models

Abstract. Loss frequency modeling concerns the probabilistic distribution of the number of insured events or claims in a specified observation period and exposure population. Its mathematical object is a count, but its valid estimation requires institutional definitions of what constitutes a claim, when a risk is exposed, and how the resulting observation enters an insurance dataset. This monograph develops a source-grounded treatment of binomial and Poisson counting, their exposure-adjusted regression forms, mixed-Poisson heterogeneity, gamma–Poisson and negative-binomial distributions, structural zeros, hurdle specifications, underdispersion, temporal dependence, and predictive validation. A central result of the analysis is that apparent departures from a Poisson distribution have distinct interpretations: heterogeneous event propensities, clustered occurrences, unstable exposure measurement, changes in reporting, dependence across time, and different mechanisms generating zeros can produce superficially similar sample signatures. Their statistical consequences diverge when a model is applied to a new portfolio or a changed exposure horizon. Exact probability-generating-function and moment derivations connect individual risk heterogeneity with marginal overdispersion and repeated-period correlation. A worked exposure example separates policy-level rates from pooled rates; further analytical counterexamples establish that matching a mean and variance does not establish the validity of a frequency distribution or of its tail probabilities. The source analysis distinguishes conditional from marginal claims, full distribution models from mean-only quasi-likelihood procedures, and observed claim counts from the underlying occurrence process. It then develops a disciplined workflow for calibration, diagnostics, scoring, stability assessment, and actuarial deployment. The original contribution is an integrated comparative interpretation of established probabilistic methods, explicit derivations and illustrative counterexamples, and a reproducible checklist linking statistical assumptions to insurance-data definitions. No new named theory, proprietary distribution, or empirical discovery is asserted.

Keywords: Insurance claim frequency; event counts; Poisson distribution; binomial distribution; mixed-Poisson process; negative binomial; gamma heterogeneity; actuarial exposure; offsets; count regression; generalized linear models; overdispersion; underdispersion; zero inflation; hurdle models; claim reporting; policyholder heterogeneity; dependence; calibration; probabilistic forecasting; model risk; ratemaking; claims science.

Introduction

A count is the simplest-looking insurance observation and one of the most demanding to interpret. A record reporting two claims may describe two independent accidents, two claims arising from one occurrence, a claim reported and reopened, or a loss allocated across separate coverage sections. A portfolio reporting 100 claims over 1,000 insured vehicle-years has a crude observed frequency of 0.1 per vehicle-year; that division alone says little about the probability of a claim on a particular vehicle, the uncertainty around the estimate, or the frequency distribution of a future portfolio with changing exposures.

The statistical problem is to estimate the distribution of a nonnegative integer-valued random variable after fixing the event definition, population, contract scope, time window and exposure denominator. The applied problem is to determine which distributional statements survive changes in risk mix, measurement systems, calendar periods and economic behavior. Both questions belong to the generalized linear modeling tradition introduced by Nelder and Wedderburn (1972) and its subsequent applications to count data and insurance pricing. The Casualty Actuarial Society's rating monograph explains how these statistical distinctions enter actual classification, calibration and model-development practice.

This dossier is concerned with occurrence counts and claim counts, their mathematical distributions and their estimation. Insurance claim amounts, large-loss severity, total indemnity, compound sums and catastrophe loss financial tails are adjacent questions with their own analytical objects. They enter here only where they clarify the event counted, claims censoring, or downstream consequences of an erroneous frequency estimate. The monograph is primarily linked to Risk Theory and shares scientific scope with Actuarial Science, Insurance Statistics and Predictive Modeling, Insurance Pricing and Reserving, and Claims Science. These are five relationships to one canonical research identity.

The historical foundations have several independent strands. Cox (1955) examined random series of events, including variation and association between events. Nelder and Wedderburn (1972) placed discrete distributions within the GLM framework. Hausman, Hall, and Griliches (1984) showed in their count-data panel analysis why persistent individual variation and innovations need different statistical treatment. Cameron and Trivedi (1990) developed regression-based overdispersion tests. Lambert (1992) gave a formal zero-inflated Poisson specification. These contributions establish distinct instruments for answering distinct empirical questions.

The research method combines formal derivation, assessment of source claims, constructed numerical illustrations, and critique of model interpretations. Mathematical equalities below are proved for the explicitly stated assumptions. Worked counts are synthetic and serve to check calculations, not to estimate real insurance portfolios. Historical articles and books are credited where their publication records or accessible manuscripts have been checked; access to abstracts is distinguished from access to full text in the accompanying evidence ledger. The level of inference for an actuarial model is always constrained by the population and observation mechanism for which its parameters were estimated.

Claim Counts and Observation Windows

The counted event and the estimand

Let N_i(W) denote the count for insured unit i over observation window W. Its probability distribution is defined only after the counting rule has been fixed. An occurrence count records insured happenings under a physical or contractual event rule. A reported-claim count records claim notifications. A payment count records payment transactions. These quantities may coincide in a carefully designed simplified sample, but they represent different random variables in operational insurance.

