Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Generalized Additive Models in Insurance Prediction

Abstract. Generalized additive models (GAMs) estimate systematic nonlinear relations between explanatory variables and the conditional distributions of insurance outcomes while retaining an intelligible additive structure. They have become a flexible bridge between generalized linear models (GLMs), whose rate relativities are often represented through prespecified categories and linear terms, and algorithms that permit more complex interactions but can be harder to interrogate. This monograph reconstructs the mathematical foundation of GAMs, penalized spline estimation, smoothing-parameter selection, response-family choices, exposure-aware claims prediction, interactions, geographic structure, longitudinal dependence, statistical uncertainty, and comparative predictive validation. It studies claim frequency, conditional severity, aggregate pure premium, reporting delay, and related insurance applications through a clear separation of response estimands and contractual exposures. Original analysis includes derivations of the penalty–complexity trade-off, the conditions for multiplying frequency and severity predictions, a reproducible hypothetical pricing calculation, and a structured comparison of modelling choices under distribution shift and small-cell uncertainty. Verified statistical sources and primary actuarial publications ground the interpretation; international supervisory materials delimit insurer governance obligations without implying one worldwide legal rule. The central finding is that the useful flexibility of a GAM depends on the specification of its prediction target, likelihood, effective exposure, smoothness constraints, training design, and decision context. A smooth curve remains an observational model component and requires separate evidence for causal interpretation. No new named theory, original empirical performance claim, or external peer review is asserted.

Keywords: Generalized additive models; GAM; insurance prediction; actuarial science; nonlinear regression; penalized splines; claim frequency; claim severity; insurance pricing; pure premium; exposure offsets; Tweedie distribution; negative binomial; smoothing parameters; REML; concurvity; tensor-product smooths; spatial smoothing; predictive validation; calibration; actuarial governance; uncertainty quantification.

Introduction

Insurance prediction begins with a target quantity. The number of claims reported during a year, ultimate paid loss from a specific accident year, the size of a claim conditional on occurrence, the probability that a claim exceeds a policy limit, and the expected loss per earned policy-year require different denominators, distributions, and forms of validation. Their predictors can include age, usage, insured value, location, coverage, calendar period, distribution channel, and past experience. The relation between many such predictors and loss outcomes is nonlinear; at the same time, insurers need traceability from observations to a defensible rating or reserving estimate.

The historical starting point is the likelihood and link-function architecture of Nelder and Wedderburn (1972), whose generalized linear model formulation makes count and continuous response regression part of a common estimation family. Hastie and Tibshirani (1986) expanded that framework by replacing specified linear covariate contributions with smooth, data-informed functions in their generalized additive model study. This substitution has a precise statistical consequence: more flexible conditional patterns can be represented without abandoning a specified response distribution or a transparent sum of terms on the link scale.

Modern GAM implementations typically build smooth functions from finite bases and control their complexity with quadratic penalties. Wood (2017) gives a systematic treatment in Generalized Additive Models: An Introduction with R. Actuarial practice supplies the complementary boundary conditions: an elegant estimated function is useful only if the insured event, coverage period, claim development, exposure base, rating granularity, underwriting decision and prospective population are defined appropriately. Ohlsson and Johansson (2010) devote a chapter to additive smooth terms in non-life insurance pricing with GLMs. The 2025 CAS insurance rating monograph provides a practical baseline for judging whether additional flexibility merits operational complexity.

This dossier examines the statistical and insurance-specific choices that make a GAM credible. It develops equations from the core probability model, considers deviations from canonical assumptions, and sets forth evaluation tests that can distinguish smoother fitted curves from demonstrated forecasting gains. The methodological contribution is interpretive and analytical: connecting established spline mathematics to contract-specific response definitions, leakage-resistant validation, and the actual uses of insurance predictions. Mathematical illustrations are constructed examples. They are not fitted to a proprietary insurer portfolio, and no comparative claims of observed predictive superiority are made.

The account has international scope. Examples emphasize non-life exposure and claims because these give clear mathematical cases; the general framework also covers binary insured events, duration, and other outcomes when a defensible likelihood and observation scheme are available. The actuarial forecasting object should remain distinct from decisions about permitted pricing factors, fair treatment, and consumer-facing explanations. Those decisions involve separate institutional assessments.

Statistical Foundations and Prediction Targets

A statistical estimand precedes a statistical model

Let Y denote the insurance response for one observation, X a vector of information available at the intended prediction instant, E a measure of contractual exposure, and C the governing coverage terms. A model of conditional expected counts, E[Y | X,E,C], addresses a different quantity from the probability P(Y>0 | X,E,C) or the conditional distribution of positive payment Y | Y>0,X,E,C. The chosen forecast horizon and observation unit enter the estimand as concretely as the covariates. A policy-period outcome cannot automatically be treated as a claim-level outcome, and an accident-date loss is not interchangeable with a settlement-date transaction.

For a one-year motor portfolio, a rating study might aim to estimate the ultimate insurer-paid loss per earned vehicle-year for coverage with fixed deductibles and limits. The response is then a loss measure reflecting the stated contract, the denominator is earned exposure, and evaluation must account for both the fraction of the period insured and the development of reported claims. A reserve study might instead estimate future payments conditional on claim information observed at valuation. The latter target changes as the claim matures; even identical fitted smooths of elapsed time would answer different inferential questions.

