SCIENCE
Insurance Statistics and Predictive Modeling
Abstract. Insurance statistics and predictive modeling studies insured-event frequencies, claim severities, aggregate losses, durations, reporting processes and contingent financial outcomes under incomplete, selected and evolving observation. This monograph provides an integrated historical, mathematical, methodological and institutional analysis of classical actuarial statistics and contemporary predictive learning. It formalizes exposure-based count modeling, conditional frequency–severity decomposition, compound distributions, survival and development processes, credibility and hierarchical inference, generalized linear models, algorithmic ensembles, calibration, proper scoring, uncertainty and dependence. It examines causal identification, selection, telematics, images, fairness, explanation and internationally differentiated supervision. The analytical synthesis establishes predictive reliability as a property of the entire chain from contractual estimand and recorded evidence to forward validation and accountable use. Claims about empirical comparisons are restricted to the cited studies and their populations; illustrative calculations use hypothetical examples. No original statistical theory or external peer-review claim is made.
Keywords: Insurance statistics; predictive modeling; actuarial data science; claim frequency; claim severity; compound Poisson; Tweedie; credibility; generalized linear models; loss development; survival analysis; hierarchical models; gradient boosting; random forests; predictive calibration; proper scoring rules; model uncertainty; distribution shift; explainability; fairness; insurance governance.
Introduction: Scope and Scientific Object
Insurance statistics is the systematic study of how insured events, exposures, contract terms, reporting processes, and financial consequences can be observed, measured, and inferred from data. Insurance predictive modeling uses those measurements to estimate conditional distributions of future events and payments under explicitly specified information, population, policy conditions, and time horizons. Its objects include claim occurrence, claim counts, individual severities, aggregate losses, payment timing, mortality and morbidity transitions, reporting delays, policyholder behavior, and insurer operational outcomes.
The discipline joins statistical inference with insurance's distinctive observation structure. An event enters a data set through policy inception, risk attachment, occurrence, notification, coverage interpretation, adjustment, payment, reopening, and closure. Each step may change what is recorded and when. A claim amount is therefore a measured outcome of an economic and legal process as well as a random financial loss. An observed zero may mean no event, an excluded event, an unreported event, or an event not yet developed. A forecast has scientific content only when its target, unit of observation, observation window, conditioning information, and uncertainty are specified.
Three commitments organize this monograph. First, the target distribution and observation process precede model choice. Second, predictions are judged through out-of-sample evidence, calibration, stability, and the decisions to which they contribute. Third, actuarial interpretation requires explicit treatment of contract boundaries, aggregation, dependence, and applicable professional or supervisory obligations. This framework synthesizes established statistical and actuarial literature rather than proposing a new statistical theory. It locates insurance statistics beside Risk Theory, Actuarial Science, Insurance Pricing and Reserving, Solvency and Capital, Claims Science, and Insurance Technology in the encyclopedia's approved architecture.
The scope is international and comparative. Standards issued in the United States or the European Union are identified as such; they are examples of jurisdictional governance, not globally binding law. Technical formulas describe mathematical populations under stated assumptions, not operational price recommendations.
Historical Development and Intellectual Traditions
Insurance statistics arose through the convergence of mortality tables, probability theory, commercial recordkeeping, collective-risk calculation, and the empirical classification of insured exposures. Early life contingencies encouraged the use of age-specific experience tables and survival probabilities; non-life insurance directed attention toward event counts, heterogeneous loss sizes, and accumulating claims. The historical transition was gradual: deterministic tabulations, empirical frequency methods, Bayesian reasoning and credibility, statistical regression, stochastic claims development, and computational learning have overlapping rather than exclusive domains.
The actuarial credibility tradition asks how experience from a particular risk should be balanced against collective experience. Bühlmann's published investigations provide influential formal and applied accounts of this question; importantly, the Casualty Actuarial Society identifies a 1967 article in ASTIN Bulletin 4(3), whereas Cambridge identifies a separate 1969 article in volume 5(2). Both belong to the history and require distinct attribution (see Bühlmann (1967); Bühlmann (1969)). Credibility subsequently connects naturally to hierarchical statistics and empirical Bayes estimation.
The regression tradition gained a general mathematical form with Nelder and Wedderburn's 1972 formulation of generalized linear models. Its combination of a probability family, systematic predictor, and link function allowed heterogeneity to be modeled with interpretable parameters (see Nelder (1972)). Insurance applications now encompass counts, conditional loss severity, and aggregate premium components; contemporary CAS practice texts still treat GLMs as an operational cornerstone (see Goldburd (2025)).
A third line concerns incomplete observations and development over time. Kaplan and Meier's 1958 product-limit estimator and Cox's 1972 hazard regression established essential methods for censored durations. Stochastic loss-reserving work, including Mack's 1993 prediction-error analysis, required practitioners to acknowledge uncertainty in outstanding claims rather than rely exclusively on point estimates (see Mack (1993); Wüthrich (2008)).
The computational learning tradition broadened the set of admissible response surfaces. Breiman's 2001 discussion of statistical cultures contrasted explicit stochastic modeling with algorithmic prediction, while his random forests and Friedman's boosting methods introduced flexible, regularized ensemble tools. Their influence in insurance follows from empirical predictive performance, while the discipline's obligations to explainability and financial accountability remain important (see Noll (2020); Aas (2024)). The scientific history is consequently one of expanding methods applied to a persistent inference problem: incomplete, heterogeneous, evolving evidence about contingent obligations.
Definitions, Units of Analysis, and Estimands
A risk unit is the entity or exposure to which a specified hazard and coverage apply during a measurement interval. Depending on the contract, the unit may be a vehicle-year, life-year, insured building, insured crop area, insured transaction, policy, claim, or coverage layer. These units cannot be exchanged silently. An exposure measure is the denominator representing time or magnitude at risk; a rating factor is an explanatory characteristic; a target variable is an observed realization of the quantity to be predicted.
An estimand is the precisely defined population quantity sought. Common estimands include expected claim count per earned exposure, conditional mean positive claim cost, expected aggregate ground-up loss, conditional distribution of paid or incurred loss at a chosen valuation date, probability of lapse in a specified interval, and a tail quantile of portfolio losses. The observable target may differ from an ultimate or counterfactual target.
Let be the number of covered claims for exposure during a stated interval; let be the financial severity of the -th covered claim under a stated valuation basis. The aggregate covered loss is
$ S_i=\sum_{j=1}^{N_i}X_{ij},\qquad S_i=0\ \text{if}\ N_i=0. $
If the claim severities are conditionally identically distributed and conditionally independent of the claim count given risk characteristics , then
$ \operatorname{E}[S_i\mid Z_i]=\operatorname{E}[N_i\mid Z_i]\operatorname{E}[X_{ij}\mid Z_i]. $
The factorization is conditional on these assumptions and the same definition of covered loss. Without conditional independence, the joint distribution and dependence structure determine the expectation. Moreover, taking unconditional expectations across heterogeneous classes need not equal the product of unconditional average frequency and severity.
A prediction uses information actually available at the forecast origin. A pricing prediction typically concerns a prospective coverage period; a reserve prediction concerns claims whose eventual ultimate amounts are incompletely observed. An underwriting score, catastrophe simulation, fraud flag, and probability of policy lapse have distinct outcome definitions and evaluation criteria. The discipline's first scientific task is to preserve those distinctions.