A collision involving three covered vehicles may produce one event and three claims. A claim reopened after another payment may add a transaction without adding a new claim. A single injury may generate multiple coverages and reserve movements. Treating every claim-system row as one occurrence can therefore violate the counting definition before distributional fitting begins. A scientific dossier on frequency must identify the counting unit, the event key, the policy and coverage linkage, the inclusion of duplicates and reversals, the treatment of reopened claims, and whether simultaneous claims are combined.

The estimand may be the conditional mean E[N_i|X_i,e_i], a marginal distribution P(N_i=n|e_i), a probability of at least one event, P(N_i>=1|X_i,e_i), or a predictive distribution of total future counts across an exposure portfolio. A mean rate does not uniquely determine the probability of any event or the upper tail. If a risk has a Poisson count with mean μ, its probability of at least one event is 1-exp(-μ); under a negative-binomial distribution with the same mean, that probability differs. Consequently a model adequate for pricing the expected claim count may be inadequate for a decision depending on a 99th-percentile count.

For a portfolio, define N_P=sum_i N_i. Its expectation is sum_i E[N_i]. Its variance is sum_i Var(N_i)+2sum_(i<j)Cov(N_i,N_j). Any approximation that ignores the covariance terms asserts a substantive independence assumption. Operationally shared catastrophe exposure, common maintenance contractors, policyholder fleets or reporting-system disruptions may make this assumption implausible.

Policies, exposure, and risk time

Exposure is the measure of opportunity for the counted event to occur. For automobile coverage it may be vehicle-years; for workers' compensation, payroll or working hours may better represent opportunity; for other product designs, insured locations, business volume, trips or transactions can be relevant. Exposure is therefore a model measurement with a specified unit rather than an arbitrary normalization. Two risks can have the same observation duration but quite different opportunities to incur a claim.

Write e_i>0 for a valid measured exposure and r_i for an event rate per exposure unit. In a homogeneous Poisson model, μ_i=e_i r_i. A one-year policy contributes 1 vehicle-year only if the coverage and risk conditions governing this exposure are active over the full period. A policy that existed for six months ordinarily contributes approximately 0.5 vehicle-years under the simplifying assumption that each day carries comparable opportunity. Seasonal driving or changing vehicle use requires a more granular exposure construction.

The distinction between calendar time and risk time is crucial. A commercial property may be exposed to wind throughout the year but to construction operations for only a shorter interval. An agricultural crop has phases with different susceptibility to particular hazards. Aggregating such intervals assumes the event intensity can be integrated meaningfully. If λ_i(t) is a nonnegative intensity, the expected number of events in an interval is the integrated intensity ∫_W λ_i(t)dt, under a counting process with the corresponding compensator. A constant annual rate is a special case.

Observation, development, and reporting

A frequency model needs a defined observation cutoff. Occurred events can be unknown to the insurer when modeling takes place. A reporting delay moves a claim into the administrative system after the event date. A valuation date thus censors or truncates the apparent frequency of a recent accident period relative to an older one. Different data systems may also apply different rules for small claims, no-payment notifications, claims below deductibles, or canceled contracts.

If each true event is independently observed with probability q under a homogeneous Poisson mechanism of mean μ, then the observed count is Poisson with mean qμ. This is the Poisson thinning identity. It explains a fundamental identifiability issue: from reported counts alone, q and μ appear through their product. A low observed frequency can reflect low incidence, weak reporting, or both. The thinning result relies on independent detection; delayed or selectively reported claims usually require a richer process.

A practical modeling dataset should preserve occurrence date, report date, valuation date, policy effective date, cancellation status, coverage and benefit identifiers, event or claim grouping and claim-system changes. Frequency estimates may need development adjustments or an event-occurrence observation model. When those adjustments are performed, their assumptions must be documented separately from the count distribution ultimately selected. A sophisticated negative-binomial fit cannot reconstruct events never admitted to the observed record.

Frequency as a stochastic process

A count at a fixed horizon is a random variable; a time-indexed sequence of counts is a counting process. The distinction determines whether a method represents a single cross-section or temporal evolution. For a homogeneous Poisson process N(t) with intensity λ>0, disjoint time intervals have independent increments, N(t)-N(s) is Poisson with mean λ(t-s), and E[N(t)]=Var(N(t)]=λt. For an inhomogeneous Poisson process, λ is replaced by a deterministic intensity λ(t) and counts over an interval have mean equal to its integral.

Observed claims may arrive in clusters, with event rates affected by weather or local conditions; repeated claims may also reveal time-stable risk characteristics. Cox's analysis of event series is relevant because randomness of interarrival times, variability of counts in fixed windows, and temporal correlation are separate checks. A count model fitted to yearly totals cannot by itself establish that accidents within the year satisfy Poisson independent-increments assumptions.

The modeler should decide whether the question is predictive count distribution over W, risk ranking per exposure, daily operational arrival rate, or long-horizon dependence. Each question specifies different data granularity and validation. Multiple-period frequency forecasts require modeling the relation between periods, not merely refitting an annual marginal distribution.

Poisson and Binomial Models

Bernoulli and binomial counts

Suppose a risk faces n specified independent opportunities, with a common probability p of one qualifying event per opportunity. Let B_j equal 1 if event j occurs and 0 otherwise. Then N=sum_(j=1)^n B_j is binomial: P(N=k)=(n choose k)p^k(1-p)^(n-k), for k=0,...,n. Its expectation is np and variance np(1-p). With 0<p<1, the variance is smaller than the mean. This is a coherent source of underdispersion relative to a Poisson law with the same μ=np.