A central statistical condition is temporal availability. Predictors recorded after the loss, including settlement-specific information, can be valuable for claims management and invalid for underwriting at policy inception. Predictors drawn from a later extract can inadvertently encode renewal decisions, retroactive edits, or future exposure. Such leakage may be stronger in a flexible smoother because it can exploit fine-grained structure left unnoticed by a simpler baseline. The predictive model must be paired with a documented data provenance timeline.

The actuarial statistical literature treats predictive work as the combination of data structure, estimation and decision-making. Frees, Derrig, and Meyers (2014) assemble these strands in Predictive Modeling Applications in Actuarial Science. A GAM is therefore best understood as a candidate conditional model inside an entire predictive design rather than as a self-sufficient rating method.

Sources of uncertainty

Three uncertainties must be kept separate. Process uncertainty describes randomness in claims even when the conditional distribution is known. Parameter uncertainty concerns the coefficients and smoothing parameters inferred from a finite sample. Structural uncertainty concerns the selected family, basis, interactions, latent risk groups, and future distribution. A confidence interval around an estimated age effect addresses only selected portions of these sources. It does not become a prediction interval for next year's aggregate losses merely by applying an inverse link.

Insurance outcomes are often heterogeneous: frequency may vary among lines; severity depends on coverage limits and inflation; claims can cluster by catastrophe or insured group. One regression likelihood may serve well for one subportfolio and poorly for another. Careful preliminary summaries of denominators, reporting lags, censoring, and distributional tails are part of the scientific design, not cosmetic preprocessing.

Additive and Smooth Predictors

The link-scale structure

A generalized additive model specifies a response distribution and a link g for its conditional mean μ_i. A useful representation is

[ g(\mu_i) = \eta_i = \beta_0 + \sum_{j=1}^{p} f_j(x_{ij}) + z_i^\top\gamma + o_i, ]

where each f_j is a smooth covariate effect, z_i contains parametric covariates, and o_i is a known offset. The functions may be nonlinear, but their contributions add on the link scale. Under a log link, the sum becomes a product of positive relativities on the response scale; under a logit link, the effects add on log odds. This retains a direct connection to traditional ratemaking factors.

Hastie and Tibshirani (1986) introduced GAMs as a class replacing specified linear terms with unknown smooth functions, originally estimated through a local-scoring and smoothing procedure. In contemporary spline implementations a smooth term typically has a finite basis representation. This representation separates function flexibility from the coefficient estimation method and makes explicit the constraints controlling complexity.

For a log-link frequency GAM, the difference f_age(25)-f_age(40) is a conditional log rate ratio between ages 25 and 40, holding all other model terms and exposure constant. Its exponential is a multiplicative rate ratio; it is not an average causal effect of aging a particular policyholder. A curve that rises at older ages can reflect vehicle usage, claim reporting, coverage choice, selection, or unmeasured characteristics in the training population. Additivity identifies an intelligible pattern conditional on the chosen regressors; it does not identify the generating mechanism.

When additivity is restrictive

A pure sum of functions f_1(x_1)+f_2(x_2) imposes that the link-scale effect of changing one predictor is invariant to the level of the other. That is appropriate when the empirical data and decision purpose support separability. It is restrictive when, for example, the mileage effect varies strongly with driver age or a coastal hazard depends on both longitude and latitude. An interaction requires its own structured smooth or parametric term; a modeler should not infer it from the fact that individual terms are nonlinear.

Every smoother has an identifiable interpretation only relative to a reference. Software often imposes a centering constraint such as the observed-sample weighted sum of f_j values being zero. The intercept then absorbs the corresponding average level. A reported smooth plot may not show the baseline premium, exposure, other terms, or even the full conditional effect of the variable. For underwriting explanations, the appropriate object is usually a contrast between two explicitly specified model predictions, including the link inverse and contractual exposure.

Spline Functions, Penalties, and Identifiability

Basis representation

A smooth is commonly represented as f_j(x)=∑{k=1}^{K_j} B{jk}(x) β_{jk}, where B_{jk} are known basis functions and β_{jk} are estimated coefficients. Cubic regression splines, thin-plate regression splines, and P-spline constructions use different bases and penalties. The dimension K_j limits the function space available to the estimator. A large K_j gives the model room to represent shape; an accompanying penalty determines how much of that room the data justify using.

Wood (2003) shows that thin-plate regression spline bases provide low-rank approximations of thin-plate smoothing problems and convenient multivariate effects. Their appeal for insurance is practical: knots or basis details can often be handled systematically rather than by manually declaring age bands whose boundaries may produce artificial premium jumps. The appropriate basis is still an analytic choice. Bounded and irregular domains, cyclic seasonality, monotone actuarial expectations, and abrupt contractual thresholds can call for different structures.

