RESEARCH DOSSIER
Insurance Data, Estimation, and Model Risk
Abstract. Insurance data are produced by policies, exposure histories, behavior, incidents, reporting incentives, contractual adjudication, settlement, reinsurance, and administrative systems. Statistical estimation performed on such records is consequently an inference about a specified insurance population and financial obligation through a selective, delayed and changing observation process. This monograph examines how representativeness, exposure denominators, missingness, censoring, reporting delays, heterogeneity, structural change, dependency and model use determine the credibility of estimates. It develops a critical synthesis of established statistical inference, actuarial reserving, predictive validation and insurance governance, with explicit mathematical illustrations linking data errors to expected claims, estimated liabilities and the distribution of capital-sensitive outcomes. International supervisory expectations are compared with particular attention to the different legal status of actuarial professional standards, Solvency II requirements, IAIS supervisory principles and U.S. state guidance. The central finding is that the scientific quality of an insurance estimate is conditional on the alignment of its estimand, observation mechanism, identification assumptions, validation population and intended decision. The cost of model error extends beyond numerical forecast loss to underwriting selection, reserve emergence, solvency, unfair consumer outcomes and false confidence in organizational decisions. The study offers an original comparative analytical synthesis and derived examples; it introduces no proprietary data, new actuarial theory or externally reviewed empirical results.
Keywords: Insurance data; exposure measurement; statistical estimation; measurement error; selection bias; missing data; censoring; claims reporting; estimation uncertainty; parameter risk; structural model risk; distribution shift; insurance pricing; claims reserving; solvency; model validation; backtesting; calibration; stress testing; actuarial governance; data lineage; model risk management.
Introduction: Research Object, Sources, and Method
An insurance estimate is a statement about a defined future or outstanding contingent financial quantity, conditioned on information available at a specified time. The data supporting that statement reflect an institutional and contractual process: a person or asset becomes covered; an exposure period begins and ends; an incident may occur; notification, coverage determination and adjustment follow; payments may be deferred or reopened. Each event changes the observable record and may change the measure that financial reporting or risk governance is intended to estimate.
The subject of this dossier is epistemic and institutional: how the investigator can know that an insurance estimate corresponds to the population, cash flows and decisions for which it is used, and how to manage the residual possibility of error. Detailed construction of generalized linear models, machine-learning architectures and individual actuarial algorithms belongs principally to the related Science on Insurance Statistics and Predictive Modeling and Actuarial Science. Here algorithms serve as examples used to examine observation quality, assumptions, inferential limits and consequential model use.
Research proceeds through critical comparison of original scholarly papers and professional or supervisory materials, explicit specification of statistical estimands, analytical counterexamples, numerical sensitivity calculations and a jurisdiction-scoped assessment of model governance. The academic literature establishes general inferential results; insurance standards establish professional or supervisory expectations within their stated remit. Illustrations are hypothetical, computed from declared assumptions and never represented as observations of an insurer or the result of a backtest on real data. Verification has been performed on accessible primary institutional texts and authoritative article abstracts or records; restricted full texts and proprietary data have not been replicated.
An essential distinction is the difference between the true insured-event process, the contracted financial obligation, the recorded claim process and the modeled statistical distribution. A model can fit recorded claims closely while giving misleading estimates of ultimate covered loss if case reserves are incomplete, exposure units drift or the sample excludes risks that would be written under a contemplated policy. The actuarial data quality standard (ASB, 2016) and the IAIS Insurance Core Principles express the professional and institutional relevance of reliable data and controlled use, while Rubin's missing-data analysis (1976) and Heckman's selection analysis (1979) explain why completeness within the observed sample does not by itself secure valid population inference.
Insurance Data and Exposure Definitions
The insurance record as a generated observation
The insurance record is jointly determined by policy provisions and events in the underlying world. Coverage exclusions, attachment points, deductibles, waiting periods and limits transform a physical or financial incident into a contractual claim. The same incident can therefore generate a paid claim under one policy and no paid claim under another. Counts of notifications, reported claims, open claims, paid claims and ultimate incurred claims are distinct random variables. Analysts must specify which is the subject of estimation, the cutoff date at which it is observed, and whether changes in definitions have been applied consistently.