The binomial framework is suitable when a finite number of well-defined opportunities exists and each opportunity can generate at most one event under the observed definition. It may describe the number of policies with at least one claim over a horizon when the outcome is intentionally capped at one per policy. It does not generally describe the number of all claims on a policy that can sustain several losses. Distinguishing claim incidence from total claim count prevents a common modeling mistake.

As n grows while p decreases and np approaches λ, the binomial distribution converges to Poisson(λ). This classical limiting approximation explains the usefulness of Poisson counts for many infrequent, approximately independent event opportunities. It also identifies where approximation error may matter: the number of opportunities must be sufficiently large and the individual probability sufficiently small. High-frequency repeated events, finite capacity, contagious clustering and common shocks are different regimes.

For heterogeneous independent Bernoulli opportunities with probabilities p_j, the resulting Poisson-binomial count satisfies E[N]=sum_j p_j and Var(N)=sum_j p_j(1-p_j), again no larger than the mean. This example demonstrates that heterogeneity in independent Bernoulli probabilities does not automatically create overdispersion: the observation structure and conditional event model decide the variance behavior.

Poisson arrivals and likelihoods

The Poisson distribution with mean μ>0 assigns P(N=k)=exp(-μ)μ^k/k! to each nonnegative integer k, with E[N]=Var(N)=μ. Its probability-generating function is G_N(z)=exp(μ(z-1)). If N_1,...,N_m are independent Poisson counts with means μ_1,...,μ_m, then their sum is Poisson with mean Σμ_i. This closure property makes the distribution especially attractive for forecasting an independently exposed portfolio.

For independent observations n_i with means μ_i, the Poisson log-likelihood (apart from a constant) is Σ_i[n_i log μ_i-μ_i]. Its score for a log-linear model produces interpretable estimating equations and a practical maximum-likelihood fit. The Poisson probability at zero is exp(-μ), which is often a severe benchmark for data containing many policies with no reported claim. The benchmark should be evaluated conditionally on covariates and exposure; a portfolio-wide single rate cannot serve as the relevant conditional baseline.

The original GLM formulation of Nelder and Wedderburn (1972) provides a unified approach to linking distributional means and linear predictors, while McCullagh and Nelder (1989) systematize GLM fitting and diagnostics. The use of a Poisson response has two different rationales: it can specify the full conditional distribution and likelihood, or it can specify a conditional mean structure whose coefficients may remain interpretable even when the variance law is wrong. In the latter use, inference needs appropriate sandwich or other robust treatment, and a full predictive distribution cannot be inferred from the mean equation alone.

Regression, offsets, and actuarial interpretation

Let X_i be the covariate vector and β a vector of coefficients. The log-link Poisson regression with a known exposure e_i>0 takes the form log μ_i=log e_i+X_i'β. The term log e_i is an offset with fixed coefficient 1. It encodes the claim that doubling comparable exposure doubles the expected number of claims. Each covariate coefficient β_j represents a multiplicative effect exp(β_j) on the conditional event rate per exposure unit, holding the other predictors fixed.

An alternative formulation models observed frequency f_i=n_i/e_i with expected value r_i=exp(X_i'β) and uses e_i as a frequency-model weight. Under the standard Poisson GLM formulation, this yields the same fitted score equations as count modeling with offset log e_i. The CAS explanation of offsets and weights also emphasizes the limitation: this equivalence does not automatically extend to a negative-binomial model with its separate dispersion or size parameter. It is essential to document whether the response is a count or a frequency and how software interprets weights.

Consider two groups. Group A contributes 100 vehicle-years with 12 claims; group B contributes 800 vehicle-years with 72 claims. Their crude rates are 0.12 and 0.09. A pooled constant-rate Poisson maximum-likelihood estimate is (12+72)/(100+800)=84/900=0.093333... per vehicle-year. The arithmetic average of the two group rates is 0.105, which targets the mean of two groups given equal group weights, not the rate for a randomly selected vehicle-year in the observed portfolio. Choice of denominator and weighting therefore defines the estimand rather than merely the precision of presentation.

If e_i=0, log e_i is undefined and the row must be handled by explicit inclusion rules; adding an arbitrary epsilon creates fictional risk time. Negative exposures indicate an error or reversal convention requiring reconciliation. Exposures measured with error can cause biased rate estimates because the offset is treated as known. If the measured denominator is actually a predictor of loss intensity, or if high-risk units accumulate systematically less exposure, a fixed-coefficient offset may require a substantive audit.

Sampling variability and model assumptions

The Poisson rate estimator λ_hat=(Σ_i n_i)/(Σ_i e_i) for a constant rate has variance approximately λ/(Σ_i e_i) under independent Poisson counts. Using the synthetic total of 84 claims and 900 vehicle-years gives an estimated standard error sqrt(84)/900≈0.01018. This quantifies sampling uncertainty conditional on the assumed independent Poisson structure and correctly measured exposure; it excludes systematic error, latent heterogeneity and changing conditions.