Why a penalty is indispensable

For likelihood ℓ(β), a standard fitting objective takes the form

[ \widehat{\beta}=\arg\max_{\beta}\left{\ell(\beta)-\tfrac{1}{2}\sum_j \lambda_j \beta^\top S_j\beta\right}, ]

where S_j is a positive semidefinite penalty matrix and λ_j≥0 controls smoothing. For a twice-differentiable univariate function, a canonical roughness penalty is proportional to the integral of [f_j''(x)]². A large λ_j discourages rapid curvature; a small value permits more localized variation. Penalization acts on wiggliness, whereas a separate shrinkage strategy may be needed if the intention is to remove an entire smooth term, including the unpenalized null space.

The penalty–likelihood trade-off is also an interpretation of how much detail the portfolio can support. When sparse cells contain a few large claims, an unpenalized high-dimensional curve can represent sampling noise as a dramatic rate differential. Excessive penalty can wash out genuine local risk structure and systematically underfit a change in hazard. Smoothing selection is therefore a predictive and inferential decision, not a choice to make a figure aesthetically pleasing.

The effective degrees of freedom (EDF) of a smooth summarize its contribution to fitted flexibility for the chosen smoothing parameters and model representation. In penalized Gaussian regression the familiar trace of the influence matrix supplies the intuition; non-Gaussian fitting has corresponding local approximations. EDF are not a literal count of meaningful risk subgroups, and an EDF near one does not prove the true relation is globally linear. Estimation uncertainty for smoothing parameters complicates further simplifications.

Identifiability and concurvity

Smooth basis columns can overlap with the intercept, other parametric terms, and each other. Centering constraints resolve some algebraic non-identifiability; numerical collinearity and nonlinear concurvity can remain. Consider insured asset age, policyholder age, and years since purchase in a narrow portfolio. One may be approximated as a function of the others, leaving several equally plausible allocations of predicted risk between smooth components.

Concurvity is especially consequential for reading individual risk curves. Predictions on the observed joint covariate surface can remain stable while the displayed partial effects differ materially between specifications. A responsible report distinguishes the prediction stability question from the decomposition stability question, examines covariance and concurvity, and tests whether projected contrasts fall within regions well supported by the training data. Smoothness does not protect extrapolation beyond that support.

Estimation and Smoothing Selection

Penalized likelihood and numerical fitting

GAM estimation usually alternates or jointly optimizes coefficient fitting and smoothing selection. Earlier GAM algorithms used local scoring and backfitting, and modern implementations exploit stable penalized iteratively reweighted least squares or related Newton-based procedures. The mathematical objective remains a response-specific likelihood with an explicit smoothness penalty. Implementation details matter because insurance portfolios may contain sparse factors, extreme exposures, large losses, and rank-deficient design matrices.

Wood's (2011) stable REML and marginal likelihood method directly optimizes an approximation to the restricted or marginal likelihood for smoothing parameters. The paper reports favorable computational stability and compares smoothing selection strategies, but a reported result for its studied settings should not be promoted into a universal guarantee for every insurer data structure. Alternative criteria include cross-validation, generalized cross-validation, and likelihood-related scores. Their effectiveness depends on sample independence, error structure, distribution choice, and the data partition used for tuning.

The general likelihood framework advanced by Wood, Pya, and Säfken (2016) in smoothing parameter and model selection allows more complex smooth models and can help when response parameters or likelihood structures extend beyond a single straightforward GLM. The method chosen should be declared alongside optimization convergence diagnostics, estimated smoothing parameters, coefficient uncertainty, and basis-rank checks.

Smoothing is itself model selection

A GAM can overfit even when every smooth is penalized. Fitting dozens of smooths, inspecting their p-values, retaining favored terms, and reporting intervals conditional on that process ignores selection. The model-search procedure must be nested within training data for valid out-of-sample assessment. The test set cannot become a repeated source of design choices. If practitioners inspect holdout calibration and repeatedly revise bases, interactions, variables, and response families, that holdout becomes a development set and a genuinely untouched test is needed.

Cross-validation must respect the insurance observation structure. Randomly allocating transactions from the same claim to different folds leaks claim identity. Randomly dividing policy renewals by row can put essentially the same risk and history into both training and test groups. A geographic hazard model tested only on nearby locations from the same event may appear far more stable than it will under a new event footprint. Dependence-aware folds and a prospective temporal test are more informative than a deceptively large number of shuffled rows.

Rank and penalty diagnostics

A basis dimension K_j set too small can limit a smooth regardless of how optimally λ_j is selected. The official mgcv basis-dimension diagnostic guidance describes a nearest-neighbor residual-differencing diagnostic, and it warns that stochastic p-values vary with repeated resampling. A low k-index with EDF approaching the available rank is a reason to revisit the basis and residual pattern, not an automatic instruction to enlarge every basis. Analysts should repeat suspicious diagnostics, compare refits, and inspect whether the result reflects residual dependence, distributional mismatch, or unmodelled interaction.

When the data span narrow regions of a variable, increasing basis rank does not create information outside the observed range. Local shape changes in a small exposure segment may be weakly identified even when numerical software reports convergence. A scientific account should state whether the final model is a stable forecast within the sampled support, a tentative extrapolation, or a specification requiring additional observations.

