Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Neural Networks for Insurance Risk Prediction

Abstract. Neural networks approximate nonlinear relations between observed insurance risk characteristics and future claim frequency, severity, aggregate loss, and development. Their utility depends on the statistical quantity being predicted, the insured exposure at risk, the observation process, the distributional objective, the integrity of model evaluation, and the controls governing insurance decisions. This monograph develops the conditional stochastic foundations of neural insurance prediction; distinguishes count rates, event probabilities, claim severities, pure premiums, and full loss distributions; compares feed-forward models, categorical embeddings, combined actuarial neural networks, and neural reserving; examines sparsity, credibility, overdispersion, tail uncertainty, policy selection, and leakage; and evaluates calibration, sharpness, generalization, explanation, discrimination, and supervisory governance. Published studies are assessed within their own data, horizons, implementation, and verification limits. Hypothetical calculations demonstrate the precise arithmetic of exposure, aggregate expectation, event probability, and actuarial shrinkage. The original contribution is an integrated analytical synthesis and practical comparative evaluation protocol, derived from existing statistical and institutional sources. No novel mathematical theory, independently trained network, insurer experiment, or universal superiority claim is asserted.

Keywords: Neural networks; insurance risk prediction; actuarial statistics; claims frequency; claims severity; aggregate loss; compound Poisson; Tweedie regression; generalized linear models; combined actuarial neural networks; embeddings; sparse claims; proper scoring rules; calibration; prediction uncertainty; model validation; fairness; distribution shift; insurance supervision.

Introduction

Insurance risk prediction estimates the distribution or a functional of a future contingent claim outcome given a defined contractual exposure, information set, and valuation basis. The object being estimated is created by insurance economics and mathematics before an algorithm is fitted. A model can predict the frequency of reported events, the rate of incurred losses, the final payment for an open claim, or the probability of reaching a policy limit. These targets may be associated with the same insured risk, but their statistical meaning and business use differ. Observed data represent prior contracts, prior risk selection, prior coverage and reporting practices, and prior claim-development states. Accordingly, an algorithmic relationship learned from the observed portfolio must be evaluated for the precise future population and decisions to which it is applied.

Neural networks enter this system as flexible supervised estimators. Their layered nonlinear transformations can represent interactions and smooth multivariate structure that are inconvenient to specify manually in classical ratemaking models. This flexibility also creates demands for regularization, out-of-time evaluation, control of high-cardinality feature memorization, and assurance that probabilities and loss distributions remain properly specified. Wüthrich and Merz (2023) present a comprehensive actuarial statistical foundation connecting generalized linear models, regularization, neural approximation, and forecast evaluation. The interdisciplinary application is distinctive because loss observations have contract dates, exposure denominators, reporting delays, extreme tails, legal eligibility conditions, and dependencies between policies.

The empirical literature has tested different model classes on particular insurance data. Noll, Salzmann, and Wüthrich (2020) study French motor third-party liability frequency modeling; Holvoet, Antonio, and Henckaerts (2025) compare technical tariff strategies and frequency–severity prediction on four datasets. Blier-Wong and colleagues (2021) review machine-learning methods for property-and-casualty pricing and reserving. Their work supports rigorous case-dependent conclusions, rather than the claim that an artificial neural network reliably outranks a generalized linear model, a generalized additive model, or boosted trees.

The present research proceeds through six scientific chapters retained from the approved research design. It evaluates public scholarly sources, official supervisory documents, a professional actuarial benchmark, and primary legislative records, connecting these sources to exact factual and methodological claims. Original analytical deductions and numerical examples are expressly identified. There is no confidential insurer data, independent empirical replication, clinical-style trial, or commissioned external review. Source access varied: some publications were fully open, whereas certain publisher pages supplied metadata and abstracts only. Conclusions based on those pages are correspondingly bounded.

The principal encyclopedic association is Insurance Statistics and Predictive Modeling, with further relations to Actuarial Science, Insurance Pricing and Reserving, Insurance Technology, and Underwriting. These are links to distinct published Science monographs; the present Research Dossier has one canonical identity.

The central interpretive proposition is that neural predictive value is conditional on target definition, exposure and sampling validity, comparative out-of-sample evidence, and accountable use. A statistically sophisticated model with an undefined outcome is scientifically ambiguous. A modest neural adjustment to a clear baseline can be useful when it captures stable nonlinear variation that improves a measurable decision.

Neural Predictors and Statistical Losses

The insured risk unit and time at risk

Index insured risk units by i. Let x_i denote admissible information available at a specified prediction date, e_i their measured exposure in reference insured-years, and N_i the number of claims incurred or reported in a clearly specified window. A standard model is

Ni∣xi,ei∼Poisson{eiλ(xi)}.N_i\mid x_i,e_i\sim\mathrm{Poisson}\{e_i\lambda(x_i)\}.

The conditional frequency rate λ(x_i) is expected claims per exposure-year, whereas e_i λ(x_i) is the expected count for the observed contract. The link with a neural function f_θ takes the form

log⁡E[Ni∣xi,ei]=log⁡ei+fθ(xi),λ(xi)=exp⁡{fθ(xi)}.\log E[N_i\mid x_i,e_i]=\log e_i+f_\theta(x_i), \qquad \lambda(x_i)=\exp\{f_\theta(x_i)\}.

The logarithm of exposure is an offset whose coefficient is fixed at one. This imposes proportionality of expected count to exposure, a substantive statistical assumption. A one-month policy without a claim and an otherwise identical twelve-month policy without a claim provide different information about an annual risk rate. A free network that treats policy duration merely as an ordinary feature need not preserve the unit convention, particularly near the edge of the exposure distribution. the actuarial motor-frequency report demonstrates why common exposure conventions and deviance measures are necessary for comparing alternative models on real motor portfolios.

The proportional-offset assumption can break down when event intensity changes predictably within the term, or when cover is issued and canceled endogenously in response to unobserved risk. Time-varying hazards may require multiple intervals per policy with correctly allocated risk exposure. Policyholder clustering, accident-event dependence, or frailty can further invalidate conditional independent Poisson claims even while a log-mean regression approximates average frequency adequately.