A useful decomposition of an observed paid outcome is a composition of exposure, event occurrence, notification, coverage recognition, settlement and payment. This decomposition is conceptual rather than a universal probabilistic factorization: certain stages interact and may be revised retrospectively. Its practical benefit is diagnostic. If the observed trend changes, the investigator can ask whether claim frequency changed, whether notification became quicker, whether deductibles were altered or whether reserve-setting practice shifted. Treating all movements as changes in underlying insured hazard confounds claims operations with risk.
A reproducible data record should preserve policy identity, insured unit, coverage form, effective dates, limit and retention terms, valuation date, event and report dates, monetary components, currency, adjustment reason, transaction provenance and the transformation history that produces the modeling table. The Solvency II documentation requirements for technical provisions expressly address data sources, characteristics and uses (European Commission, 2015). For U.S. actuarial practice, ASOP 23 on data quality (ASB, 2016) guides selection, review and reliance; neither substitutes for the empirical work of reconciling data to the actual contractual estimand.
Exposure is a denominator with legal and temporal meaning
Exposure is the measure of insured risk at risk of generating the specified event within a defined observation period. Policy-years, insured vehicle-years, payroll, sums insured, lives and insured property values are different denominators. Each carries assumptions about time at risk and intensity. Counting a half-year contract as a full policy-year can halve an estimated annual frequency if applied systematically to the denominator; counting a policy with an interruption in cover as continuously active can similarly dilute the observed rate.
For a homogeneous Poisson count model, an observation with exposure e and hazard intensity λ has E[N|e] = eλ. The elementary rate estimator is λ̂ = (Σ Nᵢ)/(Σ eᵢ), provided all Nᵢ and eᵢ refer to the same covered event definition and compatible period. This is a ratio of totals, not an average of policy-level ratios. When exposure is badly recorded, even perfectly measured claim counts yield an incorrect incidence estimate. When risks are heterogeneous, the ratio estimates an exposure-weighted aggregate rate for the observed mix, not automatically a rate for every future portfolio.
Consider 100 policy-years and five reported covered incidents: the observed frequency is 0.05 events per policy-year. If the true exposure was 80 policy-years because 20 policies were half-year contracts plus other adjustments, the corresponding frequency is 0.0625. The apparent precision of the integer numerator offers no information about this 25% upward correction. Exposure reconciliation precedes confidence interval calculation because a conventional interval conditional on an incorrect denominator quantifies the wrong uncertainty.
Units, linkage and aggregation
The appropriate statistical unit may be a policy-term, insured object, person, claim, occurrence, development cell, catastrophe event or legal entity. Aggregating across units can destroy information on dependence and produce Simpson-type reversals when the composition of the insured portfolio changes. If risks with high expected costs become less common, total average cost may fall despite worsening loss rates within every risk class. A model trained on pooled observations may then reward changes in portfolio composition rather than detect changes in class-specific claims experience.
Identifiers also generate error. One incident can produce multiple claim files, while one file can combine multiple coverages or multiple payments. Entity resolution requires explicit rules governing merged policies, reassigned claim numbers, subrogation, recoveries and catastrophe-event identifiers. The relevant control is reconciliation of the complete accounting or claims ledger to the analytical table, accompanied by an audit of unmatched, duplicated, late-arriving and excluded records.
Risk-specific estimates also depend on whether the analytic purpose is gross loss, net loss after reinsurance, incurred loss, ultimate loss or cash paid during a specified time horizon. ASOP 43 on unpaid claims (ASB, 2011 revision) requires the purpose, measure, accounting date and valuation information to be identified for the applicable actuarial assignment. An observation table built for gross incurred losses cannot be transferred unchanged into a net-of-reinsurance liquidity forecast.
