Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Generalized Linear Models in Actuarial Pricing

Abstract. Generalized linear models (GLMs) provide a common mathematical language for the frequency, severity, and expected aggregate cost of insured losses. Their actuarial value depends on a conjunction of probability-family assumptions, exposure measurement, contractual loss definitions, risk classification, estimation, validation, and the conversion of modeled costs into prospective rates. This monograph develops that conjunction from the exponential-dispersion representation through link functions, likelihood and quasi-likelihood, Poisson frequency, negative-binomial alternatives, Gamma severity, compound Poisson–Gamma/Tweedie costs, pricing relativities, diagnostic statistics, and model governance. It separates the interpretation of coefficients from the interpretation of causal effects and distinguishes conditional pure premiums from gross rates and regulatory outcomes. Independently derived identities and explicitly constructed numerical cases show how the same multiplicative factor can carry different meanings for count, severity, aggregate cost, and final premium; how a Tweedie distribution combines a mass at zero with positive continuous losses; and how exposure, reporting delay, selection, and distribution shifts modify the inferential target. International comparison uses contemporary specialist actuarial research, the adopted U.S. Actuarial Standards of Practice, European insurance-supervisory materials, and the NAIC's account of insurance AI governance. The original analytical contribution is a traceable, source-based synthesis of the mathematical and institutional transitions from a fitted regression to a defensible prospective insurance price. All numerical examples are synthetic analytical constructions. The work advances no newly named theory or proprietary empirical result.

Keywords: Generalized linear models; actuarial ratemaking; insurance pricing; exponential dispersion family; log link; offset; Poisson regression; overdispersion; negative binomial; Gamma severity; compound Poisson; Tweedie distribution; pure premium; rating relativities; maximum likelihood; iteratively reweighted least squares; quasi-likelihood; deviance; exposure; claim development; credibility; model validation; classification fairness; insurance governance.

Introduction: The Pricing Question and the Statistical Object

Insurance pricing seeks a defensible prospective transfer from a defined covered risk to a price for taking that risk over a specified period. A statistical regression can estimate conditional patterns in historical experience, but the actuarial task also identifies which exposures and claims belong to the covered contract, what development and trend adjustments are necessary, and which cost and financial provisions must accompany the projected loss cost. The starting observable response is thus never simply “risk.” It is a particular operational measure: claim counts, positive loss amounts, aggregate covered loss, or an explicitly adjusted prospective cost.

The modern GLM originated in the unified treatment of exponential-family observations and transformed linear predictors given by Nelder and Wedderburn (1972). The foundational statistical framework was subsequently developed in McCullagh and Nelder's second edition (1989). For nonlife insurance, Ohlsson and Johansson (2010) organized GLMs around tariff analysis, while the Casualty Actuarial Society's 2025 insurance-rating monograph treats the operational passage from policy and loss data to an implemented classification plan. The statistical foundations and the operational practice must be examined together.

The basic distinction is among a conditional expectation, a statistical estimator of that expectation, a prospective actuarial cost estimate, and an offered premium. They possess different units and respond to different information. An expected claim frequency of 0.12 counts per insured vehicle-year, an expected severity of $2,000 per covered claim, an expected annual loss cost of $240 per vehicle-year, and a final premium of $350 are four different statements. Their connections require explicit assumptions rather than a universal conversion factor.

This study reconstructs the probability theory, derives common estimators and cost identities, tests stylized numerical examples, and compares the resulting workflow with official professional guidance. It examines the statistical claims that generalize internationally, and assigns jurisdictional rules to their actual institutional sources. The scientific literature cited here comprises foundational original articles, specialized books and a directly inspected 2025 CAS monograph; some other original works are supported at publisher abstract or bibliography level, as specified in the evidence ledger. No insurer data set, policy-level personal record, market rate filing, or independent external peer review forms part of this research.

GLM Probability Families: Mathematical Foundations

A generalized linear model combines three mathematical components. The response has a distribution in an exponential family or a supported exponential-dispersion family; its expectation is related to covariates through a systematic linear predictor; and a link function maps the expectation to the predictor. The commonly used representation for one independent response is

[ f(y_i;\theta_i,\phi)=\exp\left{\frac{y_i\theta_i-b(\theta_i)}{a_i(\phi)}+c(y_i,\phi)\right}. ]

Here (\theta_i) is a natural parameter, (\phi) a scale or dispersion parameter, and (a_i(\phi)) commonly equals (\phi/w_i) for known positive weights (w_i). Differentiating the normalization identity with respect to (\theta) gives (\mu_i=E[Y_i]=b'(\theta_i)); differentiating a second time gives (\operatorname{Var}(Y_i)=a_i(\phi)b''(\theta_i)). Defining the variance function (V(\mu_i)=b''(\theta_i)) yields

[ \operatorname{Var}(Y_i)=\frac{\phi}{w_i}V(\mu_i). ]

These identities follow from differentiability and the regularity conditions under which differentiation may be interchanged with integration or summation. Their value is diagnostic: a regression specification chooses not only a functional shape for the mean but a relationship between mean and variance. Jørgensen's examination of exponential dispersion models (1987) develops the breadth and convolution properties of this class. The family representation does not determine whether an insurance outcome is accurately sampled from that family.