Under the stipulated Poisson distribution, the probability of at least one claim during exposure e_i is 1−exp{−e_i λ(x_i)}. This event probability is a different quantity from the annual frequency rate and expected count. Overdispersion, structural zeros, or shared latent risks can make this particular count-to-probability conversion incorrect. Model documentation must identify whether a score refers to the rate, the event probability, a conditional loss, or a general risk ranking.

Severity and aggregate losses

Let Y_ij be a strictly positive severity amount for event j on risk i, measured at a documented currency, development date, coverage and payment basis. Aggregate loss is

Si=∑j=1NiYij,Si=0 when Ni=0.S_i=\sum_{j=1}^{N_i}Y_{ij},\qquad S_i=0\ \mathrm{when}\ N_i=0.

Under conditional independence of count and identically distributed severities, the expected aggregate loss factors into

E[Si∣xi,ei]=E[Ni∣xi,ei]E[Yij∣xi].E[S_i\mid x_i,e_i] =E[N_i\mid x_i,e_i]E[Y_{ij}\mid x_i].

The proof applies the law of iterated expectation: conditional on N_i=n, the expected sum is n times the assumed conditional average severity. This factorization can fail when claim count and severity remain dependent after conditioning on the observed covariates, when severity depends on claim order or reporting status, or when insured risk limits create nonstandard event dependence. Frees, Lee, and Yang (2016) study multivariate frequency–severity relationships, reinforcing the need to examine dependence rather than automatically treating count and size as separable regression problems.

Under an independent compound Poisson model with μ expected events and severity Y having finite second moment, the law of total variance gives Var(S)=μE[Y²]. This is more than μVar(Y) because random counts add variability equal to μ(E[Y])². Two portfolios can share the same expected loss while exhibiting materially different aggregate variance and extreme-loss potential. Neither an accurate frequency mean nor an accurate severity mean alone establishes a predictive capital distribution.

In separate frequency–severity models, a count network estimates e_i λ_i, and a second network estimates positive severity m_i; the product defines estimated pure premium under the chosen assumptions. This decomposition is interpretable and may assist underwriting and operational claims planning. It also depends on a coherent positive-claim sample: observed settled payments may overrepresent fast and simple claims, creating selection bias when outstanding severe claims remain unresolved.

A direct aggregate model instead predicts the annual loss S_i as a zero-inflated and positively skewed outcome. One mathematical option is a Tweedie exponential-dispersion family with variance φμ^p and, for 1<p<2 under the appropriate distributional structure, a compound Poisson–gamma representation. the actuarial exponential-dispersion treatment explains how the likelihood family defines both an estimation target and a mean–variance relationship. A neural network fitted only to μ leaves dispersion, tail behavior, policy dependence, and large-loss accumulation unaddressed.

Observed zeros and the claim-generating mechanism

A large fraction of insured contracts contain no recorded claim. Zero counts arise naturally from rare events, short exposure, deductible thresholds, reporting choices, and heterogeneous risks. It is methodologically important not to call every zero-inflated claims dataset a binary class imbalance problem. Claims count data may be Poisson, overdispersed, negative binomial, or generated through a more complex multiple-stage reporting and payment mechanism.

For a negative-binomial parameterization with mean μ and dispersion κ, Var(N)=μ+κμ². It permits extra-Poisson variation without treating all extra zeros as products of a structural-zero population. A separate zero-inflated or hurdle model requires substantive justification for its two processes. A reported but unpaid claim, a claim paid below deductible, and no event during the observation period are distinct observations with distinct implications for the target variable.

Blindly oversampling observed claim rows in a neural regression changes the sampling distribution and may bias the portfolio mean unless sampling probabilities and losses are corrected. Conversely, using inverse-frequency weights optimized for a classification score may produce attractive event detection statistics while destroying expected-count calibration. The empirical risk used in training must correspond to the insurer's intended statistical estimand and sampling design.

The information available at decision time

Write F_t for the information legitimately available to an insurer at decision time t. A prospective underwriting forecast estimates a future loss functional conditional on F_t, not on the entire completed policy record. A claim settlement indicator recorded after the loss, a repair invoice from next month, or a final reserve assigned after litigation cannot be included in a pre-bind predictor. Such leakage can make a neural model appear exceptionally accurate during retrospective testing.

The relevant dates include policy inception and termination; exposure accumulation; incident, notification, and reporting; payment and reserve movements; feature creation; accounting valuation; and the cutoff at which historical information was extracted. Data-processing systems often overwrite current feature values, so a retrospective table may contain information unavailable on the historical quotation date. Reconstructing decision-time features is an independent research task requiring an auditable data vintage.

For reserving at development age k, information observed through that age can be legitimate. The target is then conditional on a different sigma-field and horizon. Delong and Wüthrich (2020) study joint incurred and payment histories for individual reported, unsettled claims. A claim development architecture should therefore never be confused with a prospective claim-incidence model merely because both use neural components.

Economic normalization is equally important. Historical amounts may be gross or net of recoveries, capped at contractual limits, denominated in changed currencies, or inflated to different cost levels. Changing a valuation convention alters the response distribution even when the nominal underlying hazard does not change. The scientific specification must precede model training and accompany every comparison.

Two numerical meanings of a risk score

Consider a deliberately hypothetical motor contract with exposure 0.75 year, annual frequency rate 0.12 claims per year, and conditional mean severity $9,000. Under the independent Poisson-frequency and severity assumptions, the expected count is 0.75×0.12=0.09. The event probability is 1−exp(−0.09)≈0.0861, or 8.61%. Expected loss during the three-quarter-year term is 0.09×$9,000=$810. Annualized expected loss is $1,080 per full exposure-year.

The annual rate is 12 claims per 100 exposure-years. The 8.61% probability describes one or more events over only the particular 0.75-year period. The $810 expected loss describes that period's aggregate, whereas $1,080 is its annualized pure-premium analogue. A pricing load, expenses, investment assumptions, coverage exclusions, deductible, reinsurance cost and capital charge are separate variables. These four values have different units and must remain separately named in any neural output contract.

The example is algebra, not a measured insurance premium or benchmark. Its purpose is to prevent the frequent methodological error of treating a predicted count, a classification probability, and an amount as interchangeable numerical risk scores.