Data quality as fitness for the estimand
Completeness is only one data-quality dimension. Others are accuracy, timeliness, consistency across systems and periods, legal provenance, representativeness, interpretability, lineage, version control and the materiality of errors to the intended decision. A variable can be accurate yet unsuitable: the last paid amount is accurate as a payment record but unsuitable as a proxy for ultimate claim cost for young accident years. A demographic feature may predict prior claim outcomes but reflect historical access, underwriting practices or claims-handling rules that are changing.
Quality checks should be linked to possible loss mechanisms. Date order checks protect timing models; limits and attachment checks protect severity estimates; duplicate incident checks protect frequency; signed currency reconciliation protects reported totals; and coverage-change flags protect longitudinal comparability. Data that fail one check need not be discarded indiscriminately. Their inclusion, exclusion or imputation must be justified against the target, with uncertainty propagated rather than silently transferred to a model coefficient. The Solvency II Article 82 makes appropriateness, completeness and accuracy explicit for technical-provision data within its EU legal scope.
Lineage and reproducibility
Insurance datasets are frequently snapshots. Later corrections to a claim's cause, liability, payment or reopened status make the 'same' record differ between downloads. Reproducibility requires an as-of cutoff, immutable version identifiers for each input, transformation logic, contract-version mappings and a record of manually overridden entries. A retrained model using updated claims labels without a frozen as-of view cannot reliably be compared with a prior model's historical performance: both the model and the truth against which it is judged may have changed.
A lineage ledger should identify the data owner, extraction process, independent reconciliation, transformation version, feature cutoffs, exception treatment, training population, validation population and downstream estimates affected. Such a ledger allows specific analytical questions: Was the reserve release caused by new experience, a different development assumption or a revision to data? Governance becomes scientifically substantive when those questions can be answered with evidence rather than institutional recollection.
Statistical Estimation and Sampling
Estimands, estimators and inferential populations
An estimand is the precise population quantity sought: the expected annual cost of the next policy cohort, the ultimate unpaid liability of past accidents, a one-year tail loss measure or a specified conditional risk probability. An estimator is a rule mapping observations into an estimate of that quantity. Model fit addresses how closely observations resemble estimated conditional values; identifiability addresses whether different underlying realities could generate the same observables. These are separate concerns. No reduction in standard errors can rescue a target quantity that is unavailable from the available variables without further assumptions.
Let Y represent covered loss and X recorded underwriting variables. Estimating E[Y|X=x, enrolled in historical book] generally differs from E[Y|X=x, eligible for a future offering] when enrollment is related to unobserved risk. A target-population shift also changes E[Y] even if each conditional E[Y|X] remains stable. The estimate therefore needs its population definition, data calendar, insurance cover conditions and any transfer assumptions reported together.
Point estimates can be unbiased under a mathematical model and still be unsuitable for a decision. Mean squared error decomposes into variance and squared bias relative to the stated estimand, but that decomposition is conditional on a reference data-generating process. It does not represent all risks of contractual misinterpretation or unmodeled structural events. Reporting a narrow confidence interval only describes uncertainty within the maintained inferential assumptions.
Sampling, strata and unequal selection
Representative insurance sampling requires the sample to reproduce, through design or suitable adjustment, the units and risk distribution relevant to the target. Convenience extracts from closed claims or retained customers are often selection mechanisms, not probability samples. Stratification can improve precision by ensuring coverage of rare high-risk segments; it does not guarantee external validity if some segments have zero inclusion probability. Survey-style inverse selection weighting may restore a population mean if inclusion probabilities are estimable and positivity holds.
For units with inclusion probability πᵢ > 0, a Horvitz–Thompson total takes the form Σ(SᵢYᵢ/πᵢ), where Sᵢ indicates selection. Its design-unbiasedness relies on known, correctly specified inclusion probabilities; with self-selection driven by unmeasured risk, those probabilities cannot be inferred merely from the selected sample. The general problem is captured in Heckman's study of selection bias (1979), whose regression solution depends on structural and distributional assumptions that must be examined rather than transplanted mechanically into every insurance setting.