Insurance Data Generation and Measurement
Insurance databases are produced by policies, endorsements, exposures, transactions, event reports, claim files, payments, recoveries, and external data. A reliable statistical data model identifies the entity key, time axis, valuation date, currency, coverage definition, settlement basis, policy term, and underlying administrative purpose of each field. The same numerical field can represent paid loss, incurred loss, case reserve, gross-of-reinsurance loss, net-of-reinsurance loss, or a loss after deductibles and limits. Summation across inconsistent bases creates an invalid outcome.
Time is multidimensional. Policy period locates contractual attachment; occurrence or accident date locates the event; reporting date marks recognition; transaction date records changes; valuation date is the as-of point at which development is known; accounting date controls financial statements. Long-tail claims may be measured years after the underlying incident. Data extracted today must reconstruct the information state available at a historical forecast date to prevent future-information leakage.
Exposure must be earned on a consistent time basis. A vehicle insured for six months ordinarily contributes approximately half a vehicle-year when exposure is represented by covered duration, subject to conventions and policy changes. Portfolio counts, sum insured, payroll, revenue, miles driven, and other bases can be appropriate for different questions; they are not equivalent by definition. Premium earned is a monetary revenue measure and should not automatically substitute for physical exposure.
Censoring, truncation and missingness require different treatments. Right-censoring means that an event time is incompletely observed after follow-up ends. Left-truncation means only subjects surviving past an entry condition are sampled. Claims reported with delay are subject to development and selection. Deductibles can create a lower observation threshold for reported costs; limits can cap observed insurer payments. Missing fields may reflect reporting design, contractual selection, or strategic behavior rather than random omission. Data-quality assessment should document coverage, accuracy, relevance, consistency, lineage, completeness, duplicates and the plausible missingness mechanism, with deviations evaluated against the intended estimand. ASOP No. 23 is a profession-specific United States example of disciplined data selection and review (see Actuarial Standards Board (2016)).
Third-party or externally generated information, including weather, satellite observations, geospatial images, telemetry, text and network data, can enrich risk measurement. Its collection consent, provenance, temporal matching, selection mechanism, security and representativeness affect any statistical claim. The existence of a variable in an external feed does not establish that it can lawfully or causally support a particular insurance decision.
Descriptive Statistics, Standardization, and Exploratory Analysis
Before fitting a model, the analyst should establish what the data contain. A descriptive analysis reports exposure, number of claims, proportion of policies with at least one claim, claim severity among positive claims, aggregate loss, development maturity, changes in mix, and the distribution of policy terms. A single portfolio average can conceal shifts in coverage or geography and therefore requires stratified views.
For a homogeneous observation basis, the portfolio frequency estimate is . Positive-claim severity may be summarized by , provided those claims are observed on a consistent basis and sufficiently mature. Their product equals algebraically for the same observations, but the identity of arithmetic ratios does not establish independent frequency and severity processes, predictive stability, or adequacy of reserves.
Loss ratios compare defined claims measures with defined premium measures. A paid-to-earned ratio, ultimate-to-earned ratio, gross loss ratio and net loss ratio answer different questions. Adjustments for inflation, exposure growth, rate changes, coverage terms and catastrophe years may be essential to meaningful temporal comparisons.
Exploratory diagnostics include frequency histograms, log-scale severity tails, quantile plots, cumulative development triangles, seasonal plots, residuals against exposure, frequency by rating cell, and spatial correlation. Visual abnormalities become hypotheses for a traceable investigation of data, model or operational process. Data exclusions and transformations should be documented with the business rationale and their effect on the target population.
Probability Foundations and Statistical Inference
The probabilistic core of predictive insurance science consists of conditional distributions, repeated-sampling inference, Bayesian updating, stochastic-process modeling, and uncertainty propagation. A model posits a family for an outcome , covariates , and the available information set . Estimation identifies parameters from a sample; prediction integrates parameter uncertainty and future outcome randomness as appropriate.
Likelihood estimation connects a sampling model to observed data. In a Poisson count model with , a basic log-likelihood is
$ \ell(\beta)=\sum_i{N_i\log(e_i\lambda_i)-e_i\lambda_i-\log(N_i!)}. $
Conditional independence across units is assumed in this expression. Spatial storms, multi-vehicle incidents, common insureds or repeated measurements violate that assumption unless dependence is modeled. A link such as introduces an exposure offset in the predictor of expected count, rather than treating exposure as an unconstrained rating variable.
Frequentist confidence intervals concern repeated-sample procedures under a statistical model; Bayesian credible intervals concern posterior parameter probability under prior and likelihood assumptions; predictive intervals concern future observable outcomes. These uncertainties should be described separately. Parameter precision alone is insufficient when claim process variance, selection effects, dependence or model misspecification dominate.
Sampling variability, estimation variability, model selection, scenario uncertainty, process variability and structural uncertainty all matter. A credible model validation documents their inclusion and exclusions. Insurance decisions may require tail or interval forecasts even when a central conditional expectation is reasonably well estimated.
Claim Frequency Models
Claim counts are nonnegative integers observed over defined exposure. The Poisson model assumes conditional mean equal to conditional variance, independent increments under its process interpretation, and a rate proportional to exposure at fixed covariates. For policy :
$ N_i\mid x_i,e_i\sim\operatorname{Poisson}(e_i\exp(x_i^\top\beta)),\qquad \log\operatorname{E}[N_i\mid x_i,e_i]=\log e_i+x_i^\top\beta. $
The multiplicative relativities of the log link are interpretable as conditional rate ratios. Overdispersion may arise from unobserved heterogeneity, contagion, clustering or dependence. A negative-binomial model permits variance exceeding its mean, commonly written under one parameterization with . Quasi-Poisson approaches can relax the mean–variance relationship for estimating coefficients and standard errors but do not themselves supply a complete probability distribution.
Zero-inflated models combine an additional zero-generating process with a count model; hurdle models separate zero occurrence from positive count size. Distinguishing statistical excess zeros from contractual thresholds, unreported claims and short exposure is vital. A flexible zero mechanism can mimic ordinary unobserved heterogeneity and should be validated for identifiability and operational meaning.
Repeated insured-year data motivate longitudinal generalized estimating equations, random effects or multi-level Bayesian models. Catastrophe-correlated counts motivate event or spatial processes instead of independent policy-level Poisson observations. Model comparison should include calibration by exposure and count tail, residual dispersion, concentration of forecast errors, and held-out temporal cohorts (see Nelder (1972); Goldburd (2025)).
Severity Models, Truncation, Limits, and Heavy Tails
Claim severity is usually a positive, skewed amount on a specified gross/net and paid/incurred basis. Lognormal, gamma, Weibull and inverse Gaussian models provide different hazard and tail shapes. A generalized Pareto tail may approximate exceedances above a sufficiently high threshold under extreme-value conditions. A tail fit must be assessed for threshold stability, event heterogeneity and finite-sample uncertainty; extrapolation far beyond observations is especially sensitive to assumptions.
For a severity random variable with deductible and coverage limit on payment above the deductible, an illustrative contractual payment is
$ Y=\min{\max(X-d,0),,\ell}. $
Here is a maximum insurer payment after the deductible, and other contract clauses, policy aggregates, coinsurance, inflation indexing or multiple layers may modify the expression. The distribution of contains a mass at zero and often a mass at the upper payment limit, even when is continuous. Fitting an ordinary continuous gamma distribution directly to such observed payments can therefore be misspecified.