Architecture and Training

Conditional mean architectures

A feed-forward neural network composes transformations of the form h^(ℓ)=σ(W^(ℓ)h^(ℓ−1)+b^(ℓ)), starting from encoded insured characteristics h^(0)=x. Its last layer produces a real-valued function f_θ(x). An exponential link, positive softplus transformation, or output parameterization appropriate to the desired distribution converts this output into positive count rates or severity means.

The architecture controls the family of functions approximated by the network. The likelihood or loss defines which functional of a random outcome the network is encouraged to learn. Mean-squared error targets a conditional mean under standard integrability assumptions, but is often unstable for highly skewed claim payments. A count deviance or Poisson likelihood respects a count response and exposure, while Gamma deviance may better match positive severities than ordinary squared error. The choice of statistical loss is consequently part of the scientific hypothesis, not an interchangeable training option.

For independent positive-exposure Poisson observations with log intensity log e_i+f_θ(x_i), the negative log-likelihood up to constants is

LP(θ)=∑i[eiexp⁡{fθ(xi)}−Ni{log⁡ei+fθ(xi)}].\mathcal L_P(\theta)=\sum_i\left[e_i\exp\{f_\theta(x_i)\} -N_i\{\log e_i+f_\theta(x_i)\}\right].

If the network includes a free intercept and the unpenalized likelihood reaches an unconstrained stationary point, differentiating with respect to that intercept gives ∑e_iλ̂_i=∑N_i. This is an aggregate balance equation, not conditional calibration. Weight penalties, early stopping, constraints, numerical optimization errors or offsets implemented incorrectly may disturb this identity. Thus a completed training procedure still needs an independent exposure-weighted total check.

In a zero-heavy portfolio, minimizing unweighted squared error on observed individual annualized frequencies N_i/e_i can give very large variance to short-exposure rows. The Poisson count objective retains their event-count sampling scale under its assumptions. When residual dispersion is substantial, a negative-binomial likelihood, quasi-likelihood evaluation, or explicit dependence model may be more suitable. A neural output link does not abolish the distributional obligations of insurance statistics.

Representing insurance risk characteristics

Tabular actuarial data typically contain numerical and categorical predictors: driver age, vehicle power, insured value, geography, history of claims, distribution channel, product, industry or occupation. Continuous variables can enter through standardized or carefully transformed features. For categories with many levels, one-hot encoding increases dimensionality, while a learned embedding assigns each category a vector trained jointly with the outcome.

Categorical embeddings can represent nonlinear interactions efficiently and may help integrate postal codes, product identifiers and other high-cardinality fields. the four-dataset pricing benchmark explicitly discusses categorical embeddings and autoencoders in the preprocessing of frequency–severity models. A learned embedding is a numerical approximation to predictive associations. Distances between embedded regions do not establish a causal geographic law or justify protected-trait proxy effects.

A category occurring only a few times can receive an unstable embedding. An insurer should document exposure thresholds, pooling of rare groups, handling of missing identifiers, and an out-of-vocabulary strategy for entirely new categories. Geographic administrative codes and commercial product identifiers change; a model that silently maps unseen classes to a frequent training category can distort their expected loss.

An alternative to high-dimensional individual memorization is to learn interactions within a controlled baseline. Group smoothing, explicit limits on neural residual corrections, and selective features can protect forecast stability. These are choices that must be evaluated for any loss of genuine nonlinear signal on independent data.

Nested actuarial models and combined networks

A combined actuarial neural network starts with an established model such as a GLM and learns a residual function. With a log-linked frequency baseline,

log⁡λCANN(x)=log⁡λGLM(x)+rθ(x).\log\lambda_{\mathrm{CANN}}(x) =\log\lambda_{\mathrm{GLM}}(x)+r_\theta(x).

Therefore λ_CANN=λ_GLM exp(r_θ). If the neural residual is identically zero, the baseline is recovered. A separate neural correction may learn residual interactions while retaining a meaningful reference tariff structure. Schelldorfer and Wüthrich (2019) investigate nesting classical actuarial models into neural networks. The Gabrielli–Richman–Wüthrich reserving study extends the same idea to cross-classified over-dispersed Poisson reserves and investigates bootstrap prediction uncertainty.

Model nesting creates a useful scientific experiment: compare the unmodified baseline, the standalone network, and the combined version on exactly the same data boundaries. An improvement that survives an out-of-time test reveals an incremental association not captured by the baseline under the chosen score. It does not prove that a learned adjustment represents a stable causal mechanism or remains effective after a product redesign.

CANN structures may begin from a GLM or a boosted-tree base. Holvoet and colleagues' study examines both variants and discusses conversion into a deployable technical tariff through surrogate model structures. The surrogate must be evaluated for approximation error and stability, rather than assumed to reproduce the complex source model exactly.

Sequential and alternative insurance architectures

A claim is a temporal object. Payments, estimates, reporting delays, and closure status evolve. Recurrent networks, attention-based representations, and models of temporal sequences can be useful when these transitions are available before the prediction date. They must be masked carefully to prevent future values from entering a historical evaluation. the joint claims-development model is an empirical example of a reserving question in which the individual claim history is itself the central input.

Telematics data combine time-varying driving signals with policy attributes. Duval, Boucher, and Pigeon (2024) investigate combined actuarial networks for cross-sectional and longitudinal claim counts, a setting with meaningful repeated-measure dependence. Their subject differs from ordinary static motor ratemaking: temporal aggregation and driver behavioral histories must be considered part of the estimand.

Other applications may incorporate images, text, spatial fields or weather hazards. Learned representations can increase available predictive information while introducing feature provenance, consent, missingness, vendor dependency, drift and explainability issues. The mere availability of a neural architecture does not justify including a modality whose measurement or legal use cannot be established.

Optimization, hyperparameters and model governance

Neural parameters are commonly estimated by minibatch gradient methods with regularizers and early stopping. Performance depends on random initialization, optimizer, learning rate, batch design, depth, width, activation functions, categorical encoding, stopping epoch and software versions. Hyperparameter tuning is itself a data-dependent selection process. Repeatedly choosing a model by validation-set performance introduces optimization of validation noise, even when each candidate's fitting process does not access the test labels.

Reproducibility requires fixed cohort extraction, full target definition, feature provenance, deterministic split membership, transformations fit only on training folds, loss and penalty specification, hyperparameter search budget, random seeds, package versions, and archived trained weights. An independently protected final test should be opened only after the modeling design is frozen. Operational stakeholders should receive the exact model version and input schema that generated each validated score.