Common actuarial examples are Poisson counts with (V(\mu)=\mu); Gamma positive amounts with (V(\mu)=\mu^2); and Tweedie distributions with (V(\mu)=\mu^p). A Gaussian response has constant variance, and a binomial proportion requires an explicit number of trials or denominator. The negative-binomial model for overdispersed counts is an important regression alternative, but its mean–variance relation and parametrization must be written explicitly rather than silently treating it as a fixed-dispersion Poisson GLM.

The Poisson law assumes nonnegative integers. A Gamma law puts probability only on strictly positive outcomes. A compound Poisson–Gamma distribution for aggregate cost allocates positive probability to exactly zero claims and places a continuous distribution over positive aggregate amounts. Confusing these support restrictions leads to a structural problem that small changes to the link function cannot repair. A model whose data contain exact zeros cannot use an ordinary Gamma likelihood for the entire annual loss-cost variable.

The variance function also controls information. If two cells have the same expected annual loss but one is based on many independent units and the other is based on a few high-volatility risks, treating both cells as equally precise ignores exposure and dependence. Known statistical weights and actual insurance exposures are related, but they are not interchangeable by definition. An exposure offset constrains the mean's proportionality to time or insured units; a precision weight contributes to the variance or objective. Their roles must be identified separately.

Likelihood, Estimation, and Statistical Identification

For conditionally independent observations, estimation commonly maximizes the summed log likelihood (\ell(\beta)=\sum_i\ell_i(\beta)). Write (\mu_i=h(\eta_i)=g^{-1}(\eta_i)). Under standard GLM regularity conditions the score for a coefficient is

[ U_j(\beta)=\sum_i\frac{w_i(y_i-\mu_i)}{\phi V(\mu_i)}\frac{d\mu_i}{d\eta_i}x_{ij}. ]

For an independent Poisson log-link count model with a correctly specified exposure offset, (V(\mu)=\mu), (d\mu/d\eta=\mu), and (\phi=1), so (U(\beta)=X^T(y-\mu)). The expected negative Hessian becomes (X^T\operatorname{diag}(\mu)X). Its nonsingularity requires sufficient design-matrix rank and positive information in the relevant directions. A dummy variable that is an exact sum of other dummies is aliased; highly correlated yet nonidentical predictors can still have identifiable but unstable separate coefficients.

Iteratively reweighted least squares solves a succession of local weighted regression problems. Define a working response (z_i=\eta_i+(y_i-\mu_i)/(d\mu_i/d\eta_i)) and working weight (W_i=w_i(d\mu_i/d\eta_i)^2/[\phi V(\mu_i)]). Subject to the chosen link, dispersion treatment, and numerical stabilization, a weighted least-squares update approximates a Newton or Fisher-scoring step. This is the computational insight emphasized by Nelder and Wedderburn's original account (1972). A converged numerical optimizer is only evidence that an objective has been optimized locally; it does not prove that the intended exposure, outcome, or prediction target was correctly defined.

A quasi-likelihood approach replaces a fully specified response density with an estimating relationship between mean and variance. Wedderburn (1974) established the classical quasi-likelihood formulation. It can preserve a useful mean estimating equation even when full-distribution claims would be unwarranted. It does not automatically supply a valid likelihood for likelihood-ratio comparisons, an exact probability of zero, or a fully specified predictive loss distribution. For those tasks a probability model or another justified inference procedure is needed.

Coefficient uncertainty is usually calculated from information matrices or robust sandwich estimators. When exposures share households, policies, fleets, locations, accident events, or calendar shocks, treating observations as independent can understate uncertainty. Cluster-robust covariance or hierarchical models may be appropriate, but the cluster definition must match the dependence source. Statistical significance of a coefficient neither establishes causality nor its usefulness for rating. In large insurance samples, a small and economically immaterial effect may be highly statistically significant.

Exposure, Contracts, and the Actuarial Data Unit

Exposure is the denominator that makes observed claims interpretable. A one-year motor policy, a two-month endorsement, a per-mile telematics contract, and a commercial building insured for a partial year do not necessarily create equivalent exposure units. The measurement must match the peril and cost being modeled. Annual vehicle-years can be a reasonable count offset for claims occurring approximately proportionally in covered time; insured values or payroll can be relevant rating bases in other lines, but proportionality must be justified.

A record should define the policy and coverage, observation period, effective and transaction dates, earned exposure, risk characteristics as they were known when the risk was accepted, claim event date, reporting date, loss valuation date, and the cost definition. The target may be reported claims, ultimate claims, paid loss, incurred loss, limited loss, or loss plus allocated expenses. These targets have different censoring, delay, and development properties. The insured risk classification is conditional on what was actually covered, including exclusions, limits, deductibles, and changes in policy form.