When a database contains only claims above a reporting threshold , fitting a ground-up severity model generally requires a truncation-adjusted likelihood; treating the recorded sample as complete exaggerates the body of the loss distribution. Conversely, censored limits require a probability mass or survival likelihood contribution for observations known only to exceed a threshold. Separating contractual censoring from ordinary missingness preserves the target estimand.
Tail dependence across risks can make insurer aggregate exposures much more volatile than independent-loss calculations suggest. For natural catastrophes, high-severity scenarios may be rare in historical data but economically decisive. A scientifically responsible tail forecast should identify data support, scenario assumptions, extreme-value model choice, and uncertainty in the shape of the tail rather than present a single quantile as an observed fact.
Aggregate Loss and Compound Distributions
Aggregate covered loss links frequency and severity. For a compound Poisson model with independent of i.i.d. severities with finite second moment,
$ \operatorname{E}[S]=\lambda\operatorname{E}[X],\qquad \operatorname{Var}(S)=\lambda\operatorname{E}[X^2] =\lambda{\operatorname{Var}(X)+\operatorname{E}[X]^2}. $
The variance formula depends on independence and finite second moments. It is not because variation in claim counts contributes an additional term. A frequency–severity covariance, common catastrophe factor or dependence across policyholders changes these formulas.
A Tweedie exponential dispersion model with variance function , , has a compound-Poisson–gamma representation with positive mass at zero and continuous positive outcomes. Jørgensen and Paes de Souza established insurance applications; later dispersion-focused work addressed the simultaneous modeling of expected cost and variation. The variance index, link and exposure weighting require attention: models for total loss and loss per exposure have different exposure conventions.
When closed-form compound distributions are difficult, recursive aggregation, numerical convolution, Fourier methods and Monte Carlo simulation can estimate aggregate distributions. Monte Carlo uncertainty should be distinguished from sampling and model uncertainty. The simulation's random seed, input distributions, dependence specification, and number of replicates must be recorded. Reinsurance attachment points and caps must be applied to the correct aggregation unit; a per-risk excess layer differs from a per-event or aggregate cover.
Credibility, Hierarchical Models, and Bayesian Updating
Credibility methods formalize partial pooling between limited individual experience and larger collective experience. In a simplified Bühlmann-style setting, the credibility estimate takes the form
$ \widehat m_i=Z_i\overline Y_i+(1-Z_i)m,\qquad Z_i=\frac{n_i}{n_i+K}, $
where is a collective mean, an observed mean for the risk, and a nonnegative ratio of process variance to between-risk variance under the model's assumptions. At , the estimate equals the collective mean; increasing information raises the weight on individual experience. Heterogeneous exposures and heterogeneous conditional variances require richer credibility formulations, not mechanical use of the simple factor (see Bühlmann (1967); Bühlmann (1969)).
Hierarchical GLMs introduce group-level effects for employers, fleets, territories, lines of business or agents; shrinkage avoids assigning extreme estimates to sparse cells. Bayesian methods represent uncertainty through prior distributions, likelihoods and posteriors, while empirical Bayes estimates pooling parameters from the observed data. These approaches are especially useful when sparse classes coexist with substantial portfolio-wide regularity.
The substantive judgment is whether two groups are plausibly exchangeable conditional on modeled characteristics. Pooling across risks with different coverages, claim development, legal environments or data generation may create false precision. Posterior predictive checks, sensitivity to priors, cross-group validation, and analysis of correlated effects are integral to credible inference.
Generalized Linear, Additive, and Penalized Regression
A generalized linear model comprises a random component from an exponential family, a linear predictor (including justified offsets), and a link connecting the conditional mean to that predictor (see Nelder (1972)). Log-linked Poisson frequency, log-linked gamma positive severity, and logistic event probability are established combinations. Variance specifications and the admissibility of predicted means must match the chosen response.
Generalized additive models replace selected linear effects by smooth functions, allowing continuous nonlinear relationships while preserving additive structure. Interactions, splines, spatial smooths and seasonality can be introduced where the data support them. Penalized likelihood adds a complexity charge to reduce variance: ridge penalization shrinks coefficient magnitudes; lasso can shrink coefficients to zero; elastic-net combinations help with correlated predictors. Penalization parameters must be tuned within the training data, not on the final holdout.
High-cardinality categorical variables—vehicle model, occupation, postal area, service provider—pose sparse-cell, leakage and overfitting problems. Partial pooling, validated embeddings, target encoding done strictly within training folds, and carefully structured categorical effects offer different compromises. The 2026 CAS Monograph 16 specifically treats high-cardinality factors, nonlinearities, collinearity and the blending of model components in insurance work (see Chalk (2026); Casualty Actuarial Society (2026)).
GLM coefficient interpretability is conditional on transformations, omitted interactions, collinearity and measurement quality. A rate ratio does not become a causal effect by virtue of a log link. Scientific reporting should include model matrix construction, offsets, families, dispersion, penalization, term selection, confidence or uncertainty estimates, and out-of-sample evaluation (see Goldburd (2025); Chalk (2026)).
Trees, Ensembles, Neural Networks, and Automated Learning
Decision trees partition predictor space into regions. Ensembles can improve generalization by combining diverse partitions. Random forests use randomized trees to reduce predictive variance relative to unstable single-tree rules (see Breiman (2001)). Gradient boosting builds an additive predictor sequentially by optimizing a loss in function space; its behavior depends on the chosen loss, learning rate, tree complexity and regularization (see Friedman (2001)). Neural networks offer flexible representations for interactions and high-dimensional data but require careful handling of calibration, initialization, data leakage and distribution shift.
Insurance model comparisons must evaluate the same estimand, loss basis, forecast horizon and information set. In the French motor third-party liability study, Noll, Salzmann and Wüthrich compared GLMs with tree-based and neural approaches and identified important interaction-capture differences in that case (see Noll (2020)). This is evidence about the studied data and task, not a universal ranking of algorithms. A well-specified GLM can outperform a complex model under small samples, weak signal or distribution shift.
Automated machine learning searches preprocessing, hyperparameters and model families; Dong and Quan's 2025 insurance study illustrates a tailored workflow and the relevance of imbalanced data (see Dong (2025)). Automation moves the location of model-selection decisions and therefore expands the need for nested validation, audit trails, human review, and restrictions on operational uses. It does not remove methodological responsibility.
A technically defensible comparison includes a simple actuarial baseline, regularized statistical baseline, and justified machine-learning alternatives. Report the number of tuneable choices, data split design, search budget, uncertainty in performance differences, subgroup results, and costs of deployment. Interpretability tools such as partial dependence or additive attribution estimate behavior of a fitted predictor; correlated variables and out-of-support perturbations can make their explanations misleading. Explanation should ultimately address both model mechanics and the insurance decision to which the model contributes.
Survival, Duration, and Multistate Processes
Life, health, disability, long-term-care and lapse models frequently predict time to a contingent event. With event time , survival is and hazard is when this limit exists. Kaplan–Meier estimation handles right-censored follow-up under appropriate independent-censoring assumptions. Cox's proportional-hazards model represents the hazard as (see Cox (1972)); time-varying effects or covariates may require extensions.
Insurance portfolios exhibit delayed entry, competing risks, recurrent events, multistate transitions, selective surrender, and changing benefits. A person may move from active to disabled, recovered, or dead, with transition intensities depending on age, duration and previous history. A competing event changes the interpretation of an observed cumulative incidence; using one minus a cause-specific Kaplan–Meier estimate as a competing-risk cumulative incidence can be inappropriate.