An exposure total of 900 does not create 900 independent trials if the units share weather, geography, employers, roads, reporting systems or latent risk factors. Correlation changes the uncertainty of the rate and sometimes the forecast distribution. A precise point estimate under an invalid independence model may falsely imply a narrow interval.

The central scientific question is therefore whether the Poisson model's mean and conditional variance are coherent for the defined data, not whether the observed empirical variance of all individual records happens to equal the empirical mean. Different e_i or X_i imply different μ_i; pooling heterogeneous means produces a marginal variance different from their conditional variances even when every observation is conditionally Poisson. A valid dispersion analysis conditions on the fitted mean and respects exposure differences.

Mixed-Poisson Frequency

Heterogeneity and the law of total variance

Let Λ be a nonnegative random intensity, and suppose N|Λ follows a Poisson distribution with mean Λ. The marginal count distribution is called mixed Poisson. Its expectation and variance follow directly from iterated expectation and the law of total variance:

E[N]=E[Λ], and Var(N)=E[Λ]+Var(Λ).

The second term records heterogeneity in the intensity across otherwise unobserved states. If Var(Λ)>0, the marginal count is overdispersed relative to a Poisson distribution with the same mean. This identity does not establish the cause of a given empirical dispersion excess; it describes a mechanism under a specific mixture construction. A mixture may represent variation among policies, a common calendar environment, persistent risk traits or an unobserved covariate, and those interpretations produce different dependence assumptions.

For example, suppose two equally likely types have Poisson intensities 0.1 and 0.5 per period. The marginal mean is 0.3 and the intensity variance is ((0.1-0.3)^2+(0.5-0.3)^2)/2=0.04. Thus Var(N)=0.34 and the marginal variance-to-mean ratio is approximately 1.1333. Treating the entire portfolio as a single Poisson(0.3) predicts the correct average count but a smaller variance. If type is observed and modeled, however, both conditional groups can still satisfy Poisson mean–variance equality.

This distinction matters in actuarial underwriting: explaining an observed marginal excess with covariates can reduce unexplained variance without changing the underlying physical event process. It also means that fitting a negative binomial to pooled counts is not evidence that every individual claim process has intrinsic clustering.

Probability generating functions and latent rates

The generating function of a mixed-Poisson count is G_N(z)=E[exp(Λ(z-1))]. Its factorial moments satisfy E[(N)_r]=E[Λ^r] when these moments exist, where (N)_r=N(N-1)...(N-r+1). This provides a direct mathematical bridge between moments of the latent intensity and factorial moments of the count. In particular E[N(N-1)]=E[Λ²], so Var(N)-E[N]=Var(Λ).

The generating-function form is also a warning about identification. Different mixing distributions can yield different tails and higher factorial moments even when they share the same first two moments. A diagnostic that checks only mean and variance cannot choose among these mixtures. Adequate prediction of extreme counts requires additional distributional evidence and, for scarce extremes, explicit uncertainty.

Hausman, Hall, and Griliches (1984) developed Poisson-based count models allowing persistent individual effects and extra noise in count intensities. Their empirical application concerned innovation patents rather than insurance claims, so its evidential role here is methodological: it illustrates how an unobserved source of variation can be introduced into the Poisson mean. A separate source of evidence is needed to claim that the same heterogeneity structure applies to a given insurance line.

Shared effects, panels, and conditional dependence

Suppose policy i has a stable unobserved positive risk multiplier A_i with E[A_i]=1 and Var(A_i)=α, and its counts N_it are conditionally independent Poisson variables with means A_i e_it r_it in periods t. Then, for a fixed period, E[N_it]=e_it r_it and Var(N_it)=e_it r_it+α(e_it r_it)^2. For two distinct periods t and s on the same policy,

Cov(N_it,N_is)=α(e_it r_it)(e_is r_is),

when the model's conditional independence assumptions hold. Heterogeneity thus creates positive longitudinal dependence even if the events are independent after conditioning on A_i. Refitting separate independent negative-binomial observations can match each period's marginal dispersion while discarding this cross-period structure.

If instead a single environmental factor A_t affects all policies in period t, the model creates contemporaneous covariance across policies; this is a different portfolio capital and uncertainty problem. The distinction between policy-level random effects and shared catastrophic or economic shocks must be made before summing frequency predictions. In one construction the covariance is longitudinal within policy; in another it is cross-sectional across policies.

The methods of Bühlmann and Gisler (2005), developed in credibility theory and its applications, provide a broader actuarial setting for combining experience with structural assumptions. A history of claims can inform an unobserved propensity, but a low count in a small exposure window may provide weak evidence. Hierarchical models and credibility methods should account for the amount of exposure and the uncertainty of latent risk characteristics.

Identification, credible experience, and aggregation

When an insurer observes one count per policy over a short period, it may be impossible to distinguish stable individual frailty from random calendar variation using that record alone. Repeated years, independently measured covariates, geographic structure, and a credible exposure history can improve identification. Experience rating based on past counts also raises selection effects: policyholders with very high counts may cancel, change carriers or receive altered coverage, so the continuing cohort is not an unchanged random sample.

For a portfolio with independent latent multipliers A_i and conditionally Poisson counts, summing counts yields E[N_P]=Σμ_i and Var(N_P)=Σ(μ_i+α_i μ_i²). If the multipliers are dependent, covariance terms must be added. The same number of claims and same individual marginal variance can therefore produce very different distributions for portfolio totals depending on the dependence architecture.