The CAS 2025 monograph's data-preparation chapters specifically cover joining policy and claim information, splitting modeling data, selecting targets and validating GLMs. For regulatory actuarial practice, ASOP No. 23, Data Quality distinguishes selecting, reviewing, and using data and disclosing material limitations. These requirements are especially material when a policy-claim join produces duplicate losses, when exposure is missing, or when claim development dates leak information unavailable at the time of prospective pricing.

Consider a driver whose policy was active for half a year. If the modeled claim count is one and the exposure is 0.5 vehicle-years, the observed annualized count rate is two, but the count contribution to a Poisson likelihood remains one, with (\log 0.5) as an offset. Treating the annualized rate of two as an unweighted Poisson count creates an erroneous likelihood. Similarly, a policy with no claims remains in the frequency data; it does not enter an individual positive-claim severity model as a zero-dollar Gamma observation.

Calendar leakage deserves special care. A model evaluated on random policy rows may gain information about the same event or the same risk from nearly identical rows appearing in training and test sets. A time-based prospective test can reveal deterioration masked by random folds. Claims for the latest accident periods may have systematically less time to be reported or settled. If an insurer estimates reporting-delay corrections and then uses developed ultimate cost as the GLM target, the uncertainty of the development step also belongs in the overall actuarial assessment.

Poisson Frequency Models

For risk unit (i), let (N_i) be the number of covered claims during exposure (e_i>0). A baseline model is (N_i\mid x_i,e_i\sim\mathrm{Poisson}(e_i\lambda_i)), where (\lambda_i) is the rate per unit exposure. With a log link, the specification is

[ \log E[N_i\mid x_i,e_i]=\log e_i+\beta_0+x_i^T\beta. ]

Thus (\log e_i) is a genuine offset: its coefficient is fixed to one. A policy exposed for one-half of a vehicle-year has half the expected count of an otherwise identical policy exposed for one year, provided homogeneous incidence in time and no change to risk attributes. The model assumes (\operatorname{Var}(N_i\mid x_i,e_i)=E[N_i\mid x_i,e_i]); it assigns (P(N_i=0)=e^{-e_i\lambda_i}). Both statements are empirically testable restrictions, not consequences of the word “frequency.”

For one group with constant rate (\lambda), total claim count (C=\sum_iN_i) and exposure (E=\sum_i e_i), differentiating (\ell(\lambda)=C\log\lambda-E\lambda+\text{terms independent of }\lambda) gives the maximum-likelihood estimate (\widehat\lambda=C/E), whenever (E>0). The fitted total expected count equals the observed total in that simple intercept-only situation. That equality can also occur within independently estimated strata, but it should never be promoted into an unconditional property of any complicated GLM, especially under regularization or offsets with different weights.

Take 80 claims during 1,000 vehicle-years. The annual observed frequency is 0.08; the model's predicted probability of no claim in a full year is (e^{-0.08}), approximately 0.9231. The expected count of 0.08 is not the probability of one or more claims: the latter is (1-e^{-0.08}), approximately 0.0769. Equality holds only as a small-rate approximation. For accident, theft, liability, and health-visit applications, the choice between an event count and an indicator of any event changes both the distribution and the estimand.

A cause-specific or coverage-specific count may be modeled separately. If claims from distinct perils overlap in coverage, their component counts need not be independent. Combined policy claims can also reflect reporting conventions: a single catastrophe may generate several line items or one aggregated occurrence. A valid model must carry the definition of “claim” into both the observed response and the ultimate pricing calculation.

Count Heterogeneity and Alternative Frequency Models

Frequency data frequently exhibit more variation than a conditionally Poisson model predicts. Heterogeneity in unobserved risk propensity, event clustering, repeated-claim dependence, imperfect exposure measurement, and omitted variables can all produce overdispersion. The empirical Pearson ratio (\sum_i(y_i-\widehat\mu_i)^2/\widehat\mu_i) divided by residual degrees of freedom is one diagnostic, although its interpretation depends on grouping, leverage, sample structure, and model assumptions.

An established alternative is a negative-binomial count specification with (E[N_i]=\mu_i) and (\operatorname{Var}(N_i)=\mu_i+\kappa\mu_i^2) for (\kappa>0). This variance arises, for example, when a Poisson intensity is multiplied by a unit-level positive Gamma frailty with mean one and variance (\kappa). The shape of the variance depends on what the frailty is shared across: for a single stable policy-period frailty, longer exposure increases the quadratic term in the period mean; independent short exposure segments with new frailties behave differently. That distinction is important when exposures are aggregated.

A negative-binomial specification changes predictive count probabilities and tail behavior. A quasi-Poisson specification retains (E[N]=\mu) and inflates conditional variance to (\phi\mu), but does not by itself give a full probability distribution. Zero-inflated and hurdle models allocate a separate process to zeros or to the positive-count probability. They may fit excess zeros, yet excess zero counts can also arise from inadequate covariates, exposure units, or mixing of insured subpopulations. The model choice should follow a scientific account of the data-generating process and out-of-sample evaluation.