Statistical choices must follow the contract and forecasting question. A mortality table forecasts event risk conditional on survival and portfolio composition; a long-term-care model may forecast both incidence and state duration; lapse models may be entangled with economic incentives, sales practices and management actions. An annual event probability is not automatically a continuous-time hazard, and observed attrition need not represent random censoring.
Claims Development, Reserving Data, and Forecast Uncertainty
Claims data mature through reporting and settlement. An accident-period by development-period triangle summarizes payments or incurred amounts at successive valuations. Deterministic chain-ladder extrapolates historical age-to-age development patterns; stochastic variants describe prediction error under explicit assumptions. Mack's 1993 contribution provided a distribution-free standard-error calculation within its particular model conditions, without making chain-ladder universally robust to structural changes (see Mack (1993)).
Calendar-period changes in inflation, legal practice, settlement speed or claims administration can distort extrapolations based on stable development factors. Case reserve strengthening, changes in claims handling and reporting lags also affect incurred data. A payment triangle and an incurred triangle can convey different information about the same underlying claim population.
Individual-claim transaction models allow reporting delay, payment timing, reopened claims and claim-level covariates to be analyzed separately. They can yield richer forecasts where transaction histories are complete, yet introduce sensitivity to coding consistency, claims-maturity censoring and large-claim scarcity. For both aggregate and granular approaches, a reserve estimate should identify the valuation date, gross or net basis, claim scope, tail assumption, process uncertainty, parameter uncertainty and any scenario overlays (see Wüthrich (2008)).
The statistical object is the eventual claim amount or distribution conditional on the information available at valuation. Booking a financial reserve additionally involves accounting, regulatory and management conventions treated more fully under Insurance Pricing and Reserving and Insurance Accounting and Financial Reporting.
Dependence, Catastrophes, and Systemic Exposure
Independent policy-level models can be useful for attritional claims, but common events connect many risks. Hurricanes, earthquakes, pandemics, cyber outages, supply-chain disruption and inflation shocks can create both frequency clustering and correlated severity. Aggregate portfolio uncertainty depends on marginal losses and their joint distribution.
A copula expresses a multivariate distribution through marginals and a dependence function under the conditions of Sklar's theorem. Linear correlation alone cannot characterize all tail relationships. Common-shock models, spatial random fields, hierarchical catastrophe-event simulations, dynamic factor models and network models are alternative choices for specific dependence structures. Their scientific adequacy depends on event definition, geography, insured value distribution, vulnerability functions, claims adjustment and exposure concentration.
Observed catastrophe losses reflect changing exposure and coverage as well as physical hazards. Raw insured-loss trends should not be interpreted directly as changes in hazard intensity without normalization and exposure analysis. Scenario and stress testing are particularly relevant where the historical frequency of extreme common shocks is insufficient for precise extrapolation. The 2024 IAIS Insurance Core Principles' risk-management framework supplies international supervisory context for governance of material exposures.
Predictive calibration for a common-shock portfolio cannot be demonstrated by a random split that places highly correlated policies from the same catastrophe event in both training and testing. Event-blocked, spatially blocked, and forward-time evaluations better approximate prospective generalization. Tail metrics must be interpreted with their limited effective number of independent events.
Causal Inference, Selection, and Behavioral Feedback
Prediction concerns the distribution of outcomes conditional on available information; causal inference concerns outcomes under interventions. A feature may predict a claim yet be inappropriate for estimating the effect of prevention, underwriting action or premium change. Confounding, selection, measurement error and reverse causation can all create a stable predictive association without identifying an intervention effect.
Insurance decisions affect the future dataset. Pricing changes which risks buy or renew cover; claims triage changes investigation effort and hence recorded fraud labels; safety incentives change behavior; telematics participation can select drivers with particular risk profiles. Training only on accepted and retained business creates selective labels, and off-policy evaluation requires assumptions about unobserved outcomes. An underwriting model's observed accuracy on written policies cannot establish its accuracy for declined applicants.
Randomized interventions, natural experiments, difference-in-differences with defensible trends, instrumental variables under explicit assumptions, and causal graphical analysis may help answer intervention questions. Each method requires substantive knowledge and careful identification conditions. In observational insurance settings the estimand, population affected, potential policy-induced selection, and sensitivity to unmeasured confounding deserve independent documentation.
A practical synthesis follows: when the use case is predictive, optimize valid prospective prediction; when the question is intervention effectiveness, specify an identification strategy. Combining the two tasks without their assumptions creates an avoidable source of false certainty.
Feature Engineering and Nontraditional Data
Feature engineering translates contract, exposure and environmental information into stable, forecast-origin-measurable covariates. Standard examples include age and duration, claim history over a fixed lookback, exposure-adjusted prior frequency, geographic hazard variables, and characteristics known before the coverage period. The feature definition should include the time at which it becomes available and the degree to which it will remain available in production.
Telematics produces time-series information on vehicle use, driving conditions and observed behavior, but opt-in participation, device reliability and duration of observation can induce selection. A study summarized in ASTIN Bulletin describes calibration-weight integration of traditional and telematics data and explicitly considers favorable-selection problems (see Peiris (2024)). Feature usefulness must therefore be evaluated on the target insured population, not only on the observed device-user sample.
Computer vision and learned embeddings can convert images to predictors. Blier-Wong, Lamontagne and Marceau's 2024 study explored street-view imagery within an actuarial pricing framework and discusses technical and social problems with using such data. Image features may encode neighborhood proxies, transient conditions, privacy-sensitive details and geographic access disparities. Performance on a particular peril does not justify extrapolation to other perils or locations.
Textual claims narratives and document embeddings can support classification, retrieval and operational prioritization. These data contain protected information, inconsistent narratives and post-event information. Feature availability must be aligned to the prediction timestamp, with separate restrictions for tasks involving fraud allegations, claims settlement, and adverse consumer outcomes. Provenance, permitted use, retention and human accountability are part of the statistical specification.
Training Design, Validation, and Reproducibility
A rigorous prediction study separates model construction from independent performance assessment. When observations are temporally ordered, training should precede validation, and evaluation should reflect the reporting and development delay that would have applied prospectively. Random cross-validation can provide useful estimates in genuinely exchangeable data; repeated policies, households, claims or catastrophe events demand grouped or blocked validation.
A robust workflow freezes: target definition; population inclusion and exclusions; forecast origin; variable availability; exact feature-generation code; missing-value policy; training, tuning and independent test periods; model candidates; metrics; decision thresholds; and pre-specified subgroup diagnostics. Hyperparameters, feature selection, imputation and target encoding belong inside each training fold. Tuning directly on the final test set invalidates its claimed independence.
Out-of-time performance is indispensable where inflation, claim reporting, contract wording, portfolio mix and hazard distribution change. Rolling-origin backtests can expose stability or seasonality. Model performance uncertainty should be estimated at the correct dependence unit, often by policyholder or time/event block rather than independent transaction rows.
Reproducible artifacts preserve schema versions, source data snapshots or identifiers, code revision, random seeds, parameter estimates, calibration method, evaluation population, and signed review. The ability to rerun an algorithm is distinct from having permission to reuse its underlying data; controlled access and privacy-preserving audit mechanisms may be necessary.
The central warning is label maturity: a claim that has not been reported or settled at test extraction time may have a falsely recorded zero or understated cost. Validation must recreate the as-of knowledge state and either use adequately matured outcomes or explicitly account for right-censoring and reporting delay.