The operational decision is to preserve the exact level at which a random effect is assumed to operate: event, policy, fleet, location, insurer, cohort or calendar period. The broader term "heterogeneity" is scientifically insufficient as a specification. Each level implies a different observable prediction and a different strategy for evaluation.

Negative-Binomial Counts

Gamma–Poisson derivation

Take a latent intensity Λ with gamma distribution of shape k>0 and scale θ=μ/k, so that E[Λ]=μ and Var(Λ)=μ²/k. Conditional on Λ, let N be Poisson(Λ). Integrating the conditional Poisson probabilities against the gamma density gives

P(N=n)=Γ(n+k)/(Γ(k)n!) (k/(k+μ))^k (μ/(k+μ))^n, n=0,1,2,...

The resulting negative-binomial distribution has E[N]=μ, Var(N)=μ+μ²/k, and zero probability P(N=0)=(k/(k+μ))^k. Writing α=1/k gives Var(N)=μ+αμ². The Poisson limit is approached as k→∞, equivalently α→0, with μ fixed.

One derivation is especially informative. The Poisson probability-generating function conditional on Λ is exp(Λ(z-1)). The gamma moment-generating function gives E[exp(Λ(z-1))]=(1+θ(1-z))^(-k). Substituting θ=μ/k yields G_N(z)=(1+(μ/k)(1-z))^(-k). Differentiation at z=1 recovers the mean and factorial second moment. This derivation proves that the specified gamma mixture creates the negative-binomial form; it does not prove that an empirical negative-binomial fit uniquely identifies gamma-distributed physical propensity.

Hilbe's treatment of negative-binomial regression (2011) documents the diversity of parameterizations and their implications. Reporting "negative binomial" without the definition of the mean, dispersion parameter and variance law leaves the probabilistic model underspecified. The gamma–Poisson form above is the NB2 variance convention.

Parameterizations and changing exposure

For μ_i=e_i r_i and NB2 dispersion α, Var(N_i)=e_i r_i+α(e_i r_i)^2. Doubling exposure increases the first term linearly and the second quadratically. This is coherent if the extra-Poisson heterogeneity represents a stable policy-specific multiplier over the increased exposure horizon. It is a substantive assumption that can be false if the additional exposure consists of independent risk fragments or the unobserved intensity changes rapidly.

Another convention uses a variance μ(1+δ), sometimes termed NB1, with an approximately linear mean–variance relation. A regression using NB1 and a regression using NB2 can therefore allocate uncertainty differently across large and small exposures. Neither label by itself supplies a unique generative process. Some software also defines a "size" or "shape" parameter k and other packages use its reciprocal α. Numerical comparisons require the chosen parameterization to be stated.

The count-offset and frequency-weight equivalence established for Poisson is particularly fragile here. If a normalized response N/e is passed with an exposure weight, the software's negative-binomial variance and likelihood weights may define a different estimation problem than the gamma–Poisson likelihood for integer counts with offset log e. The CAS frequency-model discussion explicitly warns of this difference. A reproducible analysis must show the response, offset, weight semantics and exact conditional variance function.

Consider a policy with annual expected count r=0.1 and exposure e=2, so μ=0.2. Under NB2 with α=0.5, the variance is 0.2+0.5(0.2)^2=0.22. The negative-binomial zero probability is (1+αμ)^(-1/α)=1.1^(-2)≈0.826446. Under Poisson(0.2), P(N=0)=exp(-0.2)≈0.818731. The same mean produces different zero and tail probabilities, and the differences grow with dispersion and horizon.

Distributional comparisons and counterexamples

A Poisson and a negative binomial can share the same mean while giving different probabilities of zero, one, or several claims. The negative binomial's variance exceeds its mean in the NB2 form; the binomial variance is below its mean under independent trials. This comparative fact helps choose candidate models, but it must not be turned into a mechanical rule that every high empirical variance requires a negative binomial.

An exact counterexample shows why matching two moments is insufficient for tail prediction. Let X equal 0 with probability 1/2 and 2 with probability 1/2. Then E[X]=1 and Var(X)=1, the same mean and variance as Poisson(1). Nevertheless P(X>=3)=0, while Poisson(1) assigns a strictly positive probability to three or more events. No test using only the first two moments can distinguish their upper tails. Here the finite support and two-point construction are deliberate; it is a mathematical demonstration, not a proposed insurance count model.

A second construction shows that a zero-heavy portfolio does not uniquely identify a separate structural-zero class. For NB2 with μ=1 and α=1, P(N=0)=1/2, compared with exp(-1)≈0.3679 under Poisson(1). A zero-inflated Poisson can also produce elevated zero probability, but with a different positive-count distribution and causal interpretation. The zero frequency alone cannot select between the two.

Likelihood criteria should be compared only on coherently defined likelihoods and the same observation sample. A quasi-Poisson fit specifies a variance multiplier and may provide coefficient estimates and uncertainty adjustments, but it does not supply the same full probability distribution for likelihood comparisons. A model may improve aggregate likelihood yet worsen extreme-count calibration or stability for a new cohort. These are different dimensions of validation.