Count models become particularly fragile under catastrophe or common-shock clustering. A regional storm can create correlated claims across policies in the same territory and event period. Even if each policy's marginal mean is well predicted, aggregation of independent Poisson predictions can underestimate portfolio volatility. The GLM's rating relativities remain useful conditional mean estimates; catastrophe scenario analysis and dependence modeling are complementary tasks.

The CAS distribution and offset treatment discusses practical frequency choices, while Jørgensen's dispersion-family analysis establishes why marginal means and variances must be understood as a pair. An assertion that a negative-binomial fit is always superior because it has a larger likelihood or extra parameter is unwarranted. Comparison must consider predictive purpose, penalization, nonnested model behavior, interpretability, and whether out-of-sample performance actually improves.

Gamma Severity and Positive-Loss Models

Let (X_{ik}>0) be the cost of positive covered claim (k) arising from risk unit (i). A Gamma regression models a positive conditional mean (m_i=E[X_{ik}\mid x_i,\text{claim observed}]) and a variance proportional to (m_i^2): (\operatorname{Var}(X_{ik}\mid x_i)=\phi m_i^2) under a conventional dispersion parametrization. A log link specifies (m_i=\exp(\alpha_0+x_i^T\alpha)), so severity relativities are multiplicative conditional mean ratios.

The model targets positive observed or developed claims, not the severity that would have occurred for policies with no claim. If higher-frequency risks are also more likely to produce larger claims, a severity fit based only on observed claims may estimate the claim-weighted severity distribution under that selection process. The resulting product of independently fitted conditional frequency and severity means is defensible under assumptions about the alignment of risk segments and conditional expectations; the assumptions must be articulated rather than inferred from the convenience of multiplying two regressions.

When several claims share an exposure unit, pooled claim-level Gamma regression may overrepresent high-frequency policies in a way that is appropriate for modeling mean cost per claim but inappropriate for estimating mean severity per policy. Alternatively, aggregate positive claim cost divided by claim count can be modeled at a risk-cell level with claim-count precision weights when the underlying assumptions justify them. Claim-count weights are not exposure offsets. Their role is to indicate how precisely each cell's mean severity has been observed.

Insurance severity often has exceptional upper-tail exposure, policy limits, deductibles, inflation, and large-loss mechanisms. A Gamma model may provide an adequate conditional mean for routine claims while materially understating extreme quantiles. The payment function is often (X=(L-d)_+\wedge u) for deductible (d), limit (u), and underlying ground-up loss (L); changing (d) or (u) changes the target itself. A single fitted historical severity coefficient does not automatically transport across different contracts or future inflation regimes.

The Gamma regression is a central case in Ohlsson and Johansson's tariff analysis (2010). The CAS 2025 GLM treatment of severity discusses the practical distribution and model-choice issues. A lognormal alternative implies a normal model for log severities, but its conditional original-scale expectation depends on the spread of those logarithms. Separate estimation of the conditional mean and tail remains an empirical modeling decision.

Compound Frequency–Severity and Tweedie Extensions

For a risk unit, total annual covered loss is (S_i=\sum_{k=1}^{N_i}X_{ik}), with the convention that an empty sum equals zero. If the conditional claim count (N_i) and severities (X_{ik}) satisfy a compound model with mutually independent and identically distributed severities of finite first and second moments, and are independent conditional on the covariates, then

[ E[S_i\mid x_i]=E[N_i\mid x_i]E[X_i\mid x_i], ]

[ \operatorname{Var}(S_i\mid x_i)=E[N_i\mid x_i]\operatorname{Var}(X_i\mid x_i)+\operatorname{Var}(N_i\mid x_i){E[X_i\mid x_i]}^2. ]

The law of total expectation and the law of total variance establish these identities. With Poisson frequency of mean (\lambda_i), the second expression reduces to (\lambda_i E[X_i^2]). For an exposure of (e_i), the annual expected total loss is (e_i\lambda_i m_i). If conditional claim count and conditional severities share latent factors not represented in (x_i), a product of marginal or improperly conditioned fitted means can be biased for expected aggregate loss. Correct exposure and conditioning are indispensable.

The Tweedie exponential dispersion family allows a direct model for a loss-cost response, often with a log link and (\operatorname{Var}(Y)=\phi\mu^p) for (1<p<2). In this range a Tweedie random variable has a point mass at zero and a continuous distribution on positive outcomes. It corresponds to a compound Poisson–Gamma construction with particular parameter relationships; (p) determines the power variance function. Dunn and Smyth (2005) studied the evaluation of Tweedie densities, including the numerical difficulties of density calculations and likelihood fitting.

With the conventional compound representation of a Tweedie random variable (Y) having mean (\mu), dispersion (\phi), and (1<p<2), the Poisson event rate is (\lambda=\mu^{2-p}/[\phi(2-p)]), Gamma event-shape is (a=(2-p)/(p-1)), and Gamma scale is (b=\phi(p-1)\mu^{p-1}). The probability mass at zero is (P(Y=0)=e^{-\lambda}). Direct substitution gives (\lambda ab=\mu), while the compound-Poisson variance is (\lambda a(a+1)b^2=\phi\mu^p). These formulas can be checked algebraically without estimating any company portfolio.