Model Evaluation, Calibration, and Proper Scores
Accuracy is multidimensional. For count rates, exposure-weighted Poisson deviance or log score assesses a particular predictive object. For positive severity, log likelihood, deviance and robust error summaries address different aspects. For a full predictive distribution, the logarithmic score, continuous ranked probability score and quantile scores offer principled comparisons; their strictly proper forms encourage truthful distributional forecasts under their assumptions (see Gneiting (2007)).
A model can rank risks correctly while overstating every probability. Discrimination metrics such as AUC assess ranking for a binary event, whereas calibration evaluates agreement between predicted and observed event rates. For binary outcomes, the Brier score is a proper score for a forecast probability . Aggregate calibration requires comparison with observed frequency over appropriately exposed groups, and conditional calibration checks whether systematic subgroup deviations remain.
Illustrative insurance calibration should compare with on a consistently defined outcome and exposure population, with temporal and segment diagnostics. A single balance at portfolio level can conceal large offsetting errors between territories, coverage classes or protected groups.
Metric selection must reflect downstream loss. Mean absolute error emphasizes absolute deviations; mean squared error magnifies large errors; quantile loss serves a specified percentile, and tail conditional metrics address extreme outcomes. Claims fraud classifiers with few positive labels need precision–recall and workload analysis in addition to ranking metrics. Economic value studies must disclose decision costs, class prevalence, interventions, and potential feedback.
A well-calibrated expected loss estimate is not itself a justified premium, reserve, capital charge or claim decision. Financial loadings, expenses, tax, market constraints, contract promises, fairness rules and governance are separate decision inputs. Reporting should preserve this boundary.
Uncertainty Quantification, Stress Tests, and Model Risk
Operational insurance models face process uncertainty, parameter estimation error, model-form uncertainty, data revisions, forecast-horizon drift, and common-shock uncertainty. Point predictions compress this structure. Predictive distributions, scenario ranges, bootstrap estimates, hierarchical posterior predictions and explicit sensitivities provide different views of the uncertainty, provided the assumptions are recorded.
Bootstrap resampling must respect dependence. Repeated policy records should be clustered by policy or insured where appropriate; catastrophe outcomes require event-level or temporal blocks. An ordinary row bootstrap can materially understate variance when rows share underlying shocks. Bayesian models require sensitivity to priors and assessment of convergence; frequentist intervals require attention to model specification and small-sample behavior.
Scenario analysis probes assumptions whose consequences cannot be learned accurately from available history: inflation changes, legal reforms, cyber accumulation, reporting delays, wildfire exposure shifts, or rare medical utilization shocks. Stress tests are conditional thought experiments; their outcomes are not empirical probabilities unless calibrated within a justified probabilistic model.
A usable model-risk register documents conceptual limitations, data dependencies, parameter sensitivity, performance deterioration, deployment changes, failure modes, review frequency and override governance. ASOP No. 56 provides a U.S.-specific professional example of responsibilities surrounding actuarial model work, with an effective date of October 1, 2020 (see Actuarial Standards Board (2019)). International IAIS standards frame risk management and controls at insurer level.
Fairness, Privacy, Explainability, and Insurance Governance
Insurance prediction participates in a market where private information, coverage access, risk differentiation and collective protection intersect. Statistical associations may capture prohibited factors, lawful risk characteristics, historic inequality, administrative bias or geospatial proxies. Fairness therefore requires a specified jurisdiction, product, legal basis and policy objective; numerical parity across groups is neither a universal legal requirement nor a complete fairness assessment.
Possible assessments include subgroup calibration, false-positive or denial error patterns, stability across regions and socioeconomic settings, proxy detection, restrictions on sensitive attributes, and analysis of availability and affordability. Fairness diagnostics may conflict because different mathematical criteria impose different constraints when base rates differ. Their use should reflect a transparent normative and legal choice, rather than an unexamined algorithmic default. The research literature now explicitly treats prediction, explanation and fairness as joint actuarial concerns (see Aas (2024)).
Explainability distinguishes global patterns, individual attributions, causal reasons and procedural disclosure. A transparent GLM can still be misleading when features are collinear, proxy-sensitive or based on selective claims labels. A complex ensemble can be audited, but a feature-attribution plot is not proof of causation. Human review should address decision consequences and the possibility of contestation.
The NAIC adopted its Model Bulletin on insurers' use of AI on December 4, 2023; a model bulletin guides state-level approaches and is not automatically law across all U.S. jurisdictions (see NAIC insurance guidance). The NAIC's subsequent 2025–2026 evaluation-tool work illustrates ongoing supervisory development, not a completed universal regime (see NAIC insurance guidance). The IAIS issued an AI supervision application paper on July 2, 2025, organized around proportionality, governance, robustness, transparency and fairness. EIOPA's August 6, 2025 Opinion concerns how existing European insurance-sector requirements apply to AI systems within its stated scope; EIOPA explicitly distinguishes this Opinion from high-risk AI systems under the EU AI Act (see EIOPA guidance). These sources must not be collapsed into a single global rule.
Privacy measures include data minimization, purpose limitation, access restriction, deletion controls, secure linkage, independent review of sensitive features, and protections against memorization or re-identification where applicable. Compliance is context-specific. Statistical validation, explanation, human accountability and consumer redress are complementary governance functions.
International Comparison and Institutional Variation
Insurance data practices vary across jurisdictions because products, legal liability rules, health systems, claims handling, availability of external records, mandatory schemes and data-protection regimes differ. A motor claim in one country may involve a different liability structure and claims development profile from a nominally similar claim elsewhere. Pooling international observations requires an explicit transferability argument, harmonized measurement, currency and inflation adjustments, and analysis of legal and institutional differences.
In the United States, CAS methodological monographs provide practical actuarial guidance, ASOPs 23 and 56 govern relevant work for actuaries subject to those professional standards, and NAIC work influences state insurance supervision according to jurisdictional adoption (see Goldburd (2025); Chalk (2026)). Within the European Union, EIOPA's 2025 Opinion articulates supervisory expectations regarding AI governance under existing insurance-sector law within its defined scope (see EIOPA guidance). At international level, IAIS Insurance Core Principles and its AI application paper establish supervisory reference frameworks rather than a uniform national statute.
These institutional documents show converging concerns—data quality, governance, validation, consumer treatment and robustness—alongside distinct legal effects. A cross-country prediction study should therefore identify observed jurisdictions, local definition of claim and exposure, legal treatment of variables, policy period, data quality, and the conditions of model transfer. Formal revalidation is warranted before changing target population, product, reporting system or law.
Illustrative Analytical Derivations and Checks
The following examples are mathematical illustrations, not insurer experience or estimates from proprietary records.
For three policy observations with claim counts and earned exposures , total count is and total exposure is risk-years. The exposure-based homogeneous Poisson rate estimate is claim per risk-year. The unweighted mean of policy-specific rates would be , a different estimand that overweights the shortest exposure. The example demonstrates why an exposure denominator and weighting convention are part of the statistical question.
For a compound Poisson model with expected count and severity taking values and equally likely, severity mean is , severity variance is , and second moment is . Thus and . Using as aggregate variance would omit count fluctuation. If an insurer adds a payment limit or common-event dependence, both moments must be recomputed for that altered definition.
For credibility and , . An observed risk mean of and collective mean of yield a pooled mean of , given the simplified Bühlmann assumptions. The number is a model-based estimator; it supplies neither a risk margin nor a contractual premium.