Suppose a portfolio has 900 low-risk exposures with expected claim count 0.05 per exposure and 100 high-risk exposures with expected count 0.30. Expected total claims are 900(0.05)+100(0.30)=75; the whole-book expected rate is 0.075. A convenience sample containing 90 low-risk and all 100 high-risk exposures has expected count 34.5 over 190 exposures, or approximately 0.18158 per exposure. Naive extension of that sample rate to the complete book overstates the expected rate by about 142.1%. This arithmetic illustration isolates compositional bias and assumes fixed group rates; no observed insurer dataset is involved.
If the two true group rates are known, standardizing them to the target composition yields 0.9(0.05)+0.1(0.30)=0.075. In real applications, uncertainty about group rates, inclusion probabilities and the risk composition must also enter the result. Standardization cannot remedy missing risk groups for which there are no supportable estimates.
Bias, variance and effective sample size
In actuarial work large transaction counts may conceal a small number of independent risk-generating units. Thousands of claims associated with a handful of catastrophe events provide less information about tail-event frequency than thousands of independent observations. Correlation within employer groups, households, fleets, geographic zones and underwriting cohorts reduces effective information, sometimes dramatically. Treating each payment transaction as an independent claim can produce spuriously narrow uncertainty intervals.
For n equally correlated observations with pairwise correlation ρ and common variance, the variance of their mean is σ²[1+(n−1)ρ]/n. The corresponding effective sample size heuristic is n_eff=n/[1+(n−1)ρ]. With n=100 and ρ=0.05, n_eff≈16.81. This is an illustration of a specific exchangeable-correlation model, not a universal formula for catastrophe experience. It shows why the number of records is not equivalent to the number of independent risk realizations.
The conventional bias–variance trade-off must therefore be viewed alongside clustering, exposure credibility and process instability. Bootstrap replications that sample individual claims independently can break a policyholder or accident-year dependence structure. Resampling should preserve the unit whose variability is being evaluated—portfolio cohort, insured group, event or time block—while accounting for the limited count of independent clusters. The experimental work in Kohavi's validation-method comparison (1995) is useful for understanding resampling assumptions, while its classifier-specific findings should not be generalized as a universal insurance validation prescription.
Missingness and the limits of imputation
Rubin (1976) distinguished inferential conditions under which missing-data mechanisms can be ignored from those where the mechanism must be modeled. In the standard terminology, missing completely at random means that the occurrence of missingness is independent of both observed and missing values; missing at random means that, conditional on observed variables, missingness no longer depends on missing values; missing not at random means the missingness mechanism retains a dependence on unobserved values. The detailed ignorability result additionally requires parameter distinctness in likelihood or Bayesian analysis. The original missing-data paper does not warrant assuming that insurance files satisfy those conditions.
A repair estimate missing mainly for large, litigated claims is likely to have a different mechanism from one missing because a file was corrupted independently of severity. Filling missing amounts with the mean of observed payments can bias a severity estimate when missingness correlates with unresolved claims or high settlement amounts. Multiple imputation expresses uncertainty about missing values only to the extent that its model and missingness assumptions are defensible; adding many imputations to a misspecified mechanism creates numerical stability without identification.
The most transparent approach begins by tabulating missingness by policy vintage, claim state, adjuster, coverage, severity proxies and report date. Where missing-not-at-random scenarios remain plausible, sensitivity analyses should specify how outcomes among missing records might systematically differ from observed ones. An interval of outcomes across plausible assumptions is an analytical result with explicit premises, not a statistical confidence interval with guaranteed coverage.
Parameter estimation and credibility
Estimating a frequency parameter from Poisson events, a claim severity tail from sparse large claims or a development factor from a short triangle yields sampling uncertainty. Penalization, hierarchical pooling and credibility procedures can reduce estimator variability by introducing restrictions or sharing information across groups. These methods trade some bias for possible gains in out-of-sample precision, subject to assumptions about group relatedness and stability.