For an illustrative (\mu=200), (\phi=20), and (p=1.5), the Poisson event rate is (\sqrt{200}/10), approximately 1.4142; the Gamma shape is 1 and its scale is about 141.4214. Hence (P(Y=0)=e^{-1.4142}), approximately 0.2431; the expected amount is 200; and the variance is (20(200)^{1.5}), approximately 56,568.54 in squared monetary units. This is a constructed demonstration of the family parametrization, not an estimate of any actual claim frequency or portfolio variance.

A direct Tweedie cost model and a frequency–severity model answer different questions about transparency and calibration. The two-part approach exposes distinct frequency and severity drivers and can handle separate development assumptions, while a direct cost model fits an aggregate conditional mean and variance with a coherent zero atom. Neither approach automatically enforces compatibility with contract limits or guarantees superior prediction. The insurance GLM applications chapter (Ohlsson and Johansson, 2010) and CAS's revised monograph discuss pure-premium models in actuarial practice.

Relativities, Marginalization, and Rate Indications

For frequency with log-link coefficients (\beta), define (\lambda(x)=\exp(\beta_0+x^T\beta)). For severity with corresponding coefficients (\alpha), define (m(x)=\exp(\alpha_0+x^T\alpha)). Under the compound assumptions, the annual conditional pure premium is (q(x)=\lambda(x)m(x)=\exp[\beta_0+\alpha_0+x^T(\beta+\alpha)]). A risk factor's pure-premium relativity is therefore the product of its frequency and severity relativities when both models use aligned definitions, factors, and no relevant interactions omitted from that expression.

Suppose an indicator has conditional frequency relativity 1.50 and severity relativity 1.10. Its compound expected-loss relativity is 1.65. It does not follow that an insurer's offered premiums must have a factor of 1.65: expense allocations, discounts, minimum premiums, territorial constraints, policy fees, reinsurance costs, risk-load policy, and jurisdictional limits can modify the rate structure. A factor in an internally estimated statistical pricing model and a factor in a filed or approved rating manual are different institutional objects.

Conditional and marginal comparisons also differ. A fitted coefficient for a territory controls for the other covariates as specified. An observed territory loss ratio averages over different mixes of vehicles, drivers, deductibles, and policy durations. If the mix differs, its ratio is not the same thing as a conditional territory factor. Portfolio-level prospective cost is computed by summing expected costs over the actual or projected exposure distribution: (\sum_i e_i q(x_i)). Multiplying a single unweighted “average relativity” by the number of policies generally fails when exposure and covariate compositions vary.

Rate indication asks how projected prospective loss, expense and other provisions relate to expected premium at current rates. In an illustrative pricing equation with a fixed expense (F) per exposure unit, a variable expense fraction (v) and a target underwriting profit fraction (r), the gross premium might be expressed as (P=(q+F)/(1-v-r)), assuming the denominator is positive and the chosen definitions are consistent. This equation represents a specified illustrative rate-construction convention; it is not a general legal prescription. ASOP No. 53's prospective-cost guidance addresses exposure bases, classification, loss development, trend, expenses, credibility, catastrophe, reinsurance and profit/contingency provisions in its professional scope.

A log-linked regression enables a coherent conditional comparison but cannot determine normative choices about affordability, permissible discrimination, or the financing of shared risk. Statistical differentiation and acceptable institutional differentiation must be evaluated through their respective evidence and legal requirements.

Worked Synthetic Actuarial Pricing Cases

A first constructed example has four equal 1,000-vehicle-year exposure cells formed by two binary characteristics: territory A/B and vehicle class standard/high. Claim counts are deliberately set to exhibit a purely multiplicative Poisson mean; the data are an algebraic test case and contain no sampling-based claim of fit to real experience.

Territory Vehicle class Vehicle-years Constructed claims Fitted annual frequency
A Standard 1,000 80 0.080
A High 1,000 100 0.100
B Standard 1,000 120 0.120
B High 1,000 150 0.150

Using A/standard as the baseline, the frequency intercept is (\log(0.08)), the B territory coefficient is (\log(1.50)), and the high-vehicle coefficient is (\log(1.25)). The B/high predicted frequency is (0.08\times1.50\times1.25=0.15). Counts equal fitted means cell by cell, so the additive log specification has zero residual discrepancy for this contrived four-cell table. Because the cell values were created to satisfy the model, this fit supplies no evidence of forecasting performance. In a real sample, random error, selection, and omitted factors would remain.

A second constructed component assigns base positive-claim severity of $1,200, a territory-B severity factor of 1.10, and a high-vehicle severity factor of 1.20. Under conditional independent frequency–severity construction, the baseline pure premium is (0.08\times1{,}200=$96) per vehicle-year. For B/high, severity equals (1{,}200\times1.10\times1.20=$1{,}584); pure premium equals (0.15\times1{,}584=$237.60). The pure-premium relativity is (237.60/96=2.475), which also equals (1.50\times1.25\times1.10\times1.20).