For a binary event with forecast , observed outcome gives Brier loss ; an observed gives . Averaging over independent event realizations drawn from an event probability yields . Forecasts that differ from the true probability have a higher expected Brier score. This illustrates proper scoring, not actuarial pricing.
These computations were independently re-evaluated in the editorial test routine associated with the source package. They depend only on arithmetic and the stated hypothetical assumptions.
Synthesis: What Reliable Insurance Prediction Requires
The established literature supports a unified account of reliable insurance prediction across data regimes. The central object is a conditional future loss or event distribution measured on a contractual and temporal basis. Predictive modeling succeeds when the estimand, observation mechanism, evidence, functional assumptions, decision use and institutional constraints remain aligned.
Classical statistical tools contribute explicit likelihoods, interpretable rates, partial pooling and coherent estimation. Modern learning tools contribute interaction discovery, nonlinear approximation and the ability to integrate complex data. Neither category settles a problem whose labels are immature, exposures misdefined, censored losses untreated, or population selection unexamined. The selection between tools is an empirical and scientific judgment grounded in the same out-of-sample target.
A second synthesis concerns the direction of uncertainty. Insurers observe developed histories but must act on future risks whose risk factors, climate, law, technology and behavior may change. Retrospective fit cannot establish prospective validity where the data-generating system changes. Stability, forward testing, stress scenarios and reasoned governance must therefore accompany the model's numerical output.
A third synthesis concerns the relationship between prediction and action. More accurate prediction may improve risk understanding, yet insurance is a contract, a social institution, and a supervised financial activity. Decision procedures mediate the transformation of forecasts into coverage, price, claims handling, reserves or capital. Statistical validity is necessary for quantitative confidence in those procedures; it does not alone determine their legitimacy, welfare effect or financial prudence.
These are analytical conclusions derived from existing research. They are offered as an integrated assessment of the field and not as a claim of a newly established theory, doctrine or priority.
Open Research Problems and Future Inquiry
Important unresolved problems remain. Distribution shift is persistent in climate-sensitive catastrophe risk, cyber accumulation, healthcare utilization and inflation-driven claims costs. Reliable forward prediction requires detecting new regimes without falsely treating an observed recent period as a stable model of the next. Methodologically promising directions include online recalibration, interpretable time-varying hazard models, robust hierarchical pooling and scenario-conditioned forecasting, with careful external validation.
Claims censoring and selection challenge standard machine-learning benchmarks. The field needs evaluation protocols that jointly recognize immature claims, reporting delay, selective underwriting, fraud-investigation feedback and policyholder churn. Novel labels and external data must be assessed for availability and lawful use at the actual decision point.
Rare-event prediction remains difficult because a portfolio may contain millions of policies but only a small number of effectively independent catastrophic events. Better calibration, defensible tails and explicit uncertainty in dependence require collaboration between physical science, actuarial statistics, engineering and economics.
The trade-offs among predictive accuracy, explanation, risk classification, consumer access and fairness continue to demand transparent, jurisdiction-specific analysis. Current research on images, telematics, AutoML, embeddings and algorithmic governance demonstrates growing methodological diversity (see Blier-Wong (2024); Aas (2024)). No single class of models has established universal superiority across insurance products, data regimes and policy contexts.
Future evidence will be strongest when studies disclose target populations and information timestamps, use prospective and grouped tests, report negative as well as positive results, compare against credible baselines, release reproducible code when permissible, and separate model performance from social and regulatory claims.
Conclusion
Insurance statistics and predictive modeling is a mature, rapidly developing science of inference about contingent insured events and their financial consequences. Its core quantities are frequencies, severities, aggregate losses, durations, dependencies and evolving claim obligations. Its methods span descriptive statistics, credibility, generalized regression, survival analysis, stochastic claims development, hierarchical modeling, ensembles and neural learning. Their suitability depends on observable data, contract definitions, mathematical assumptions and prospective use.
The decisive scientific result of the synthesis is that predictive reliability is an end-to-end property. Defined exposures, correct temporal cutoffs, treatment of censoring and selection, defensible distributional assumptions, honest validation, uncertainty quantification and accountable deployment jointly establish the evidentiary meaning of a forecast. A statistically sophisticated model cannot compensate for an incoherent target or an inaccessible future label.
Internationally, the professional and supervisory literature increasingly places robust modeling beside governance, fairness, explanation and data integrity. Those expectations operate through particular jurisdictions and institutions. The science advances by testing claims against appropriate evidence and preserving the distinction between observation, inference, intervention and financial decision. This monograph establishes an integrated scholarly map of that discipline for the Insurance by Bogdanova encyclopedia.
Research Dossiers
183 approved research subjects are organized into 7 thematic families. Individual titles become links when their complete scholarly monographs are published and verified.
Foundations, History, and Concepts7
- Actuarial Credibility Theory and Experience Rating
- Extreme Value Theory for Insurance Losses
- Fire Insurance Archives and the History of Property Risk
- Mortality Tables and the Origins of Actuarial Valuation
- On-Level Premium Adjustments and Rate History Reconstruction
- Sources and Methods in Insurance History
- The History of Insurance Technology and Digitalization
Law, Regulation, and Governance6
- Actuarial Model Governance and Validation
- Claims Anomaly Detection and Fraud Prediction
- Commercial Insurance Exposure Schedule and Asset Data Governance
- Modeled Loss Triggers and Calculation Agent Governance
- Solvency Capital Disclosures and Regulatory Data Quality
- Traffic Enforcement and Insured Crash Frequency
Risk Science, Data, and Modeling64
- Actuarial Capital Modeling and Aggregation
- Actuarial Catastrophe Loss Modeling
- Actuarial Climate Scenario Analysis
- Actuarial Communication and Disclosure
- Actuarial Data Quality and Provenance
- Actuarial Experience Studies and Assumption Setting
- Actuarial Rate Indications and Profit Provisions
- Actuarial Reinsurance Pricing and Experience Analysis
- Actuarial Statistical Inference and Parameter Uncertainty
- Aggregate Loss and Compound Risk Models
- Bayesian Risk Assessment and Decision Updating
- Catastrophe Bond Risk Modeling and Expected Loss Calibration