Insurance Frequency, Severity, and Pure Premium

The multiplicative decomposition and its assumptions

Let S_i=∑{k=1}^{N_i} X{ik} denote total loss over a defined exposure interval. Under a conditional compound Poisson model with event rate λ_i per unit exposure and identically distributed claim sizes having conditional mean m_i, and with the relevant conditional independence, E[S_i|x_i,e_i]=e_i λ_i m_i. A separate frequency GAM for λ_i and severity GAM for m_i can then be combined into a predicted loss cost. When both use log links, the log expected aggregate loss equals log e_i plus the sum of the log-frequency and log-severity predictors, provided their definition and conditioning are compatible.

The general identity is more demanding. Let m(n,x)=E[X_k|N=n,x] for exchangeable positive claim sizes conditional on count. Then E[S|x]=∑_{n≥0}n P(N=n|x)m(n,x), for such an exchangeable construction. It reduces to E[N|x]×m(x) only when the conditional severity expectation is constant over N under the relevant conditioning. Portfolio losses can violate that simplification when the same event causes multiple claims, when claim count signals a latent hazard, or when claim handling changes after multiple events. Garrido, Genest, and Schulz (2016) document an approach in their frequency-severity dependence study; this established literature justifies explicit testing rather than automatic multiplication.

The mathematics matters to underwriting decisions. If a model underestimates total loss when claims cluster, a smooth relativity may be interpreted as evidence of an age or location effect when the substantive problem lies in dependence. A separate severity GAM trained only on observed positive claims also conditions on claim occurrence and may be selected on an atypical subpopulation. A pure-premium study should compare the combined two-stage prediction with directly modeled total-loss alternatives, using identical claims development, exposure definitions and prospective evaluation periods.

Coverage and economic translation

A smooth risk function is not itself a complete premium. Contract design determines covered claims, deductibles, coinsurance, indemnity limits, inflation clauses, ancillary benefits, and claim settlement rules. Expected paid loss plus expenses, capital cost, reinsurance cost, and commercial constraints may generate a final rate. The CAS comprehensive insurance pricing study emphasizes real-world obstacles such as nonlinearities, correlated predictors and component blending. Fitted claims relativities are therefore inputs into decisions rather than a mechanically authorized tariff.

For a contractual change, the model must be reevaluated on the altered loss response; a rate derived from low-deductible policies does not directly estimate payment under a high deductible. Likewise, a claims-severity smoother fitted to settled claims might be affected by settlement delays or inflation that correlate with policyholder tenure. The scientific question remains the precise conditional loss cost under the stated contract and period.

Thresholds and structural heterogeneity

Some sharp changes are genuine: a coverage limit, legal eligibility condition, or product boundary can alter the expected paid loss discretely. Automatically smoothing across such a boundary may introduce bias. Models can accommodate known structural indicators or interact smooth terms with product segments where justified. Conversely, arbitrary rating age bands can introduce discontinuities unsupported by the data. The choice between a step and smooth relationship should be justified by the insurance mechanism and prospective validation, not by software defaults.

Interactions, Spatial Effects, and Portfolio Heterogeneity

Tensor-product and factor-specific smooths

A GAM with separate smooths of age and mileage assumes additive effects on the selected link scale. If the mileage effect depends on age, an interaction f(age,mileage) may be fitted. Wood (2006) develops low-rank tensor-product smooths with distinct marginal penalties; these are especially convenient when variables have different physical scales. A bivariate mileage–age effect can describe joint nonlinear patterns without creating a huge table of rating cells, although its flexibility demands enough observations throughout the joint covariate space.

Interaction plots should be accompanied by density or support information. A modeled premium ridge at combinations absent from the portfolio may be an artifact of the basis and penalty. The appropriate predictive question is whether the interaction improves held-out performance and calibration for exposure-supported combinations, and whether the added behavior is stable across calendar periods. A weakly supported localized interaction can have real regulatory or affordability consequences if used to allocate premiums.

Factor-specific smooths allow the same continuous predictor to have different functions across coverage groups or geographic regions. Such stratification may improve a fit, but separate functions in small groups can be unstable. Partial pooling through suitable penalties or random-effect constructions can often make more efficient use of data. The analyst should disclose whether groups share smoothing strength and whether the reported uncertainty reflects that pooling.

Geographic surfaces and boundaries

Risk geography is multidimensional. Flat two-dimensional smoothing can help represent hazard gradients, but geographic boundaries have meaning: coastlines, rivers, building-code zones, or abrupt microclimates may make cross-boundary smoothing implausible. Wood, Bravington, and Hedley (2008) develop soap-film smoothing for complex domains to respect difficult region boundaries in appropriate settings. For insurance maps, the method demonstrates that topology and geography can matter to the statistical penalty; it does not establish that every boundary requires soap-film methods or that using one eliminates ecological bias.

Regional data may also be correlated due to weather events, catastrophes, repair-cost inflation, local regulation and market selection. A spatial smooth of coordinates can absorb some variation without explicitly identifying the cause. Claims from a single catastrophe should not be split randomly across evaluation folds when the intended deployment concerns future catastrophes. A spatially blocked or event-held-out test better addresses that question, though it may necessarily contain fewer independent units than the row count suggests.

Time and changing risk

Calendar trends and accident-year effects can be modeled smoothly when the conditional relation changes gradually. Seasonality may call for cyclic smooths; known shocks may require discrete regime indicators; inflation may change the claim severity response. Smoothing a policy-level time trend cannot automatically repair a mismatch between underwriting-time predictors and claims developed long afterward. The prospective validation window should follow the time horizon for which the model will be used, and the feedback process should monitor covariate and outcome shifts after deployment.