A shrinkage estimate λ̃=Z λ̂_group+(1−Z) λ̂_collective, for 0≤Z≤1, illustrates the information balance without prescribing a particular actuarial credibility formula. The individual and collective targets must refer to comparable coverage, exposure and calendar definitions. The epistemic question is what information justifies Z and whether the collective experience transfers to the specific risk. Better predictive performance in historical tests establishes a conditional empirical advantage; it does not prove that the shrunk estimate will remain suitable after a material underwriting rule or claim practice changes.
Censoring, Selection, and Change
Truncation, right censoring and delayed reporting
A censored observation reveals that an outcome exceeds or precedes a boundary without disclosing its exact value. A large insured loss capped by a policy limit can be observed only as a limited payment; a claim still open at valuation is censored with respect to ultimate settlement amount or time-to-closure; and a late notification may not appear in a current reported-claim sample. Truncation differs: an observation may be absent altogether because it falls below a reporting threshold or retention. These mechanisms require distinct likelihoods and cannot be repaired by the same adjustment.
Let ultimate severity be X and a payment limited to L be Y=min(X,L), ignoring deductibles and expenses. Then E[Y] generally differs from E[X]; indeed E[Y]=∫ from 0 to L of P(X>t)dt for nonnegative X under standard integrability conditions. Recovering E[X] from Y demands assumptions or independent information about the unobserved tail beyond L. Such assumptions are particularly material for excess-of-loss, catastrophe and liability portfolios, where high quantiles may be dominated by poorly observed events.
If accidents occur during calendar year t but are reported with delay D, then the expected observed count by valuation date v is an integral over occurrence times weighted by P(D≤v−occurrence time). Comparison of recent and mature accident years without development adjustment confounds reporting maturity with event incidence. A shift to digital notification that reduces D can create a temporary increase in reported claims even when occurrence intensity is stable. Conversely, a claims-handling backlog can create an apparent frequency reduction.
Claims development and valuation time
Claims reserving requires estimating payments and obligations after the current valuation date. Development triangles summarize information by accident or origin period and development age; they compress individual event histories and rely on assumptions about stability in development patterns. Mack's chain-ladder error analysis (1993) derives uncertainty measures within particular moment conditions. England and Verrall's stochastic survey (2002) examines predictive distributions and a variety of reserving specifications. Their contributions establish why the point estimate alone is an incomplete description of the future reserve outcome.
A development factor calibrated on earlier reporting practices can be misapplied after a change in notification, litigation, claims staffing or inflation. Although the model may match the older triangle, its forecast of the latest accident year incorporates a structural extrapolation. A serious review distinguishes estimation error in a factor from error in the assumption that the same factor remains appropriate. These sources of uncertainty generally require different evidence: bootstrap procedures address some sampling variation; scenario studies, operational change records and expert challenge address the structural premise.
Three dates deserve separate treatment: occurrence date identifies the insured event or origin cohort; valuation date determines how much information is already available; review date identifies which later information was considered by the analyst. The unpaid-claims estimation standard distinguishes accounting, valuation and review dates in applicable communication. Without these dates, a reserve comparison is vulnerable to hindsight and cannot establish what information supported the original estimate.
Behavioral selection and endogenous observation
Insurance data arise from an interaction among prices, policyholder choices, underwriting and claims handling. A premium increase can cause lower-risk customers to lapse; a stricter underwriting rule may alter both the portfolio's risk distribution and the observed relation of recorded features to loss; enhanced fraud screening can modify claim reporting or investigation outcomes. Historical observations thus reflect an equilibrium of prior insurance decisions, not an unfiltered sample of prospective risks.
Causal inference separates the consequence of an intervention from the association seen in historical data. Hernán and Robins' causal methods (2020) emphasize specification of the intervention, study population and identification assumptions. A model may predict higher claims among customers receiving an inspection because high-risk customers were preferentially inspected. It does not follow that inspections cause higher claims. For policy or pricing decisions, selection mechanisms and potential behavioral responses require consideration before associations are interpreted as effects.