- Causal Inference versus Predictive Association in Insurance
- Claims Reserving Tail Factor Estimation
- Climate Nonstationarity and Insurance-Linked Securities Tail Risk
- Climate Risk Uncertainty and Nonstationarity
- Cohort-Level Life Insurance Exposure and Credibility Adjustment
- Compound and Cascading Hazard Risks
- Cox Regression and Duration Prediction in Insurance
- Cross-Method Reserve Triangulation and Actuarial Judgment
- Cyber Loss Modeling for Actuarial Applications
- Decision Trees and Random Forests in Insurance Prediction
- Distribution Shift and Transfer Learning in Insurance Prediction
- Economic Scenario Generators in Insurance
- Explainability and Fairness in Actuarial Models
- Explainability Methods for Insurance Predictive Models
- Exposure Measurement and Actuarial Risk Units
- Generalized Additive Models in Insurance Prediction
- Generalized Linear Models in Actuarial Pricing
- Gradient Boosting for Insurance Loss Prediction
- Health Medical Claim Frequency and Severity Modeling
- Health Risk Adjustment Models and Calibration
- Incurred but Not Reported Claims Estimation
- Industry Loss Index Triggers in Catastrophe Bonds
- Insurance Data, Estimation, and Model Risk
- Insurance Prediction with Images and Remote Sensing
- Insurance Predictive Modeling Reproducibility and Open Benchmarks
- Insurance Predictive Training Design and Validation Splits
- Insurance Statistical Estimands and Prediction Targets
- Loss Trend, Inflation, and Exposure Adjustment
- Model Drift Detection and Forecast Backtesting in Insurance
- Model Selection and Predictive Benchmarking in Insurance
- Morbidity Classification and Predictive Health Risk Scoring
- Motor Insurance Claim Frequency and Severity Modeling
- Motor Vehicle Years Mileage and Exposure Measurement
- Neural Networks for Insurance Risk Prediction
- Predictive Uncertainty Quantification and Insurance Prediction Intervals
- Probabilistic Calibration of Insurance Forecasts
- Proper Scoring Rules for Insurance Prediction
- Property Insurance Building-Level Damage Estimation from Remote Sensing
- Property Insurance Catastrophe Model Vulnerability Curves
- Property Insurance Geocoding and Exposure Data Quality
- Prospective Loss Cost Estimation and Pure Premium Ratemaking
- Risk Model Sensitivity and Robustness
- Road Traffic Crash Exposure and Insurance Risk Denominators
- Selection Bias and Exposure Representativeness in Insurance Data
- Simulation and Numerical Risk Evaluation
- Solvency Capital Calibration and Tail Probability Assumptions
- Spatial Statistics and Geographic Risk Prediction for Insurance
- Stochastic Claims Reserving and Predictive Distributions
- Stochastic Mortality Forecasting
- Structural Estimation of Insurance Demand and Costs
- Survival Analysis and Kaplan-Meier Estimation in Insurance
- Telematics and Usage-Based Insurance Prediction
Finance, Capital, and Markets23
- Algorithmic Price Discrimination in Insurance Markets
- Board-Level Insurance Risk Reporting and Decision Dashboards
- Comparative Insurance Statistics and Market Indicators
- Cross-Jurisdictional Insurance Capital Comparability
- Cyber Insurance Rate Adequacy under Data Scarcity
- Empirical Tests of Selection in Insurance Markets
- Fairness in Insurance Classification and Pricing
- Health Insurance Rate Development and Medical Trend
- Insurance Data Advantages and Market Power
- Insurance Data Generating and Reporting Mechanisms
- Insurance Rate Monitoring and Pricing Model Validation
- Insurance Risk Data Architecture and Reporting Lineage
- Insurance-Linked Securities Model Risk and Disclosure
- Loss Ratio Ratemaking and Indicated Rate Changes
- Nonstationary Risk and Forward-Looking Insurance Pricing
- Portfolio Mix Shifts and Prospective Rate Adequacy
- Predicting Claim Reporting and Settlement Delays
- Price Competition and Market Power in Insurance
- Pricing of New Coverages and Sparse Insurance Experience
- Reinsurance Treaty Experience Data and Bordereaux
- Reserve Diagnostics and Backtesting
- Solvency II Standard Formula and Modular Risk Charges
- Welfare Identification in Insurance Markets
Insurance Products, Coverage, and Claims58
- Algorithmic Fairness Measurement in Insurance Predictions
- Artificial Intelligence in Insurance Decision-Making
- Automated Machine Learning for Insurance
- Bornhuetter-Ferguson Claims Reserving
- Calendar-Year Effects and Inflation in Claims Reserving
- Chronic Disease and Multimorbidity Risk in Health Insurance
- Claims Development Triangles and Chain-Ladder Reserving
- Coverage Relativities and Classification Rating Plans
- Deep Representation Learning in Insurance Analytics
- Disability Insurance Claim Incidence Trend and Rating Credibility
- Disability Insurance Recovery Relapse and Claim Termination Models
- Distributional Evaluation of Insurance Policies
- Enterprise Insurance Loss Runs and Claims Experience Reconciliation
- Enterprise Model Risk Inventory and Validation Oversight in Insurance
- Enterprise Risk Management and Insurance Firm Value
- Exploratory Data Analysis and Statistical Standardization in Insurance
- Field Experiments of Insurance Choice Architecture
- Health Claims Cost and Utilization Models
- Hierarchical and Mixed-Effects Models for Insurance Losses
- High-Cardinality Feature Engineering in Insurance
- Individual Claim Reserving and Micro-Level Loss Development
- Insurance Data Linkage and Entity Resolution
- Insurance Demand Elasticity and Coverage Choice
- Insurance Demand Experiments and Elicitation Methods
- Insurance Development and Economic Growth
- Insurance Model Ensembling and Stacking
- Insurance Operational Risk Event and Loss Data Collection
- Insurance Risk Communication and Board Challenge Effectiveness
- Insurance Risk Expert Judgment and Qualitative Assessment
- Large-Loss, Latent Claims, and Long-Tail Reserving
- Latent Consideration Sets in Insurance Choices
- Liability Claim Frequency and Severity by Accident Year
- Life Insurance Cause-of-Death and Excess Mortality Analysis
- Long-Term Care and Disability Insurance Experience Data Harmonization
- Missing Data, Truncation, and Censoring in Insurance Analytics
- Mortality Data Provenance and Coding Quality in Life Insurance
- Mortality Trend Validation and Backtesting in Life Insurance
- Multistate Models for Life and Health Insurance
- Natural Experiments in Insurance Economics
- Paid versus Incurred Claims Development
- Penalized Regression and Variable Selection in Insurance
- Privacy-Preserving Statistical Learning for Insurance
- Property Insurance Frequency and Severity Decomposition
- Property Insurance Model Uncertainty and Validation Across Perils
- Quantile and Distributional Regression for Insurance Losses
- Rare Event Learning and Class Imbalance in Insurance
- Risk Numeracy and Probability Comprehension in Insurance
- Scientific Methods in Insurance Research
- Stated and Revealed Insurance Preferences
- Statistical Cleaning and Reconciliation of Insurance Claims Data
- Statistical Leakage and Proxy Contamination in Insurance Models
- Stochastic Claim Processes
- Structural Identification of Behavioral Insurance Preferences
- Subjective Risk Beliefs in Insurance Decisions
- Tweedie Compound Poisson Models for Insurance Costs
- Unsupervised Risk Segmentation and Clustering in Insurance
- Welfare Analysis of Insurance Decision Frictions
- Zero-Inflated and Hurdle Models for Insurance Claims
Institutions, Operations, and Technology1
Specialized and Interdisciplinary Research24
- Age-Period-Cohort Models for Insured Mortality
- Aleatory and Epistemic Uncertainty
- Annuitant Mortality Experience Data Provenance
- Cape Cod and Expected Loss Ratio Reserving
- Dependence Structures and Copula Models