Prediction, Uncertainty, and Statistical Inference

Contrasts, rate relativities, and covariance

GAM coefficients often do not convey a useful standalone actuarial interpretation. Predictions and their contrasts do. Suppose the log-frequency linear predictor for risk profile A is η_A and for B is η_B, with equal exposure and otherwise specified conditions. Then the fitted rate ratio is exp(η_A-η_B). For a smooth contrast, coefficient covariance can be used to approximate uncertainty on the difference of linear predictors, which can then be transformed to the response scale. The distinction between link-scale uncertainty and dollar-loss uncertainty should be explicit.

The official mgcv prediction documentation explains multiple prediction types and a linear predictor matrix that supports derived contrasts. Analysts should confirm how formula offsets and new-data exposures enter predictions. A term plot alone may omit the intercept, a required offset, or the covariance with other terms. A self-contained explanation should state the reference exposure, all material covariates, the contractual assumptions, and whether an interval concerns a mean, relative rate, total aggregate expected loss, or an individual future observation.

Intervals and selection

Marra and Wood (2012), in their interval coverage study, analyze how spline confidence intervals perform across smooth functions. Wood (2013) discusses inference for smooth component tests. These statistical contributions justify careful use of smooth-term intervals and tests, but their existence does not guarantee coverage when the data are dependent, smoothing selection is ignored, the family is misspecified, or the final model has been chosen after extensive exploratory screening.

Inference must reflect the purpose. If a policymaker asks whether mileage has predictive value, a conditional smooth-term test may inform an association question. If an actuary asks how many dollars next year's portfolio may lose, a distribution for aggregate losses including process, parameter and accumulation uncertainty is required. Those are different outputs. For long-tailed insured losses, ordinary approximate intervals around a mean smoother should not be presented as extreme-quantile or capital estimates.

Interpretability is conditional

The apparent transparency of a sum of smooths is a practical advantage: one can inspect shapes, offsets, and coefficients, and can compare predicted profiles. Interpretability is still conditional on feature engineering, identifiability, selection and correlation. A curve called a “driver age effect” is more precisely a conditional fitted contribution on the model scale, given training data, chosen adjustment set and centering convention. Regulatory or ethical interpretations require explanation of feature provenance, legal permission, exposure comparability and group outcomes. A readable curve does not settle those questions.

Diagnostic and Predictive Validation

Diagnostic adequacy

A complete fit report should assess the relationship between observed and fitted counts, zero frequency, dispersion, heavy tails, extreme residuals, basis rank, concurvity, and optimizer stability. Plots of residuals by exposure length and calendar period can reveal offset mistakes. Residual patterns across geography may show omitted dependence or a wrong smooth surface. Large claim payments can dominate mean models; robust analyses need to distinguish a genuine heavy-tail property from incorrect data units, duplicates, and inconsistent claims handling.

The mgcv k-index guidance offers one practical test for insufficient smooth basis dimension. Its limitations support triangulation with residual visualizations and refits. A model with stable optimization and plausible smooth curves may still use the wrong family, represent reporting selection rather than loss incidence, or fail outside its training period. Separate diagnostics are necessary for each layer of validity.

Out-of-sample design

Predictive evaluation must compare models on precisely the same estimand, eligible records, loss development basis, and exposure-weighted metric. For insurance policies beginning at time t, a credible design trains on earlier inception cohorts, tunes within those cohorts, freezes model decisions, and evaluates on a later cohort when outcomes are sufficiently developed or when the evaluation properly adjusts for incomplete reporting. A comparison that gives the GAM more current information or different claim maturation than the GLM does not isolate model flexibility.

A useful baseline ladder includes an intercept/exposure model, a conventional actuarial GLM with defensible factors, a GAM with selected smooths, and a substantially more flexible method when relevant. Friedman (2001) introduces gradient boosting of additive functions that can serve as a distinct function-learning comparator, particularly with tree base learners. Boosting performance cannot establish a GAM's superiority or inferiority in a new portfolio without data. The CAS 2026 actuarial modeling guidance explicitly stresses validation and practical judgment in high-dimensional nonlinear rating.

Scoring and calibration

For Poisson counts, one may report exposure-aware out-of-sample log likelihood or Poisson deviance, together with observed-to-expected ratios for coherent segments. The Poisson deviance for nonnegative counts with fitted means m_i>0 is

[ D_{\mathrm{Poi}}=2\sum_i\left[y_i\log(y_i/m_i)-(y_i-m_i)\right], ]

with the convention 0 log(0/m_i)=0. This formula evaluates conditional mean predictions under a Poisson criterion; its interpretation must be adjusted when excess dispersion, claim clustering or a different target is central. For positive severity, the scoring function should match the target and potential tail relevance. For full distributional forecasts, proper scoring rules distinguish statistical fit from merely producing narrow intervals.