Training convergence establishes approximate optimization for the selected objective. It does not certify calibrated premiums, fair treatment, resilience under distribution shift, predictable claims development, or capital adequacy. Every such claim requires its own evidence and threshold.

Regularization for Sparse Claims

What makes insurance data statistically sparse

The number of insured contracts can be much larger than the number of informative claims. Rare losses, large limits, uncommon risk factors and long claim-development processes produce sparsity even in databases with millions of rows. Effective sample size depends on the number and dependence of events, not simply the size of a policy table. Multiple claims from one catastrophe may share a common cause; multiple periods for one policyholder may share an unobserved propensity; repeated claim payments are correlated with claim development.

Neural models can fit irregular patterns in these correlated records, particularly when high-cardinality categorical features identify narrow underwriting groups. A strong training score can then reflect memorization. Regularization and test design must match both parameter complexity and the data-generating structure. the P&C machine-learning survey documents the broader consequences of flexible learners in insurance pricing and reserving.

A small neural architecture may be more reliable than a deep one for a thin count target. Weight shrinkage, dropout, early stopping, pooling of rare levels and smoothness restrictions can control different forms of instability. None guarantees valid extrapolation or fully quantifies uncertainty under unseen economic regimes. A regularizer is a modeling assumption whose effectiveness is established by held-out performance and stability.

Heterogeneity and credibility

Insurance portfolios contain heterogeneous levels of expected loss. A fitted frequency relationship may be precise for a large, homogeneous class and noisy for a group with a handful of claims. Classical credibility and hierarchical statistical models make partial pooling explicit. A neural predictor can borrow strength implicitly across similar covariates, but its degree of borrowing and uncertainty are usually less transparent.

For a hypothetical rate class with five exposure-years and one claim, the raw rate is 1/5=0.20 claims per year. Consider a Gamma prior on a constant Poisson event rate with shape 2 and rate 20, giving prior mean 0.10. Combining this prior with one observed claim in five exposure-years produces a Gamma posterior with shape 3 and rate 25, giving posterior mean 3/25=0.12. This illustrates shrinkage due to a specified prior and sampling law.

The calculation is an algebraic illustration with invented parameters. It does not certify the prior, the rate, or a premium for actual insured persons. A neural network trained on that small class may generate a different rate because it uses covariate similarities and a different regularizer. Comparative assessment requires testing both estimators on subsequent claims and evaluating the suitability of their assumptions.

Heavy-tailed severity and coverage transformations

Positive insurance claims can be sharply skewed. Policy limits, deductibles, litigation, repair costs, medical inflation, and the time needed to settle large claims shape the observed data. Fitting a flexible conditional mean does not specify the severity tail. A network can improve ordinary mean-based deviance while underestimating the frequency or severity of losses above a large threshold.

Consider a stylized contract payment Y=min(max(X−d,0),u), where X denotes ground-up loss, d a deductible, and u a payment cap under the assumed wording. Modeling Y predicts paid losses under that contract, while modeling X requires information about ground-up loss that may be censored by the payment rule. Observing many capped payments does not reveal their underlying uncapped tail beyond the limit. A more flexible neural architecture cannot statistically identify information unavailable in the likelihood.

Claim development also selects which severities are observed as final. Fast-settling small claims may dominate a dataset of closed claims; severe open claims remain partially developed. A model built only on final paid claims can misstate prospective severity if it ignores closure selection. Claim-level reserve changes and settlement dynamics may therefore be required, especially for liability and other long-tailed lines.

An insurer using a neural conditional mean for capital assessment must attach additional assumptions concerning dispersion, high-loss distribution, dependence, and catastrophic event accumulation. The same conditional mean is compatible with many distinct tail outcomes.

Missing inputs and portfolio selection

Missingness can be a measurement artifact, a product feature, a privacy choice or a decision-dependent absence. Median imputation changes the implied data distribution; separate missing indicators may be informative if missingness and the training–deployment mechanism are stable. An underwriting predictor must not encode fields generated only after the claim occurs, even if their missingness status is highly predictive in historical extracts.

Policy selection creates another identification problem. A historical insurer observes claims mainly for risks previously accepted under its earlier underwriting standards. Rejected applications generally lack comparable insured claim outcomes. The conditional loss relationship for previously rejected applicants is not identified merely by fitting a nonlinear function to accepted-policy data. Extending quotations to a new segment therefore requires explicit transfer assumptions, supplementary evidence, or more cautious manual underwriting.

Bias can also arise from claims-reporting thresholds, data truncation, exposure errors and post-event backfills. A model may learn the probability that a loss is recorded by the insurer rather than the probability of its occurrence. These phenomena affect causal interpretation and operational pricing, even where purely predictive performance is high on similarly recorded historical data.

Validation with correlated risks

Resampling must respect the level at which observations are independent. Randomly distributing different policy-years of one customer across training and test sets can leak persistent behavior. Splitting claims from one catastrophe across folds can produce inflated geographic accuracy. Randomly withholding development cells within a triangle can expose information from the same origin-year outcome. Proper evaluation may split by customer, accident event, geography, accident year, or underwriting vintage according to the research question.

The over-dispersed-Poisson neural reserving paper discusses bootstrap prediction uncertainty in a particular reserve context. Its resampling assumptions cannot automatically be transported to cross-sectional underwriting. Bootstrap intervals from independent rows should not be treated as valid where large-scale loss dependence exists.

Prediction uncertainty contains process variability, estimation uncertainty, and structural or scenario uncertainty. These components can be separately assessed when suitable data and assumptions are available. A neural point forecast is one element of a stochastic insurance risk assessment, not a substitute for its distribution.

Calibration and Probabilistic Outputs

Forecast error, proper scoring and calibration

A predictor can be compared using mean squared error, deviance, negative log-likelihood, proper distributional scores or policy-relevant economic objectives. The right choice is determined by the target and information required. For counts with exposure, Poisson deviance measures a particular conditional mean discrepancy. For positive severities, a Gamma-based score or robust tail-oriented functional may be appropriate. For complete probability distributions, scoring rules should encourage accurate uncertainty rather than only accurate central values.