Overdispersion testing and alternative counts

For conditional Poisson regression, the variance equals the fitted mean. Cameron and Trivedi (1990) developed auxiliary-regression dispersion tests that assess departures through a specified mean–variance relation. Dean and Lawless (1989) proposed tests for extra-Poisson variation, and Dean (1992) examined overdispersion in Poisson and binomial regression. These results belong to related but distinct test constructions; a reported dispersion statistic should identify its null hypothesis, alternative, and sampling assumptions.

The Pearson statistic X²=Σ_i (n_i-μ_hat_i)^2/μ_hat_i is often divided by nominal residual degrees of freedom as a rough conditional dispersion diagnostic. An elevated ratio suggests remaining extra variation, but it is not a substitute for checking the data structure, correlation, influential observations, exposure, incomplete reporting and fitted mean specification. Small expected counts, estimated dispersion parameters, unbalanced exposures and correlation can distort simple asymptotic approximations.

Underdispersion has its own explanations: finite Bernoulli opportunity sets, event capping, negative association, constraints on simultaneous claims or narrower conditional distributions. Efron (1986) described double exponential families allowing additional variance flexibility within regression settings. A negative-binomial model, whose NB2 variance is at least its mean, cannot represent a genuinely underdispersed conditional count law by choosing an allowable positive α. Model selection should therefore preserve the sign and the mechanism of any dispersion departure.

A more demanding analysis considers residuals stratified by fitted mean and exposure, the count distribution within comparable risk cells, changes of fit with observation window, predictive tail frequencies, and the persistence of residual dependence. Evidence from any one test does not uniquely select a generating process.

Exposure and Zero-Inflation

Denominator construction and calibration

Exposure denominators are a central source of structural error. An insurer may store a policy as one record even when the insured risk changes during the period. A building can change occupancy, a driver can use a vehicle differently, and a business may expand or contract. If frequency is stated per policy-year while the true opportunity is proportional to transactions or distance traveled, a stable estimated rate per policy-year may conceal a changing underlying hazard.

Choose an exposure unit based on the intended risk and its observability. When available, disaggregate a policy into homogeneous segments with measured e_ij and intensity r_ij. The expected count is Σ_j e_ij r_ij, whereas replacing the segments by one annual rate assumes a valid exposure-weighted average. Aggregation may be harmless for a single expected total while being damaging for event timing, covariate attribution or within-policy dependence.

Calibration can be checked by grouping records into predicted-risk bands and comparing Σn_i with Σμ_hat_i, while retaining exposure totals. Plotting average observed frequency per unit exposure against predicted rate makes the denominator explicit. A risk-decile chart based on individual rows can be misleading if one decile contains mostly short policies and another mostly full-year policies; the relevant denominator is total observed opportunity.

The CAS monograph's model-building and validation guidance addresses the need to reconcile datasets, define targets, examine actual-versus-predicted relationships and track model stability. Statistical validation is only as coherent as the policy and claim data preparation that precedes it.

Excess zeros and zero-inflated Poisson

Suppose a policy belongs to a structural-zero class with probability π, 0<=π<1, and otherwise produces a Poisson count with mean λ>0. A zero-inflated Poisson (ZIP) variable has P(N=0)=π+(1-π)exp(-λ) and, for n>=1, P(N=n)=(1-π)exp(-λ)λ^n/n!. Its mean and variance are

E[N]=(1-π)λ, and Var(N)=(1-π)λ+π(1-π)λ².

The derivation applies the law of total variance to a latent two-state mixture. The second component of variance reflects uncertainty about latent zero-class membership. Lambert (1992) formalized this specification in zero-inflated Poisson regression, using manufacturing defect data as the original application. The transfer to insurance is a methodological analogy, not evidence that insurance zero counts necessarily arise from permanently immune policyholders.

The phrase "structural zero" carries a causal commitment. It may be appropriate for policies with genuinely inactive coverage during some state, units with no physical opportunity for the counted event, or a verified administrative class that cannot generate the measured claim. But many zero-claim policies simply had no event during their finite window. A short-term Poisson(0.1) risk produces no claim with probability roughly 0.9048 without any structural-zero process.

ZIP parameters should be linked to evidence beyond the observed zero proportion. Time repetition, known eligibility, risk states, external exposure indicators and out-of-sample behavior can help distinguish excess zeros caused by population composition from ordinary rare-event randomness. ZIP should not become a descriptive label substituted for a mechanism.

Hurdle, negative-binomial, and underdispersed alternatives

A hurdle model separately defines the probability of crossing from zero to a positive count and the conditional distribution of positive counts. If P(N>0)=q, and f_+(n) is a distribution on n>=1, then P(N=0)=1-q and P(N=n)=q f_+(n) for n>=1. With a zero-truncated Poisson positive component, f_+(n)=exp(-λ)λ^n/[n!(1-exp(-λ))]. The presence of a zero is determined entirely by q; every member of the positive component necessarily crosses the hurdle.

A ZIP model differs because the Poisson component itself can generate a zero. A hurdle model can therefore be appropriate when claim initiation and repeat claim count reflect separable processes, provided the data support that distinction. Mullahy (1986) studied modified count-data specifications, providing historical context for the methodological development. The original paper's application was outside insurance, and its count-model distinctions are used here independently of that domain.