- Disability Incidence and Recovery Models
- Empirical Measurement of Insurer ERM Maturity
- Gompertz-Makeham Mortality Models
- Insured-Life Mortality Experience and Selection Effects
- Liability Loss Triangle Calendar-Year Distortion
- Life Tables and Survival Functions
- Long-Term Care Entry Incidence by Functional State
- Longevity Model Backtesting for Annuity Liabilities
- Loss Frequency Models
- Loss Severity and Heavy-Tailed Distributions
- Mortality Heterogeneity and Frailty in Life Portfolios
- Multiple Decrement and Competing Risks Models
- Network Models of Risk Propagation
- Nonstationarity and Time-Varying Risk
- Objective and Subjective Interpretations of Probability
- Probability and Risk under Limited Information
- Risk Classification and Underwriting
- Systematic Retirement Drawdowns and Rule Backtesting
- The Economics of Risk Information Acquisition
References
- Nelder, J. A., and R. W. M. Wedderburn. 1972. “Generalized Linear Models.” Journal of the Royal Statistical Society, Series A 135 (3): 370–384. DOI: 10.2307/2344614. https://academic.oup.com/jrsssa/article/135/3/370/7110572
- Bühlmann, H. 1967. “Experience Rating and Credibility.” ASTIN Bulletin 4 (3): 199–207. Casualty Actuarial Society archival record. https://www.casact.org/abstract/experience-rating-and-credibility
- Bühlmann, H. 1969. “Experience Rating and Credibility.” ASTIN Bulletin 5 (2): 157–165. DOI: 10.1017/S0515036100008023. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/experience-rating-and-credibility/0EC60711D65BFF225845CBA50C4B04D4
- Mack, T. 1993. “Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates.” ASTIN Bulletin 23 (2): 213–225. DOI: 10.2143/AST.23.2.2005092. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/distributionfree-calculation-of-the-standard-error-of-chain-ladder-reserve-estimates/E8D207F9A4DCE30300A76780FE510437
- Jørgensen, B., and M. C. Paes de Souza. 1994. “Fitting Tweedie's Compound Poisson Model to Insurance Claims Data.” Scandinavian Actuarial Journal 1994 (1): 69–93. DOI: 10.1080/03461238.1994.10413930. https://www.tandfonline.com/doi/abs/10.1080/03461238.1994.10413930
- Smyth, G. K., and B. Jørgensen. “Fitting Tweedie's Compound Poisson Model to Insurance Claims Data: Dispersion Modelling.” ASTIN Bulletin. Original publisher record. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/fitting-tweedies-compound-poisson-model-to-insurance-claims-data-dispersion-modelling/DEF0B49F96FC015C7FBE076BC0A5C3AC
- Kaplan, E. L., and P. Meier. 1958. “Nonparametric Estimation from Incomplete Observations.” Journal of the American Statistical Association 53 (282): 457–481. DOI: 10.1080/01621459.1958.10501452. https://www.tandfonline.com/doi/full/10.1080/01621459.1958.10501452
- Cox, D. R. 1972. “Regression Models and Life-Tables.” Journal of the Royal Statistical Society, Series B 34 (2): 187–202. DOI: 10.1111/j.2517-6161.1972.tb00899.x. https://rss.onlinelibrary.wiley.com/doi/10.1111/j.2517-6161.1972.tb00899.x
- Breiman, L. 2001. “Statistical Modeling: The Two Cultures.” Statistical Science 16 (3): 199–231, including discussion and rejoinder. DOI: 10.1214/ss/1009213726. https://www.scienceopen.com/document?vid=d7090733-eed9-4040-ad16-82730f65c9d3
- Breiman, L. 2001. “Random Forests.” Machine Learning 45: 5–32. DOI: 10.1023/A:1010933404324. https://doi.org/10.1023/a%3A1010933404324
- Friedman, J. H. 2001. “Greedy Function Approximation: A Gradient Boosting Machine.” Annals of Statistics 29 (5): 1189–1232. DOI: 10.1214/aos/1013203451. https://www.jstor.org/stable/2699986
- Gneiting, T., and A. E. Raftery. 2007. “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102 (477): 359–378. DOI: 10.1198/016214506000001437. https://sites.stat.washington.edu/people/raftery/Research/PDF/Gneiting2007jasa.pdf
- Wüthrich, M. V., and M. Merz. 2008. Stochastic Claims Reserving Methods in Insurance. Chichester: John Wiley & Sons. Print ISBN 978-0-470-72346-3. https://www.wiley.com/en-gb/stochastic-claims-reserving-methods-in-insurance-p-9780470723463
- Hastie, T., R. Tibshirani, and J. Friedman. 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. New York: Springer. DOI: 10.1007/978-0-387-84858-7. https://doi.org/10.1007/978-0-387-84858-7
- Noll, A., R. Salzmann, and M. V. Wüthrich. 2020 revision. “Case Study: French Motor Third-Party Liability Claims.” SSRN working paper. DOI: 10.2139/ssrn.3164764. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3164764
- Goldburd, M., A. Khare, D. Tevet, and D. Guller. 2025 revised second edition. Generalized Linear Models for Insurance Rating. CAS Monograph 5. Casualty Actuarial Society. ISBN 978-1-7333294-4-6. https://www.casact.org/monograph/cas-monograph-no-5
- Chalk, A., D. Deacon, M. Guillén, and M. Martinelli. 2026 release. From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges. CAS Monograph 16. Casualty Actuarial Society. ISBN 978-1-7370028-9-5. https://www.casact.org/publications-research/publications/flagship-publications/cas-monographs/monograph-no-16
- Blier-Wong, C., L. Lamontagne, and E. Marceau. 2024. “A Representation-Learning Approach for Insurance Pricing with Images.” ASTIN Bulletin 54 (2): 280–309. DOI: 10.1017/asb.2024.9. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/representationlearning-approach-for-insurance-pricing-with-images/E442676A2F597A9DBBDFC9FB986C3070
- Aas, K., A. Charpentier, F. Huang, and R. Richman. 2024. “Insurance Analytics: Prediction, Explainability, and Fairness.” Annals of Actuarial Science 18 (3): 535–539. DOI: 10.1017/S1748499524000289. Editorial. https://www.cambridge.org/core/journals/annals-of-actuarial-science/article/insurance-analytics-prediction-explainability-and-fairness/CF5EDDEA60798783CEFD5FBE6A51A8FF
- Dong, P., and Z. Quan. 2025. “Automated Machine Learning in Insurance.” Insurance: Mathematics and Economics 120: 17–41. DOI: 10.1016/j.insmatheco.2024.10.002. https://www.sciencedirect.com/science/article/pii/S0167668724001057
- International Association of Insurance Supervisors. 2025. Application Paper on the Supervision of Artificial Intelligence, release announcement July 2, 2025. https://www.iais.org/2025/07/the-iais-publishes-application-paper-on-the-supervision-of-artificial-intelligence/
- International Association of Insurance Supervisors. Insurance Core Principles and ComFrame, current online framework; amendments adopted December 5, 2024. https://www.iais.org/activities-topics/standard-setting/icps-and-comframe/
- European Insurance and Occupational Pensions Authority. 2025. Opinion on Artificial Intelligence Governance and Risk Management. August 6, 2025. https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en
- National Association of Insurance Commissioners. 2023. Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers
- Actuarial Standards Board. 2016 revision, effective 2017. Actuarial Standard of Practice No. 23: Data Quality. https://www.actuarialstandardsboard.org/asops/data-quality/
- Actuarial Standards Board. 2019 adoption, effective 2020. Actuarial Standard of Practice No. 56: Modeling. https://www.actuarialstandardsboard.org/asops/modeling-3/
- National Association of Insurance Commissioners. 2026 current topic page. “Artificial Intelligence.” https://content.naic.org/insurance-topics/artificial-intelligence
- Peiris, H., H. Jeong, J.-K. Kim, and H. Lee. 2024. “Integration of Traditional and Telematics Data for Efficient Insurance Claims Prediction.” ASTIN Bulletin 54 (2): 263–279. Cambridge University Press journal issue description. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/issue/546FBEBDDA8027D4EE2F210453ECE09E
- Casualty Actuarial Society. 2026. “New Exam 8 CAS Monograph Available.” February 25, 2026. https://www.casact.org/article/new-exam-8-cas-monograph-available