Gneiting and Raftery (2007) establish the role of proper scoring rules for probabilistic forecasts. Strict propriety means that the expected score favors truthful specification of the relevant forecast distribution under the rule's assumptions. This is particularly important if the forecast will support reserving capital or risk decisions whose consequences depend on the distribution beyond its mean.

A primary model-selection score should be specified before comparing many alternatives. Otherwise, selecting a metric that favors the desired model confounds model evaluation. Secondary metrics may document complementary qualities, but should not be presented as independent confirmatory outcomes when they were themselves chosen after inspecting results.

Calibration of annual frequency and portfolio totals

Given exposures e_i and fitted annual claim rates λ̂_i, the predicted total count is ∑_i e_iλ̂_i. The observed count is ∑_i N_i. Their ratio can provide a portfolio-level diagnostic. Agreement is useful, but it is not enough: a model may predict too many events in one segment and too few in another, with errors canceling in the total.

Conditional calibration can be evaluated by bands of fitted risk, exposure duration, product, geographic region, distribution channel, inception period, and other allowed business variables. For each segment, counts must be assessed against their expected variability and dependence structure. If a group has 100 exposure-years at expected rate 0.02, it expects only two claims. A realized count of zero or four does not, by itself, prove the model is badly calibrated.

Event probability calibration asks a different question. The network's annual frequency rate cannot be interpreted as a probability of at least one claim without a count distribution and exposure transformation. Under Poisson assumptions the map is 1−exp(−eλ). Under a negative-binomial distribution with the same mean and different variance, the zero probability generally differs. A correct rate fit and an incorrect assumed count distribution can therefore generate a miscalibrated event probability.

Guo and colleagues (2017) study neural classifier confidence calibration and evaluate temperature scaling in classifier applications. Their results are evidence that network classification confidence can be miscalibrated. They do not prove that temperature scaling fixes Poisson insurance rates, Gamma severity, or compound-loss quantiles. Calibration procedures must match the actuarial target and target distribution.

Sharpness and full predictive distributions

Gneiting, Balabdaoui, and Raftery (2007) distinguish forecast calibration and sharpness and investigate distributional diagnostics. Sharpness refers to concentration of predictive distributions, evaluated jointly with calibration. A very wide interval may achieve high empirical coverage but be operationally uninformative; an implausibly narrow interval may look attractive until uncommon losses occur.

For continuous outcomes, probability integral transform diagnostics can reveal systematic bias or variance misspecification. Discrete claim counts require adjustments such as randomized PIT to avoid mistaking discreteness for poor calibration. For insurance amounts with point mass at zero and a positive component, evaluation should respect both mass and severity. Portfolio dependencies also complicate interpreting independently produced policy-level predictive distributions.

If the business decision concerns aggregate tail capital, a count-and-mean-severity model requires additional information about positive severity variance and dependence. The compound Poisson identity Var(S)=μE[Y²] makes this explicit. The same mean loss can arise from frequent small claims or rare large ones, but these portfolios have different financial risks.

Recalibration without test leakage

A simple count-rate correction is c=(∑N_i)/(∑e_iλ̂_i), computed on a legitimate calibration sample with nonzero predicted total. Multiplying all fitted rates by c matches that sample's total. It does not correct mistakes in relative group rates. Estimating c on the final test data and reporting recalibrated test performance as if the predictions were frozen is information leakage.

A valid workflow uses training data to fit model parameters, a designated calibration or validation set to choose and recalibrate forecasts, and an untouched later test for final evaluation. Repeated cycles of adjustment against the same final test amount to additional training. When claims experience emerges slowly, the study may need multiple vintages and a maturity-aware evaluation scheme.

Calibration should be studied both statistically and in the final insurance decision. A fitted rate of 0.14 and a final premium change are linked by technical pricing rules, expenses, reinsurance, and commercial constraints. Testing the rate alone does not assess whether the tariff changes are acceptable or stable. Decision-relevant calibration requires visibility into these additional transformations.

Tail uncertainty and model use

Extreme aggregate loss events are too rare for a small selected portfolio to estimate every tail parameter from its own history. For underwriting with modest limits, a mean forecast may be the principal quantitative output. For capital, catastrophe aggregation, reinsurance design or extreme liability, a distributional or scenario model is necessary. A neural predictor may provide covariate-sensitive mean intensity or hazard parameters to that model, but cannot replace independently supported tail and dependency assumptions.

Quantile regression, parametric distributional networks, simulation, bootstrapped model refits and scenario analysis can each supply additional uncertainty information under different assumptions. They should be compared on the quantity required: conditional quantile coverage, predictive interval adequacy, portfolio capital sensitivity, or stability after environmental changes. Reporting a numerical interval without documenting its coverage population, horizon, and dependency structure offers limited actuarial evidence.

Insurance Benchmark Comparisons

Interpretation of insurance-specific experiments

the French motor case study compares statistical regression and machine-learning approaches for claims-frequency data. It supports the conclusion that interactions and flexible predictors can materially affect out-of-sample results within the case. It does not establish that one model class is universally superior across risk lines or even across different motor books.

The German Actuarial Association benchmark reports 678,031 French motor policies at the source stage and 678,013 after a cleaning operation to address claim amounts. It compares GLM variants, regularized GLMs, GAMs, deep nets, and tree-boosting families and includes cross-validation and deviance. In the studied implementation, certain boosted-tree approaches outperform the tested neural alternatives. This conclusion is specific to that portfolio, feature set, tuning, split and loss.

the frequency–severity four-dataset investigation evaluates GLM, boosted trees, feed-forward networks and CANN strategies and considers calibration and translation toward technical tariffs. Its detailed empirical setting illustrates the need to compare performance beyond one scalar deviance score. The present dossier attributes only the study scope and publicly accessible reported findings; it does not invent missing individual model score tables or claim to have rerun the authors' experiments.

The telematics longitudinal count study explores combined actuarial networks with repeated behavioral information; the individual claims-development study addresses incurred and paid amount trajectories. Their input structures, target horizons, and dependence are distinct. A gain found with longitudinal risk information cannot be attributed simply to adding neural depth when a static benchmark lacked the same temporal information.