A zero-inflated negative-binomial model combines a latent zero class with an overdispersed nonstructural count. Its flexibility can improve in-sample fit, but additional components demand stronger evidence and create estimation challenges. If the finite opportunity model is binomial or counts are constrained, an underdispersed distribution may be preferable. In each case the scientific question is what process produces the observed frequency, rather than how many parameters a flexible family can absorb.

Zero probabilities should be evaluated together with the full positive-count spectrum. A model that matches the proportion with no claims but mispredicts single-claim and multi-claim policy counts may be unreliable for expected claims or claims operations. A positive-count tail is particularly influential when high-frequency policyholders are rare.

Measurement, reporting, and structural zeros

Insurance systems create zeros through business rules as well as stochastic event incidence. A policy may have no recorded claim because the event occurred after the observation cutoff, because the policy was not in force, because the event was handled outside coverage, or because a claim fell below an administrative reporting threshold. Such zeros are observational categories that may require data correction or a separate observation model.

Let N* be the true number of qualifying occurrences, and let each occurrence be reported independently with probability q. Under Poisson thinning, N_obs|X has mean qμ if N*|X is Poisson(μ). If q differs by distribution channel, coverage or calendar period, a frequency regression may attribute differences in claims reporting to risk classification. With independent detection, the Poisson family cannot separate q from μ without additional information. If report probabilities depend on event severity, the observed frequency may also be selectively connected to the separate severity distribution.

Causal stories need to be checked against claim-system design. An apparent structural-zero segment might actually be a portfolio subgroup with no recorded exposure, inconsistent policy identifiers, or a different data pipeline. A standard validation program audits record reconciliation and data-generating mechanisms before preferring a zero-inflated likelihood.

A related caution concerns the direction of causality. An insurer's claims handling, deductibles and notification requirements can influence the observed number of claims independently of physical risk. Frequency estimates used to evaluate future product designs must account for possible behavioral and contractual responses rather than assume a fixed reporting mechanism.

Validation and Use

Research design, estimation, and benchmarks

A defensible modeling study begins by recording the estimand, event definition, exposure unit, development horizon, population, admissible covariates, cohort selection rule and intended decision. It should preselect benchmark distributions and document a validation split aligned to the prospective use. For instance, training on earlier accident years and validating on a later accident year evaluates calendar transport more directly than a random split that distributes the same unusual weather episode across both partitions.

Baseline models should include a constant-rate Poisson with exposure offset, a covariate-adjusted Poisson, and a properly specified negative-binomial candidate where overdispersion is plausible. Depending on the process, add mixed effects, time-varying intensity, hurdle or zero-inflated models. A carefully documented simple model often provides more reliable information about temporal stability than a highly parameterized fit without a coherent deployment target.

The full likelihood is suitable for comparing properly normalized distributional predictions; mean-only quasi-likelihood estimates require a different evaluation logic. Report the likelihood convention, estimator, treatment of group effects, uncertainty calculation and inference assumptions. Where observations are clustered, split data and compute uncertainty at the appropriate cluster level. Do not interpret an ordinary row-based confidence interval as a population-level guarantee when random effects are shared.

Diagnostic plots, scores, and out-of-sample testing

Evaluation should separately address the conditional mean, the probability of zero, the distribution over positive counts, the upper count tail, and the dependence of counts across units and periods. For means, compare observed-versus-expected counts across exposure-stratified and risk-stratified segments. For distributions, examine observed and forecast frequencies for n=0,1,2,..., using pooled bins only where sample support demands it. Assess whether the predicted number of zero-claim policies, multi-claim policies and large-count policies has reasonable held-out coverage.

The logarithmic score for a realized count n under a forecast mass function p is -log p(n), with lower values better when p(n)>0. This score evaluates the full predictive distribution and penalizes forecasts assigning very small probability to an observed event. A Poisson forecast calibrated only in the mean can therefore perform poorly when counts are overdispersed or have unexpectedly frequent zeros. Proper scoring assesses both sharpness and calibration, but the score should be reported by relevant cohorts to avoid a large, low-risk group dominating the aggregate.

A useful randomized probability integral transform for discrete counts is U=F(n-1)+V P(N=n), with V uniform on [0,1] independent of the observed count. Under a correctly specified predictive count distribution and corresponding conditional independence, U is uniform. The randomization is needed because an ordinary continuous PIT applied to integer counts is not uniform. Such diagnostics should be interpreted with independence and estimation effects in mind.

Residual comparisons can locate misspecified mean structure, missing interactions, nonlinearity and heterogeneity. Temporal plots can show drift that a global dispersion estimate hides. For multi-period histories, assess within-policy residual correlation and changes in the frequency of consecutive high-count years. A marginal negative-binomial distribution can fit annual histograms while failing these longitudinal tests.

Stability, governance, and deployment

A fitted frequency model depends on historical exposure mix, underwriting selection, claims reporting, prevention programs, contract triggers and economic conditions. Stability assessment therefore needs more than a holdout random sample. Repeat validation across calendar periods, territories, distribution channels, underwriting cohorts, policy age, exposure sizes and, where lawful and scientifically relevant, groups affected by the model's classifications.