A third constructed calculation inserts a fixed expense of $20 per vehicle-year, variable expenses of 25% of gross premium, and a target underwriting margin of 5% of gross premium. The illustrative gross premium is ((96+20)/0.70=$165.71) for A/standard and ((237.60+20)/0.70=$368.00) for B/high, rounded to cents. Their gross-premium ratio is approximately 2.221, rather than the pure-premium factor of 2.475. Fixed expenses change the relativity even though the loss-cost model is unchanged. In practice the margin definitions, taxes, expense allocation, smoothing and legal constraints would require explicit separate treatment.

A fourth construction tests a count-deviance expression. For an observed Poisson count (y) and predicted mean (\widehat\mu>0), the unit deviance is (d(y,\widehat\mu)=2[y\log(y/\widehat\mu)-(y-\widehat\mu)]), with (0\log0=0). For (y=3) and (\widehat\mu=1.5), the deviance is (2[3\log2-1.5]), approximately 1.1589. For (y=0) and (\widehat\mu=0.1), it is 0.2. These are arithmetic verification cases, not likelihood tests across observed portfolios.

These examples establish the importance of carrying units and financial definitions throughout a GLM calculation. A coherent result can be independently reproduced from its assumptions; statistical reliability would additionally require historical estimation data, prospective validation and appropriate uncertainty assessment. The example structure follows the practical priorities identified in the CAS classification-ratemaking monograph without reproducing any proprietary data or example from that source.

Diagnostics and Statistical Validation

Model validation begins with the target and observation mechanism. A good count GLM should reproduce calibration by exposure-weighted risk segments, reporting period, coverage, and important classes without systematically underestimating large or rapidly changing cells. Residual analysis asks whether deviations correlate with risk factors or time; it does not become convincing merely because average residuals equal zero by a fitted intercept constraint.

For Poisson observations, Pearson residuals are (r_i=(y_i-\widehat\mu_i)/\sqrt{\widehat\mu_i}); the signed deviance residual has magnitude (\sqrt{d_i}) and sign of (y_i-\widehat\mu_i). These statistics are useful for flagging poor fit but are heteroskedastic, leverage-sensitive, and discrete for small expected counts. For Gamma severity and Tweedie aggregate costs the relevant deviance, weights and residual behavior differ. The CAS 2025 validation chapters cover actual-versus-predicted plots, lift charts, double lift charts, loss ratios, stability and documentation.

An information criterion such as AIC compares models by a penalized likelihood when likelihoods and estimation samples are comparable. A model with quasi-likelihood in place of a full probability model cannot be mechanically ranked using an ordinary AIC derived from unavailable full likelihood values. Deviance comparisons also require attention to nested hypotheses, dispersion and sampling assumptions. A low training deviance does not measure prospective stability.

A robust research and practice workflow uses a development sample for specification, a validation sample for model selection, and a held-out assessment or genuinely future time period for reporting prospective performance. When a time-ordered exposure process matters, a rolling-origin evaluation is often more relevant than a random split. Metrics should be aligned with the financial target: calibration of expected aggregate dollars, frequency calibration, severity conditional mean, claim concentration and tail sensitivity address different failure modes. Gini or lift improvements alone do not certify unbiased portfolio-level total premium adequacy.

Consider the difference between two candidate rating variables. A model may assign territory with a statistically stable effect but produce severe pricing swings for small spatial units; an alternative may have slightly lower in-sample fit yet more stable geography and better prospective aggregate calibration. The preferred production choice requires an explicit decision criterion, exposure volume, uncertainty, operational feasibility and constraints on allowable factors. Model governance should preserve the rejected alternatives and the reasons for selection.

Portfolio Drift, Selection, and Predictive Uncertainty

A GLM fitted on past policies estimates a conditional feature of the historical portfolio under the data and estimation regime. The future insured mix can differ because of product changes, marketing, underwriting rules, inflation, claims handling, legal reform, and competitor responses. If (P_{\mathrm{future}}(X)) changes while (E[Y\mid X]) stays stable, the aggregate prospective rate changes through the new exposure mix. If the conditional outcome process changes, the regression's factor coefficients can also cease to transport.

Exposure trend and severity inflation are distinct. An increase in sums insured or repair prices can raise severity even when claim count per vehicle-year is stable. Changes in deductibles or policy limits alter the payment distribution; increasing frequency of small reported claims may arise from reporting procedures without comparable changes in serious losses. Actuarial adjustment must specify whether a trend factor is applied before or after contract transformations, and whether the historical observation has reached a sufficiently developed valuation basis.

Uncertainty has at least four practical sources: finite-sample parameter uncertainty, observation variability conditional on fitted parameters, model and specification uncertainty, and future environment uncertainty. A confidence interval for one coefficient reflects only a defined subset of these. A predictive interval for individual cost is much wider than a confidence interval for the conditional mean because individual insurance loss remains variable even if the mean were known exactly. Aggregate portfolio uncertainty additionally depends on correlated events and concentration.