Lessons from general tabular machine learning

The Grinsztajn–Oyallon–Varoquaux benchmark examines why tree-based methods perform strongly on many medium-scale tabular tasks, including the effects of feature structure and uninformative predictors. The later Holzmüller–Grinsztajn–Steinwart analysis demonstrates the competitiveness of carefully prepared and pre-tuned multilayer perceptrons with strong boosted-tree baselines in designated tabular benchmarks. The apparent difference is intelligible when evaluation protocols, architecture choices and training budgets are included.

Both studies concern cross-domain tabular data rather than a controlled multinational insurer trial. Insurance introduces exposure, claim development, coverage design and event dependence not present in many generic benchmarks. General tabular evidence is useful for preventing weak baselines, but it cannot authorize general claims about insurer pricing accuracy. A valid study compares models using the same target and information available at the actual decision date.

Design of a reproducible actuarial comparison

An experiment should begin with a question such as: which candidate predicts twelve-month reported motor claim counts for new insurance policies, using only legitimate quotation-time information, on policies written after the training period? The research specification must fix the target, exposure definition, insured population, calendar periods, valuation conventions and metrics before training. Changing the goal mid-experiment invalidates a simple winner narrative.

A comprehensive benchmark can include an intercept-only portfolio rate, a fitted exposure-offset GLM, a tuned GAM, a well-prepared boosted-tree learner, a feed-forward neural network, and a CANN with a documented GLM baseline. Every candidate must have a justified search budget. All preprocessing, categorical regrouping, imputation and normalization must be learned within training folds. The final test should be time-separated from all fitting and tuning, with additional insured-level, region-level or catastrophic-event holdouts if the intended use requires transport.

A single random row split is often inadequate. Multiple policy-years from one insured can be correlated; claims from the same weather event are shared exposures; a changing product version can shift the target mechanism. Resampling or uncertainty estimation should respect these clusters and periods. Tuning for a target metric should not silently use the final test population.

The report should retain primary and secondary evaluation criteria: exposure-weighted count deviance, aggregate predicted versus observed events, group and time calibration, positive severity prediction, pure-premium adequacy, tail-sensitive performance where relevant, predictive variance, stability of tariff relativities, fairness and economic consequences. Runtime and reproducibility may also matter. Results should include uncertainty for performance differences and an account of selection among tried candidates.

Why an average score is an incomplete decision rule

Suppose two hypothetical models yield similar Poisson deviance. The first predicts steady relativities and can be represented by an interpretable tariff table. The second achieves a marginally smaller test score but shifts rare occupations or postal codes sharply. An insurer may need to examine thin-class credibility, underwriting consequences, subgroup fairness and operational override costs before determining whether the incremental fit is useful.

A different distinction appears between small and large claims. A model could improve frequent low-cost loss predictions while missing extraordinary payments that dominate capital risk. It may still be appropriate for a narrow average-cost application but inadequate for a solvency or reinsurance decision. The business function and required risk measure should determine the evaluation protocol.

The original analytical interpretation of the benchmark literature is therefore that the relevant comparison is between complete model-and-decision systems. These systems include the data cohort, target definition, loss, preprocessing, architecture, calibration method, governance restrictions and operational output. Ranking architectures alone suppresses most of the conditions that make actuarial evidence transferable.

Evidence needed before a commercial implementation

A finished insurer research report should attach the model inventory, source data and permission basis, statistical assumptions, variable availability dates, contract definitions, exposure reconciliation, split indices, complete search history, frozen prediction files, loss calculations, subgroup calibration, uncertainty estimates, stress tests and implementation review. It should distinguish source-reported findings from independently reproduced numbers and should preserve raw model scores before tariff adjustments.

This dossier supplies that specification as an editorial analytical synthesis. It does not manufacture score tables, train fictional models, or attribute a single winner to every insurance sector.

Extrapolation, Fairness, and Controls

Portfolio drift, support and extrapolation

Neural networks learn relations on the support of their training information. When the new portfolio contains previously unseen products, regions, claims inflation or behavioral patterns, outputs reflect trained functions and structural assumptions rather than new empirical evidence. Smooth extrapolation on age or insured value does not show that claim hazards remain smooth in a new demographic or high-limit segment.

A change in observed policy features P(X) can alter the frequency of inputs even if conditional claim behavior remains stable. A change in P(Y|X) can affect expected loss at the same features. In insurance, exposure and reporting changes can shift the observed target independently of underlying risk. Monitoring should separate composition effects, conditional relationship changes, and measurement artifacts.

Time-based holdouts and forward-looking stress tests are important where economic inflation, law, climate or underwriting practice evolves. New category handling and feature-range limits should be included in the model use contract. Inputs beyond a designated support boundary can trigger review or fallback to a vetted traditional method. These controls should be treated as explicit implementation choices, not as guarantees of statistical truth.

Fairness under risk classification

Insurance prediction affects premiums and sometimes access to coverage. A correlation useful for a predictive loss model may also encode an unfair or prohibited differentiation. Excluding a protected attribute does not necessarily remove its information because geography, occupation, commercial transactions or behavioral features may serve as proxies. Complex learned embeddings can produce indirect discrimination even when the displayed feature list appears innocuous.

Lindholm and colleagues (2022) distinguish direct and indirect discrimination and analyze a statistical discrimination-free pricing approach. Their subsequent multitask-network investigation develops pricing estimation when protected information is partly missing. The paper demonstrates concrete methodological choices, while the applicable legal framework, permission to process sensitive data, and societal fairness goals must be established in the relevant jurisdiction.

Different notions of fairness can conflict in settings with different observed loss rates, selected insured populations, uneven data quality or disparate access. A responsible evaluation identifies the precise protected classes, legal rules, outcome being judged, counterfactual or statistical standard used, and the effect on tariff outcomes. It should distinguish the actuarial best estimate from an insurer's final pricing and pooling decisions.

The IAIS application paper discusses fairness, ethics and consumer redress as part of a proportionate insurance AI supervisory perspective. A highly granular predictor can improve measured fit and still create unacceptable concentration of prices or exclusion among certain groups. Such effects require explicit consideration alongside statistical error.

Explanations and their limits

Lundberg and Lee (2017) develop additive feature attribution methods that can describe contributions to an individual model prediction under a reference specification. These contributions help inspect neural behavior and identify proxy features or unexpected interactions. They are model-attribution quantities, not direct causal effects of interventions on policyholder risk.