There is a feedback problem when predictions change the data used to evaluate the predictions. If a model directs manual scrutiny toward certain policyholders, its own recommendations influence which fraud cases are investigated and confirmed. A subsequent model trained on confirmation labels can treat operational scrutiny as evidence of underlying fraud. Independent review can use randomized audits where appropriate, adjudication controls, alternative labels and assessment of differential investigation rates; these methods have legal and ethical constraints and require situation-specific justification.
Temporal drift and regime change
Distribution shift may affect marginal exposure characteristics P(X), conditional outcomes P(Y|X), the definition of Y, or the downstream decision function. Covariate shift is limited to changes in P(X) under maintained conditional stability; concept drift involves changes in the predictor–outcome relation. Gama and colleagues' drift survey (2014) shows that methods for identifying change depend on the form and operational setting of that change. In insurance, price inflation, health utilization, weather extremes, judicial trends and underwriting revisions can move multiple parts of the system simultaneously.
A monitor that compares only average claim costs cannot identify all distribution changes. Rising deductibles may reduce count frequency while increasing observed conditional severity; a high-limit book may preserve its mean but become more exposed to a heavy tail. Monitoring should therefore consider stratified distributions, exposure adjustments, coverage changes, reporting delays, tail behavior and model residuals. Thresholds require materiality and expected false-alert rates, rather than arbitrary universal percentages.
The validation population must reflect the planned use. A random split of a claims table often allows records from the same policy, individual catastrophe or evolving claim to appear in both training and test sets. Such leakage makes measured performance look better than truly forward use. Where the decision will occur next year, forward-chaining or origin-time validation provides a more relevant experiment, subject to the need for sufficiently mature outcomes and adequate effective sample size.
Predictive Validation and Backtesting
Validation begins with the decision and horizon
Validation asks whether the model and its use are reliable for their intended purpose. A model that estimates reporting probability, one-year premium cost, ultimate claim severity and one-year solvency capital can share inputs yet require distinct targets, horizons, calibration checks and decision losses. It is important to specify what is available at prediction time; a claim-status field created after the insured event cannot be used to score the customer at policy inception merely because it appears in the final historical extract.
The ASOP 56 modeling guidance (ASB, 2019) addresses model design and appropriate intended use, while Solvency II Article 124 requires periodic internal-model validation against experience and sensitivity to assumptions for its regulated scope. The methodological implication is that no single discrimination score, fit statistic or overall backtest suffices for every insurance use. A convincing validation portfolio examines the complete measurement and decision chain.
Backtesting must also distinguish observed outcomes from ultimate outcomes. For young claims, a model may correctly predict ultimate expense but appear inaccurate when assessed against incomplete paid-to-date amounts. Conversely, it may look accurate against a provisional proxy yet fail when outcomes mature. Truth labels therefore need a documented maturity convention, and changes in the convention need to be treated as changes to the evaluation task.
Calibration, discrimination and proper scoring
Calibration evaluates whether predicted probabilities or distributions correspond to eventual frequencies or outcomes under the relevant information set. Discrimination evaluates separation or ranking across units. A classifier can rank higher risks well but systematically understate every group's probability. Ranking metrics then coexist with material errors in aggregate pricing or capital projection. Likewise, a well-calibrated average can conceal severe subgroup errors.
For binary events Y∈{0,1} with forecast probability p, the Brier loss (p−Y)² is a proper scoring rule. Logarithmic loss assesses probabilistic forecasts with especially severe penalties for confidently incorrect extremes. Gneiting and Raftery's scoring framework (2007) establishes the theoretical basis for proper scoring, and Gneiting and Katzfuss' review (2014) examines calibration and sharpness of distributional forecasts. These works support multi-dimensional evaluation rather than a universal choice of metric.
For an illustrative homogeneous validation cohort with true frequency 0.20, a constant prediction p=0.10 has expected Brier loss 0.2(0.9²)+0.8(0.1²)=0.17. A correctly calibrated constant p=0.20 has expected loss 0.2(0.8²)+0.8(0.2²)=0.16. The gain of 0.01 in expected Brier loss is mathematically real under the specified population; its financial relevance must be assessed through the underwriting or claims decision and need not translate into a fixed premium amount.