Risk selection can undermine the interpretation of regression effects. A premium change may change which risks renew, and a new classification plan may alter insurer underwriting decisions. Historical covariates can become endogenous to previous prices or claims handling. Causal claims about a price intervention require a separate design; a predictive coefficient alone describes conditional association in the observational sample. This limitation is central whenever a rating model is used to evaluate policy interventions rather than to forecast costs under a roughly stable business process.

The governing professional counterpart is ASOP No. 56, Modeling, which includes evaluating and mitigating model risk, model testing, validation of output, documentation and governance in its U.S. professional context. This is the adopted 2019 revision effective in 2020, rather than an earlier exposure draft with a similar page title. The scientific requirement to disclose the limits of transportability applies independently of the jurisdiction's exact professional rules.

Credibility, Interactions, and Model Extensions

Sparse rating cells create unstable observed ratios. Actuarial credibility methods borrow information across related groups, and multilevel models introduce group-level distributions or random effects. The objective is to prevent extreme results driven by small denominators while allowing evidence of genuine persistent differences to matter. Ohlsson and Johansson's treatment of multilevel factors connects GLMs to credibility ideas. Partial pooling requires exchangeability or an appropriately structured hierarchy; it should not imply that two territories with materially different exposures can always be merged.

Interactions carry substantive meaning in insurance. A vehicle-age factor may depend on territory, a coverage-limit effect on line of business, and a commercial property peril on construction type. In a log-linked model, adding (\beta_{AB}A_iB_i) means that the combined relativity includes (\exp(\beta_{AB})) beyond the product of the two main-effect relativities. Interactions increase flexibility and also consume exposure support. Model selection should consider cell volume, interpretability, extrapolation and multiplicity, not merely whether an interaction term crosses a conventional significance threshold.

Nonlinear effects can be represented through factor bands or prespecified spline bases inside a GLM. Generalized additive models extend systematic terms to estimated smooth functions, and hierarchical GLMs allow more structured group variation. Penalized GLMs can stabilize correlated or high-dimensional factors, but regularization alters inferential and calibration properties; an intercept and aggregate calibration may require deliberate treatment after penalization. When operating controls require transparent rating tables, smoothly estimated effects may be translated into documented, legally permissible factor schedules.

The revised CAS GLM monograph covers interactions, multicollinearity, validation, generalized additive models, mixed models and penalized extensions. These are established methods and potential substitutes in a model-development comparison, not a license to claim a new theoretical architecture. Complexity is useful when it resolves a demonstrated pattern and remains stable under the use for which the price is intended.

International Standards and Institutional Governance

Actuarial pricing is both a technical and institutional practice. Mathematical GLM identities are internationally portable; legal permission to use a characteristic, required documentation, rate oversight, consumer remedies, and fairness standards are jurisdiction specific. This monograph uses U.S. actuarial standards and selected EU supervisory instruments to illustrate that division, without treating one jurisdiction's obligations as universal law.

In the United States, ASOP No. 12, Risk Classification addresses the design, review, and change of risk-classification systems within its professional scope and includes considerations of relevant characteristics, class design, practical use, and applicable law. ASOP No. 23, Data Quality addresses selection, review, use and disclosure of data constraints. ASOP No. 53, Prospective Costs addresses prospective property-casualty cost analysis. ASOP No. 56, Modeling addresses model purpose, testing, reliance, governance, disclosure, and controls. The standards have defined professional applicability and cannot be converted into binding rules for every insurer worldwide.

The National Association of Insurance Commissioners insurance AI oversight overview records adoption of a model bulletin in December 2023 and subsequent state-level and supervisory work. A model bulletin is not identical to enacted law in every state. The presence of a GLM in a pricing workflow does not itself determine whether a specific AI-related bulletin or other regime applies; the implementation, use, jurisdiction, and instrument determine the obligations.

For the European Union, EIOPA's 2025 AI governance Opinion explains supervisory expectations about insurer AI systems under existing sectoral law and distinguishes AI Act categories. The European Commission's life-and-health insurance AI guidance describes the Annex III category for AI systems intended for risk assessment and pricing for life and health insurance. A conventional regression or a broader insurance IT system cannot be presumed to meet the legal definition of an AI system solely from the mathematics of GLMs. Nor should the life-and-health category be projected mechanically onto all nonlife rating models.

Transparent pricing demands more than a list of positive coefficients. The analyst should document whether a characteristic is available at quote time, how missingness is handled, whether it is legally permissible, the consequences of correlated proxies, the stability of effects across groups, and the sensitivity of prices to model version changes. Predictive association can produce classification patterns correlated with protected traits or historic exclusion even without the inclusion of an explicit protected attribute. Fairness assessment therefore requires both statistical examination and interpretation under applicable law and institutional rules.

An independent audit trail should connect source data and valuation cutoffs to transformations, fitted coefficients, selected factors, testing results, underwriting implementation and effective rating period. Approval and deployment are separate events. The supporting statistical research permits technical scrutiny; governance records permit reconstruction of which version of the model was used for which insured risks at which time.