Gneiting and Raftery (2007) develop strictly proper scoring rule theory for evaluating probabilistic forecasts. Their principle supports scoring a claimed predictive distribution by a rule whose expected value favors truthful probabilities. Calibration checks whether predicted risks and observed outcomes align over the intended population. Average observed-to-expected agreement can conceal compensating subgroup errors, so calibration should be examined across predicted-risk deciles or bands, exposures, large and small groups, protected or vulnerable segments where lawful, and future periods. Discrimination and calibration describe different aspects of quality.

A pure-premium study may also compare predicted and realized insurer-paid loss per exposure, with uncertainty around high-loss segments. The choice of score and aggregation level should be specified before performance comparison. Business measures such as retention, price elasticity or commercial conversion belong to an additional decision objective and should not be silently blended into a model presented as forecasting claims cost.

Drift and deployment

Model validation is not finished on the day of estimation. Repair costs, judicial decisions, hazards, policy terms, insured behavior, and coverage portfolios change. Monitoring should define thresholds for changes in exposure mix, fitted-versus-realized loss, latent geographic concentration, individual smooth behavior, claims development, and operational override patterns. A detected change warrants diagnosis: retraining can restore fit under some shifts, while a changed contract or claims process may require redefining the target.

Model Comparison and Operational Governance

GLMs, GAMs, distributional GAMs, and tree models

A conventional GLM offers relatively simple parametric features and often familiar rating factors. A GAM adds smooth functional degrees of freedom and a visible conditional effect decomposition. A distributional GAM can model more than the mean. Tree ensembles provide different interaction mechanisms and may perform well with heterogeneous variables but introduce their own explanation and stability requirements. Comparison should begin by aligning data, likelihood target, financial meaning and validation regime, then assess error, calibration, instability and operating complexity.

The GAM’s intermediate position is analytically important. The model can capture curvature without discarding a statistical error law and can add a small number of meaningful interactions. It can also accommodate too many spline terms, misrepresent sharp underwriting boundaries, and produce impressive in-sample figures from time leakage. No model class has an inherent claim to actuarial validity. The 2025 CAS GLM practice monograph and 2026 CAS pricing challenges monograph together emphasize that method choice is part of a complete rating workflow.

Fairness, permissible factors, and supervision

An insurer may predict risk with a variable that policy or law forbids it to use in setting an individual premium. Predictions may also correlate with socially sensitive attributes, whether or not they appear explicitly as features. Model explainability can facilitate an audit of pricing effects, but permitting a factor, applying it consistently, and testing its wider consequences are separate questions. A statistical study should therefore report the technical target and distinguish it from the operational decision rule applied to the prediction.

The IAIS (2025) paper on AI supervision organizes supervisory discussion around proportionality, governance, robustness, explainability, fairness, and redress. EIOPA (2025) describes risk-proportionate AI governance expectations under existing insurance-sector law; its Opinion expressly clarifies supervisory interpretation rather than creating a new independent statute. These are primary institutional sources with specific scopes. They do not establish that every GAM is necessarily classified as an AI system or that the same insurer obligations apply across every country, insurance line, and deployment.

A workable model governance record states the training extraction date, intended deployment horizon, approved predictor list, response and exposure definitions, fitted family, basis and penalty choices, calibration results, segment audits, version control, documented limitations, human challenge rights and remediation process. Model modifications should be traceable to re-estimated predictions and new validation. Technical explanations must be understandable to users of the rate or claim forecast rather than relying only on function names and statistical coefficients.

Decision consequences

A statistically calibrated smooth predicts expected loss conditional on recorded information; it does not decide risk-sharing goals, limits on classification, or what counts as acceptable price volatility. When smooth predictions are converted to rating steps, caps, territories or premiums, the transformation should be independently tested. An insurer can create new discontinuities after fitting a smooth by rounding, capping, or segmenting the final tariff. Governance should include the decision mapping, not solely the regression output.

Insurance Rating Case Studies and Worked Calculations

Worked illustration: earned exposure and smooth relativities

Consider a purely hypothetical portfolio in which a fitted annual claim frequency for a reference risk is 0.08 claims per earned policy-year. Suppose the sum of three log-scale GAM contributions for a specified profile is 0.25 for one smooth, 0.15 for another, and −0.10 for a territory term, holding other terms fixed. The net contribution is 0.30. Under a log-link Poisson model, the predicted annual rate becomes 0.08×exp(0.30)=0.1079887046, approximately 0.10799 claims per policy-year.

For a policy with half a year of otherwise proportional exposure, the predicted count is 0.5×0.1079887046=0.0539943523. If the mean paid loss per claim is assumed to be $3,400, identical under the relevant conditional assumptions, the illustrated annual pure premium is $367.1616 and the corresponding half-year expected total loss is $183.5808. Rounded to cents: $367.16 and $183.58. These values are direct algebraic consequences of assumed rates and loss size. They demonstrate how the exposure offset and inverse link combine; they are not empirical estimates from an insurer, and they omit expenses, inflation uncertainty, dependence, and reinsurance or capital loadings.

Now suppose a single age smooth takes the centered values f(20)=0.25 and f(40)=−0.15 with other predictors fixed. The age-20 versus age-40 rate ratio would be exp(0.40)=1.4918246976, approximately 1.492. The 49.2% relative difference is conditional on that illustrative model configuration and specified covariate contrast. Because the values are centered function contributions, neither corresponds directly to a base policy rate.