The model documentation should distinguish predictive use from causal attribution. A covariate may improve forecasts without measuring an intervention effect. Claim counts can also be affected by coverage choices and incentive changes. Changing deductibles, claim forms or eligibility may alter the observable outcome and invalidate an apparently stable historical coefficient.

Governance records should specify data lineage, event definitions, the treatment of exposures with zero or erroneous values, model version, parameter estimates, sampling and distributional uncertainty, diagnostics, overrides, operational monitoring and rebuild triggers. The CAS monograph on rating models offers an applied point of reference, while general statistical foundations remain those of the GLM family. An insurance company deciding on real classification and pricing must additionally examine the law, regulatory guidance, consumer impacts and contractual definitions relevant to its own jurisdiction.

No single loss-frequency family has an exclusive scientific claim to future performance. A Poisson model can be appropriate when conditional means are well specified and counting opportunities are approximately independent. Gamma-mixed Poisson and negative-binomial models are suitable for some persistent heterogeneity, with explicit dispersion assumptions. Hurdle and zero-inflated models represent more elaborate zero-generating stories; binomial or other families may represent bounded event opportunities and underdispersion. The choice is conditional on the observed process and the intended predictive use.

Integrated worked examples and research implications

The following four calculations are synthetic and independently derived. They illustrate the inferential differences that matter in insurance without claiming empirical market frequency.

First, the two-group exposure sample above produced 84 claims over 900 vehicle-years. The pooled maximum-likelihood constant Poisson rate is 0.093333...; a 1,500-vehicle-year comparable prospective portfolio has a mean forecast of 140 claims. This extrapolation assumes identical effective risk composition and rate per vehicle-year. A new portfolio with a higher proportion of group A should instead combine group-specific rates and exposures.

Second, let N|Λ be Poisson(Λ), with Λ taking 0.1 and 0.5 with equal probabilities. The mean count is 0.3 and its variance 0.34. The zero probability is [exp(-0.1)+exp(-0.5)]/2≈0.75568, greater than Poisson(0.3)'s exp(-0.3)≈0.74082. The distinction arises from mixing, without adding a structural-zero class. A future portfolio in which every type is observed has different conditional probabilities from a randomly selected untyped record.

Third, with an NB2 count mean 0.2 and α=0.5, the variance is 0.22 and the zero probability is about 0.82645. The Poisson with the same mean has variance 0.2 and zero probability about 0.81873. Small-looking probability gaps can become important in large portfolios or when evaluating segment-specific zero rates, but this numerical contrast alone supplies no criterion for choosing one family from real data.

Fourth, define an insurance count X that is either 0 or 2 with equal probability. It has the same mean and variance as Poisson(1), but cannot ever equal 3. Thus exact matching of E[X] and Var(X) does not control tail probability P(X>=3). A model adequate for expected frequency cannot claim adequate tail calibration without an additional test. This is the central limitation of moment-only diagnostics.

The four cases connect to the primary research literature rather than assert an original distribution theory. The overdispersion testing literature clarifies what can be tested about the conditional variance; the negative-binomial regression treatment specifies extra dispersion families; the zero-inflation formulation separates a latent zero-generating process. Their integration supports a pragmatic inference: counting rules, exposure and reporting should be audited before attributing residual structure to latent risk types.

A research program capable of extending these results would combine linked occurrence and reporting data with repeated risk histories, independently measured exposure intensity, calendar hazard covariates and prospective distributional backtests. It would compare the ability of alternative models to predict changes in claim rates, zero counts and upper count frequencies after observable exposure changes. Such data would separate competing mechanisms more effectively than retrospective overall goodness-of-fit alone.

Conclusion

Loss frequency is a distributional property of a precisely defined observation process. Its basic variable is the count of insured occurrences or claims during specified exposure, and its interpretation depends on the identity of each counted event, the time window, contractual coverage, reporting mechanisms and population composition. The exposure denominator is part of the scientific definition of the estimand.

The binomial distribution describes finite independent event opportunities, including cases with conditional underdispersion. The Poisson distribution offers a coherent count law, event-process representation and regression mean structure under explicit assumptions. Mixed-Poisson models describe latent rate variation; the law of total variance explains how this creates marginal overdispersion. The gamma mixture yields the negative-binomial NB2 form with a variance that grows quadratically in the mean. Other variance conventions, correlation structures and zero-generating mechanisms require separate specification.

The analysis establishes four recurring methodological consequences. First, conditional and marginal frequency distributions must be distinguished. Second, reported frequency and physical occurrence frequency can differ through thinning, reporting delay and administrative selection. Third, matching conditional means or the first two moments does not establish adequacy of zero probabilities, count tails or multi-period dependence. Fourth, prospective validation should align with the actual actuarial decision and reflect exposure changes, calendar transport and operational reporting.

These results are grounded in established probability, econometrics and actuarial modeling. The original synthesis joins formal moment identities, explicit finite counterexamples, exposure-weighting analysis and source-specific methodological limits into a coherent review of claim-frequency inference. Its contribution is analytical and editorial; it establishes no new named theory. A research-ready frequency model is one whose target count, exposure, assumptions, parameterization, calibration and range of defensible use can be stated precisely and evaluated against observed evidence.

References