Integrated Interpretation and Research Findings

The mathematical results and institutional evidence converge on a first finding: the GLM is a controlled conditional-mean and variance specification, while actuarial pricing is an interpreted forecast of covered cost integrated with a financial and regulatory decision. A GLM does not supply the coverage definition, exposure base, claims maturity, allowable factors, future trend, rate margins, or governance verdict. Each is an independent input whose effect should be traceable through to the proposed price.

A second finding follows from the equations for log-linked frequency and severity. A conditional pure-premium factor can be decomposed into frequency and severity components only under compatible conditioning and aggregation assumptions. Once a fixed expense, coverage limit, selection process, or dependence structure enters, the simple factor product can cease to be a ratio of final premiums or even of correctly specified expected covered losses. The constructed four-cell example demonstrates the exact algebra; its deterministic construction establishes mathematics and says nothing about predictive quality.

A third finding is epistemic. Fitted coefficients, model calibration and prospective utility are different forms of evidence. A convergence diagnostic supports numerical estimation; an information matrix supports a specified sampling-uncertainty calculation; a held-out time test supports a conditional forecast claim over the tested period; and a documented institutional review supports responsible use. None individually substitutes for the others. This interpretation synthesizes the source-established separation of GLM mathematics, CAS model development and validation, and professional modeling controls.

A fourth finding concerns statistical transparency. A log link makes conditional ratios easy to read but cannot make a claim about causal mechanisms, fairness, or stable transport across environments. A simple model may yield a clear but inadequate estimate, while a richer model may improve calibration yet require additional control of interactions, proxy effects, and uncertainty. Interpretability is achieved by documenting the relationship between predictors, estimands, contracts and intended decisions, not by the mere absence of complicated algorithms.

A fifth finding concerns institutional comparability. The same Poisson or Gamma equations can operate across countries, but the operational pricing product is composed under local contract forms, data-access rules, regulatory constraints and market conditions. International analysis must therefore align mathematical objects and separately describe the national or regional rules that govern use. That separation also prevents erroneous assumptions that U.S. ASOPs or selected EU AI materials impose identical obligations everywhere.

Together these findings form an original analytical synthesis of existing theories and established actuarial methods. They are interpretive conclusions supported by derivation and cited material, without the presentation of a new named doctrine, new distribution family or verified empirical discovery.

Research Boundaries and Reproducibility

This dossier is a foundational analytical monograph, not an estimated insurer rating model. It uses no individual insurance records, confidential business data, experiments with policyholders, market filing sample, or proprietary performance measurement. The numerical cases are deliberately specified and reproducible by arithmetic, which validates the displayed identities but cannot validate a model on unseen insurance outcomes. Observational data are necessary for external calibration, claims-development assessment, prospective performance, and line-specific inference.

Its references include source-level inspection of a CAS 2025 primary specialist monograph, readable publisher metadata, and directly readable professional-standard and supervisory HTML. For some foundational works, the inspected evidence consists of publisher abstracts and bibliographic records rather than full critical access to every original printed page. The bibliographic entry supports the documented proposition at that access level; it is not represented as a line-by-line critical edition. The evidence ledger records the per-source verification basis and limits.

An analyst seeking to reproduce an operational GLM pricing study should additionally preserve: a precisely defined contract and target variable; the data extraction and evaluation dates; the exposure denominator; claim joins and development adjustments; the fitted design matrix and categorical baseline definitions; preprocessing code; distribution and link parameters; coefficient estimates and covariance; training and prospective test splits; calibration and residual summaries; rate transformations; and the legally reviewed implementation. Publishing a coefficient table alone is insufficient for reproducibility.

The document distinguishes internal scholarly and editorial assessment from independent peer review and from later production release verification. The proposed Research Dossier canonical URL is reserved and no first-publication date is assigned in the manuscript. The evidence and editorial passport in this source package record the exact revision and tests performed during this research session.

Conclusion

Generalized linear models supply an essential statistical foundation for actuarial pricing because they connect insurance responses, exposure, covariates, expected outcomes, and variance assumptions within a coherent inferential structure. Their strongest applications arise when the response distribution matches the support of the insurance data, the offset reflects a justified exposure law, and the link supports an economically interpretable conditional mean. Poisson counts, Gamma positive losses and Tweedie aggregate costs illuminate distinct facets of the same covered portfolio, each with characteristic assumptions and diagnostic vulnerabilities.

The mathematical derivations establish the compound-loss expectation and variance under explicit independence assumptions, clarify the algebra of multiplicative relativities, and show why final gross premiums can differ from pure-premium ratios. The numerical cases verify formulas for a constructed four-cell rating table, a Tweedie zero probability and a gross-premium transformation. Their evidence is demonstrative, with no representation as measured market experience.

The principal scientific conclusion is that an actuarial price acquires meaning through a traceable chain from contractual coverage and exposure to data, statistical modeling, expected prospective cost, rate construction, testing and governance. Each link has different assumptions and different evidence requirements. GLM coefficients become decision-relevant when this chain is explicit, calibrated to the insured population, and examined under the professional and legal framework in which the resulting rates will operate. The complete argument remains a synthesis of established statistical, actuarial and institutional knowledge, with the manuscript prepared for a separate, later batch publication decision.

References