These computations reveal a common actuarial interpretation error: exp(f(age)) cannot be treated as a complete premium or as a causal effect when the model includes an intercept, exposure, correlated risk features and coverage-specific benefits. Reliable reporting uses fully specified comparative profiles and discloses the prediction scale.

Case design: smooth vehicle age versus rating age bands

A hypothetical motor portfolio uses vehicle age recorded in completed years. A GLM might include several age categories; a GAM might use a penalized cubic regression spline, holding coverage limits and usage factors fixed. Compare the two under identical exposure and claims definitions and inspect any changes at former band boundaries. The GAM could reduce arbitrary rate jumps, but a small sample at very high ages may create large uncertainty and penalty-sensitive curvature. A well-designed comparison would display exposure density and a temporal holdout’s residual losses by vehicle age, followed by calibration and sensitivity checks.

The scientific finding would be conditional on that test, and no number is reported here as an actual empirical advantage. A modeler should also consider the insurance reason for a possible threshold: vehicle condition, repair access and anti-theft technology need not vary smoothly or exclusively with age. An observed age curve may proxy for those unmeasured features.

Case design: geographic storm loss and event dependence

Imagine property insurance losses observed over several seasons. A two-dimensional spatial smooth could capture gradual regional changes in expected claim frequency. A boundary-aware specification might be relevant around coastlines or a complex insured territory, following the statistical motivation of boundary-respecting soap-film smoothing. Yet one storm event can produce many spatially adjacent claims; evaluating by randomly scattered individual policies would recycle the same event signature between training and test. An event-held-out validation and explicit catastrophe structure may be more informative for future-event prediction.

A smooth that predicts average annual occurrence may be poorly suited to severe tail capital estimation, especially if its response distribution gives insufficient weight to extreme losses. A responsible application distinguishes frequency and expected-cost forecasting from catastrophe simulation and reinsurance aggregation.

Case design: positive severity, zeros, and claim development

A third conceptual case concerns a portfolio with many zero paid totals, some settled positive amounts, and recent claims still under development. One approach fits a Tweedie mean model to an explicitly defined period outcome; another uses separate frequency and positive severity models; another builds a distributional model or incorporates development. Their target quantities must be made equivalent before performance is compared. A Tweedie model's positive mass at zero does not by itself correct missing reports or right-censoring. Likewise, fitting a positive severity smooth to early-settled claims and multiplying it by projected frequency may introduce settlement-selection bias.

The research design would compare models on ultimate loss at a consistent valuation basis, test representation of zeros and tails, document which claims remain immature, and reserve an independent evaluation cohort. An interpretable curve has scientific value only in conjunction with these conditions.

Implementation and Reproducibility

Transparent estimation specification

A reproducible GAM study should store the exact response expression, earned exposure definition, source-date cutoffs, training and validation keys, permitted covariates, cleaning transformations, distribution family, link, basis type and rank for each smooth, penalty estimation method, fitted coefficient vector, uncertainty matrix, random seeds used for diagnostics, software versions and preprocessing decisions. A fitted model object alone may conceal earlier data filtering, invalid date joins or undocumented cap rules.

For a simple conceptually appropriate count model, the formula may be expressed in R-style notation as: claim_count ~ s(driver_age, k=10) + s(annual_mileage, k=12) + territory + offset(log(exposure_years)), with family=poisson(link="log") and method="REML". This is an illustrative specification, not runnable code for a real portfolio without supplied columns, data cleaning, exposure checks, model fitting and validation. Choosing k values requires diagnostic assessment; the illustrative integers are not defaults recommended for all applications.

The software's prediction-scale and offset behavior should be checked directly when forecasts are converted from model output to actuarial rates. The family specification reference should similarly be used to confirm parameterization before mixing count, severity, Tweedie or multiple-parameter distributional families. Compatibility and defaults may change with software releases, so a scholarly reproduction record should identify the tested version, not merely the package name.

Minimal auditable comparison record

For each contender, a valid technical record contains the exact frozen dataset or data provenance manifest; the prediction target at forecast inception; fold-generation rules; model hyperparameters and smoothing choices; fitting logs and convergence status; test predictions with time and exposure identifiers; metric definitions; calibration tables; subgroup and shift assessments; and an explanation of any excluded observations. The out-of-sample comparison should preserve the same data eligibility across models, or explicitly document why comparability is unattainable.

Sensitivity analyses are integral: change basis rank, inspect penalty robustness, compare Poisson versus negative binomial counts where overdispersion warrants it, test factor-specific smooths against pooled ones, and evaluate how incomplete claims development changes the target. An identifiable research finding can then be framed as conditional on a documented response, portfolio and validation design rather than claiming universal performance from one dataset.

Reproducibility of the present monograph

This research product is a documented, source-based methodological synthesis. Its numeric illustration uses explicitly stated inputs and elementary exponentiation; no claims are made about fitting actual insurer data. The source ledger specifies source titles, checked claims, URL provenance, accessible verification scope, and the limits of what was inspected. Its manuscript and evidence file have cryptographic SHA-256 fingerprints to preserve the exact editorially reviewed edition. Reproduction of an empirical GAM study would additionally require permission to use claim-level data and a version-locked analysis environment.