A surrogate GLM can translate parts of a neural function into a tariff structure; its ability to approximate the source model must be evaluated on held-out predictions and relevant risk segments. An explanation of a final underwriting decision may need to include business rules and overrides as well as the neural prediction. A high expected loss score is not evidence that a claimant is dishonest, and a predictive feature is not independent proof of responsibility for an insured event.

Effective explanation has an audience. Internal actuaries may require residual deviance, calibration and partial-dependence diagnostics; supervisors may require source provenance, model governance and adverse-impact evidence; policyholders may require a plain-language account of a material price or coverage decision. These obligations require different levels of detail but must all remain faithful to the real decision process.

Supervisory documents and their legal functions

The IAIS 2025 AI application paper describes risk-based supervision and proportionality, governance and accountability, robustness, security, transparency, explainability, fairness and redress under existing Insurance Core Principles. It provides international supervisory guidance, while binding operational rules arise under applicable domestic law and regulation.

EIOPA's 2025 governance opinion addresses EU insurance-sector AI oversight within the opinion's own defined scope, which expressly differs from the prohibited or high-risk categories under the EU AI Act. The EU Artificial Intelligence Act identifies certain life and health insurance risk assessment and pricing systems in Annex III among high-risk uses, subject to statutory qualifications, implementation dates and amendments. The General Data Protection Regulation also governs lawful personal-data processing and contains Article 22 safeguards concerning certain solely automated decisions of legal or similarly significant effect. Real deployment requires an as-of-date analysis of role, application, legal basis and relevant domestic implementation.

In the United States, the NAIC model bulletin expresses supervisory expectations for AI governance, documentation and unfair discrimination under existing insurer regulation. Its legal effect is not uniform across states merely because the NAIC adopted a model bulletin. Firms should assess the actual rules and requirements of each relevant jurisdiction.

The NIST AI Risk Management Framework supplies voluntary cross-sector functions for governing, mapping, measuring and managing AI risks. These provide a helpful organizational structure alongside actuarial, prudential and consumer-protection controls, but they do not replace substantive insurance laws or mathematical model validation.

A documented lifecycle for neural insurance models

A deployable model inventory should record responsible owner, approved purpose, scientific target, data lineage, exposure basis, training and valuation dates, model version, permissible information fields, validation evidence, known limitations, and review authority. Any new data source, classification category or substantial model retraining can change actual risk outcomes and should activate change management proportional to its consequences.

Monitoring begins with input integrity: exposure distributions, missingness, unknown categories, policy mix, inconsistent timestamps and major data revisions. It proceeds to outputs: prediction distributions, risk relativities, override frequency, customer effects and divergence from prior approved models. Finally, as outcomes develop, monitoring assesses cohort claim counts, incurred and paid losses, subgroup calibration and adverse decision effects. Reporting lag should be documented, because early claims experience can be a biased representation of ultimate losses.

Escalation triggers may include a material change in claims definition, significant deviation in a large insured segment, evidence of unfair discrimination, a newly restricted data feature, or severe mismatch in catastrophe-affected cohorts. Fallback may mean a previously validated GLM, direct human review, a separate stochastic capital model, or a temporary restriction of model use. The fallback must itself be technically available and legally appropriate.

Audit logs should preserve the input record and timestamp, the selected version, the original neural output, postprocessing and tariff rules, any human override, and the final decision. Such traces permit dispute resolution and reproducibility. External vendor models require the same substantive attention even if some implementation internals are contractually inaccessible.

Limits of the present evidence and research directions

The reviewed studies address different data sets, statistical outcomes and validation designs; they do not form one globally controlled model comparison. Some publication records were accessible at abstract or publisher metadata level and were treated accordingly. No independent experimental network was fitted in this manuscript, and no proprietary insurer file was accessed. The numerical examples are hypothetical mathematical demonstrations, not real-policy predictions.

Future primary empirical studies can compare the same actuarial GLM, GAM, boosting, feed-forward and CANN candidates across risk lines and economic regimes with predeclared targets, matching tuning budgets, proper out-of-time holdouts, uncertainty intervals, and lawful fairness audits. They could measure how results change when reported claims are replaced by mature incurred losses, when major event dependence is modeled, and when distributional calibration rather than point-loss accuracy governs evaluation. Such research requires explicit data access and privacy provisions and should preserve the distinction between insurance operations and scientific experimentation.

The present analytical synthesis produces an integrated checklist of research questions and controls using existing statistical and supervisory methods. It introduces no newly named theoretical system or regulatory instrument.

Conclusion

Neural networks make a contribution to insurance risk prediction when they estimate a scientifically specified conditional target and demonstrate useful out-of-sample performance for the intended portfolio and decision. Their mathematical flexibility is especially relevant to interactions, heterogeneous categorical risks, sequential claims experience and residual correction of established actuarial models. Its value is measurable only through the loss distribution, exposure convention, data maturity, evaluation horizon and controls attached to the prediction.

The first conclusion is statistical: annual claim rate, expected count, event probability, positive severity, aggregate expected loss and capital-tail quantiles are different quantities. The offset Poisson model expresses an exposure-adjusted rate under explicit assumptions; frequency–severity factorization requires conditional assumptions about claims; loss valuation and reporting histories must match the prediction date. Neural architecture cannot repair ambiguity in any of these definitions.

The second conclusion is methodological: neural and combined actuarial architectures should be compared with strong GLM, GAM and boosted-tree baselines. Published French motor, four-dataset pricing, longitudinal telematics and reserving studies establish significant case-specific methods and outcomes. General tabular benchmarks also demonstrate that model family rankings change with data structure and training design. No universal ranking follows.

The third conclusion concerns statistical evidence: accurate mean predictions and matching portfolio totals do not establish calibrated event probabilities, group fairness, valid severity tails or a correct aggregate loss distribution. Proper scoring, group and time calibration, exposure-aware comparison, dependence-sensitive uncertainty, and untouched test periods establish the meaning of each measured improvement. Decision makers need distributions and stress assumptions when the objective extends to capital or catastrophe losses.

The fourth conclusion concerns transport: a model trained on historically accepted risks, earlier coverages and earlier claims administration estimates a relation within that record. New products, new risk classes, claims inflation, reporting changes, economic stress, and evolving legal requirements can change the target mechanism. Model monitoring, support boundaries and an operational fallback are elements of its substantive scientific reliability.