Research Method, Source Criticism, and Limitations

Search and critical reading

The source set joins foundational statistical articles, later methodological developments, actuarial monographs, official software documentation and primary insurance supervisory materials. Original methods are attributed to their historical literature rather than to this dossier. The research inspection checked publisher or institutional bibliographic identity and the source-described claims; several book and journal records were accessible as abstracts or publisher previews rather than complete proprietary texts. Consequently, this study does not assert independent verification of every proof, simulation table or empirical result in restricted works.

The methodological sequence proceeded from a conditional insurance estimand through family/link selection, basis construction, penalized estimation, validation and financial decision use. Points of inferential friction were made explicit: exposure offsets presume proportionality within the defined interval; a factor smooth is conditionally associative; a frequency×severity product requires appropriate dependence assumptions; Tweedie formulation is sensitive to response normalization; smoothing-parameter estimation contributes uncertainty; claim clustering and distribution shift can undermine standard independent-observation inference; and decision fairness lies beyond model accuracy alone.

These limits are substantive. They explain why a general smooth predictor, though mathematically powerful, should not be treated as a universal model of all lines, tails, claims delays, benefits and geographic hazards. The actuarial literature represented by Ohlsson and Johansson's insurance treatment and the CAS nonlinear-pricing monograph is evidence of active use and methodological guidance, not evidence that the same GAM specification wins on all datasets.

Empirical and jurisdictional scope

No insurer-owned microdata, independently reestimated published study, randomized intervention or cross-country benchmark is presented. No exact predictive improvement, treatment effect, observed fairness difference or novelty priority is claimed. The numerical examples are transparent algebra, not data analysis. Links to regulatory materials are used to describe their stated scope: IAIS documents supervisory principles; EIOPA offers EU supervisory interpretation under existing law. The legal admissibility of any factor depends on the product, market, jurisdiction and applicable law at the time of deployment.

A future empirical study could compare pre-registered GLM, GAM, GAMLSS and tree-based specifications across insurer-held datasets using coherent exposure, contract, claims maturity and catastrophe-event definitions. Such a study would improve the evidence for line-specific predictive comparisons while still requiring source criticism and regulatory review. That proposal is a research agenda, not a new conceptual system or a result already established by this monograph.

Integrated Analytical Findings

The preceding analysis yields a connected set of conclusions. First, defining the prediction target is logically prior to specifying spline flexibility. A beautifully fitted age curve cannot repair an ambiguously measured pure premium or a misaligned policy exposure. Second, an additive log-link predictor gives an auditable conditional decomposition, but a projected rate is meaningful only after reconstructing its intercept, offset, all covariates, contractual setting and response family.

Third, spline penalties regulate the complexity of the fitted relationship, and basis rank determines which relationships are available to be fitted. Both belong in technical documentation and validation. A seemingly stable curve can be the product of insufficient basis dimension, selection on the test set, correlated covariates, or smooth extrapolation beyond support. Fourth, forecast uncertainty extends beyond a smooth-term confidence band to response randomness, parameter estimation, design selection, dependence and future shift.

Fifth, frequency, severity, and aggregate pure premium can be modeled in complementary ways. The familiar product of expected claim count and expected claim size has precise conditions and is not an algebraic truth for every portfolio. A Tweedie mean model or a distributional GAM can be useful alternatives where their response laws and zero/positive behavior match the observed process. Sixth, geographic and joint effects require substantive support: a tensor product expresses statistical interaction; a spatial boundary treatment expresses geometrical structure; neither establishes cause by itself.

Seventh, GAMs should earn their place through prospective and dependence-aware comparisons against appropriate actuarial baselines, judged by scores and calibration aligned with the cost target. Finally, translating an estimate into a price, underwriting decision or reserve introduces institutional obligations that statistical prediction alone does not satisfy. The original analytical synthesis of this dossier is the integration of these established conditions into one insurance-specific line of reasoning. It is not claimed as a newly discovered named theory, doctrine, or general predictive law.

Conclusion

Generalized additive models furnish a principled method for representing nonlinear insurance risk relationships within a specified regression likelihood. Their value lies in the conjunction of an interpretable additive predictor, controlled spline flexibility, appropriate distributional modeling, and a reproducible workflow of diagnostics and prospective evaluation. The method is particularly useful when conventional rating factors hide smooth changes in risk and when an insurer needs to inspect conditional relativities, but its success cannot be inferred from curve smoothness or in-sample fit alone.

A valid insurance GAM has a named response estimand, a defensible exposure denominator, a fitted family consistent with the outcome, explicit basis and penalty choices, attention to dependence and selection, and a comparison against credible models on an independently defined future sample. Partial effect curves describe conditional associations in the fitted portfolio. The translation from expected loss to premium, contract eligibility, capital or claims practice requires its own controls, explanations and governing authority.

The enduring scientific conclusion is methodological: nonlinear flexibility gains meaning through the complete structure of observation, estimation, validation and decision use. When these elements are documented, GAMs can support actuarial analysis that is simultaneously more adaptive to complex risk patterns and more auditable in its uncertainty and limitations. This monograph establishes a source-checked synthesis and a research agenda, preserves the distinction between mathematical deduction and observed performance, and leaves public release to the separately authorized publishing process.

References