The fifth conclusion is institutional: granular predictive models can generate proxy discrimination and opaque rate effects. Fairness research, explanation methods and international insurance supervision offer usable analytic and control instruments. Applicable law, legal force and national implementation remain specific; they cannot be inferred from a model's statistical fit.

This monograph contributes an original source-attributed integration of stochastic loss mathematics, neural modeling, comparative actuarial evidence, calibration and insurance governance. Its hypothetical derivations and benchmark specification make the conclusions operationally testable. Neural networks are consequently judged by the rigor of the insurance question, the credibility of held-out evidence, the stability of the resulting forecast, and the accountability of its economic use.

References

  • Blier-Wong, Christopher, Hélène Cossette, Luc Lamontagne, and Etienne Marceau (2021). “Machine Learning in P&C Insurance: A Review for Pricing and Reserving.” Risks 9(1): article 4. Official publication, DOI, or source record.

  • Delong, Łukasz, and Mario V. Wüthrich (2020). “Neural Networks for the Joint Development of Individual Payments and Claim Incurred.” Risks 8(2): article 33. Official publication, DOI, or source record.

  • Duval, Francis, Jean-Philippe Boucher, and Mathieu Pigeon (2024). “Telematics Combined Actuarial Neural Networks for Cross-Sectional and Longitudinal Claim Count Data.” ASTIN Bulletin 54: 239–262. Official publication, DOI, or source record.

  • European Insurance and Occupational Pensions Authority (2025). “Opinion on Artificial Intelligence Governance and Risk Management.” EIOPA-BoS-25-360, August 6, 2025. Official publication, DOI, or source record.

  • European Parliament and Council of the European Union (2016). “Regulation (EU) 2016/679: General Data Protection Regulation.” Official Journal of the European Union L 119, May 4, 2016. Official publication, DOI, or source record.

  • European Parliament and Council of the European Union (2024). “Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence.” Official Journal of the European Union, July 12, 2024. Official publication, DOI, or source record.

  • Frees, Edward W., Gee Lee, and Lu Yang (2016). “Multivariate Frequency-Severity Regression Models in Insurance.” Risks 4(1): article 4. Official publication, DOI, or source record.

  • Gabrielli, Andrea, Ronald Richman, and Mario V. Wüthrich (2020). “Neural Network Embedding of the Over-Dispersed Poisson Reserving Model.” Scandinavian Actuarial Journal 2020(1): 1–29. Official publication, DOI, or source record.

  • Gneiting, Tilmann, and Adrian E. Raftery (2007). “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102(477): 359–378. Official publication, DOI, or source record.

  • Gneiting, Tilmann, Fadoua Balabdaoui, and Adrian E. Raftery (2007). “Probabilistic Forecasts, Calibration and Sharpness.” Journal of the Royal Statistical Society, Series B 69(2): 243–268. Official publication, DOI, or source record.

  • Grinsztajn, Léo, Edouard Oyallon, and Gaël Varoquaux (2022). “Why Do Tree-Based Models Still Outperform Deep Learning on Typical Tabular Data?.” Advances in Neural Information Processing Systems 35, Datasets and Benchmarks Track. Official publication, DOI, or source record.

  • Guo, Chuan, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger (2017). “On Calibration of Modern Neural Networks.” Proceedings of Machine Learning Research 70: 1321–1330. Official publication, DOI, or source record.

  • Holvoet, Freek, Katrien Antonio, and Roel Henckaerts (2025). “Neural Networks for Insurance Pricing with Frequency and Severity Data: A Benchmark Study from Data Preprocessing to Technical Tariff.” North American Actuarial Journal 29(3): 519–562. Official publication, DOI, or source record.

  • Holzmüller, David, Léo Grinsztajn, and Ingo Steinwart (2024). “Better by Default: Strong Pre-Tuned MLPs and Boosted Trees on Tabular Data.” Advances in Neural Information Processing Systems 37. Official publication, DOI, or source record.

  • International Association of Insurance Supervisors (2025). “Application Paper on the Supervision of Artificial Intelligence.” IAIS, July 2, 2025. Official publication, DOI, or source record.

  • König, Daniel, and Friedrich Loser (2024). “Claim Frequency Modeling in Insurance Pricing Using GLM, Deep Learning, and Gradient Boosting.” Deutsche Aktuarvereinigung, Actuarial Data Science, January 3, 2024. Official publication, DOI, or source record.

  • Lindholm, Mathias, Ronald Richman, Andreas Tsanakas, and Mario V. Wüthrich (2022). “Discrimination-Free Insurance Pricing.” ASTIN Bulletin 52(1): 55–89. Official publication, DOI, or source record.

  • Lindholm, Mathias, Ronald Richman, Andreas Tsanakas, and Mario V. Wüthrich (2024). “A Multi-Task Network Approach for Calculating Discrimination-Free Insurance Prices.” European Actuarial Journal 14: 329–369. Official publication, DOI, or source record.

  • Lundberg, Scott M., and Su-In Lee (2017). “A Unified Approach to Interpreting Model Predictions.” Advances in Neural Information Processing Systems 30: 4765–4774. Official publication, DOI, or source record.

  • National Association of Insurance Commissioners (2023). “Model Bulletin on the Use of Artificial Intelligence Systems by Insurers.” NAIC, adopted December 4, 2023. Official publication, DOI, or source record.

  • Noll, Alexander, Robert Salzmann, and Mario V. Wüthrich (2020). “Case Study: French Motor Third-Party Liability Claims.” SSRN working paper, revised 2020. Official publication, DOI, or source record.

  • Schelldorfer, Jürg, and Mario V. Wüthrich (2019). “Nesting Classical Actuarial Models into Neural Networks.” SSRN working paper. Official publication, DOI, or source record.

  • Tabassi, Elham; National Institute of Standards and Technology (2023). “Artificial Intelligence Risk Management Framework (AI RMF 1.0).” NIST AI 100-1, January 26, 2023. Official publication, DOI, or source record.

  • Wüthrich, Mario V., and Michael Merz (2023). “Statistical Foundations of Actuarial Learning and its Applications.” Springer Actuarial. Cham: Springer, 605 pp. Official publication, DOI, or source record.