Calibration can be studied by forecast bins, population segments, exposure weights, calendar cohorts and quantiles. Bin-level charts have substantial sampling uncertainty when events are rare. A nominally precise calibration curve is unreliable when plotted with no attention to the number of independent outcomes in each bin. Distributional checks may include probability integral transforms and interval coverage under the relevant assumptions, with special scrutiny of upper-tail forecasts for liability and catastrophe uses.
Temporal and grouped experimental design
A defensible insurance split holds back future origin periods or independent insured units according to the decision being simulated. Cross-validation can help compare candidate methods within historical data, but it must be nested appropriately when the same data are used for model selection and evaluation. Features created from future or linked-entity outcomes must be excluded from every training fold that would not have had them at deployment. Kohavi's resampling study (1995) documents important biases and variance properties of common validation methods under its experimental conditions; insurance-specific splits must additionally account for time, cluster and loss maturation.
At portfolio scale, group-level error analysis is essential. The aggregate performance metric can favor abundant low-risk policyholders even if the model fails for small high-severity commercial groups. Conversely, a capital model's adequacy may depend on the tail of event aggregation rather than average policy-level accuracy. A validation report should preserve results at the level where the insurer incurs financial loss and makes decisions, even when the model itself is trained at a finer unit.
Backtests of liability and capital estimates
Reserve backtesting compares a valuation-date estimate with subsequent paid and reported development, controlling for new information and accounting or claim-processing changes. The observed result is one realization from a distribution; a difference between forecast and outcome is therefore neither proof of faulty modeling nor grounds for unqualified reassurance. Over many independent origin periods and comparable portfolios, systematic directional bias and excessive prediction-interval failures provide stronger evidence. Sparse independent development periods limit the power of such tests.
Under EU Solvency II, Article 121 statistical quality and Article 122 capital calibration attach specific requirements to approved internal models. They do not prescribe a universal diagnostic that can be copied into every insurance estimate. The delegated validation process rule concerns internal model validation and independence; it should be read alongside the applicable Directive and delegated-regulation scope. A capital estimate also has to be assessed against the probability distribution of aggregate risk, not just the expected claim cost.
A tail-risk backtest is intrinsically information-constrained. An annual 99.5th-percentile loss benchmark, for example, corresponds under a correctly specified continuous distribution to an exceedance probability of approximately 0.5% each year. Few exceedances over a short historical window provide little statistical power to validate the tail. Stress testing, sensitivity analysis, independent model comparisons, scenario narratives and conservative limitations are therefore complementary sources of evidence rather than substitutes for mathematical calibration.
Reproducibility and methodological audit
Validation should freeze the prediction timestamp, training sources, labeling policy, software revision, hyperparameters, manual overrides, exposure weights and the portfolio membership evaluated. A model can be technologically reproducible yet scientifically non-reproducible if the target label is later revised or if a source system silently reclassifies closed claims. Reproducible review requires both computational and institutional versioning.
Every evaluation should state the model's intended population; the target and its maturity; exclusions; the available sample size and effective number of independent units; the baselines used; the metrics; uncertainty intervals; drift and subgroup diagnostics; exceptions; and the operational consequence of failing a criterion. This permits qualified independent challenge and makes validation a decision record rather than a ritual scorecard.
Parameter and Structural Model Risk
Process risk, parameter risk and model uncertainty
Process variability is the randomness of future outcomes conditional on a stipulated model and its parameters. Parameter uncertainty is the uncertainty in estimated parameters given finite observations. Model uncertainty concerns competing or incorrect structural assumptions, including distributional family, dependence, reporting mechanism and behavioral stability. These components interact, and a single numerical range seldom separates them uniquely without assumptions.
Consider an aggregate loss S=Σ from i=1 to N of Xᵢ, where N is Poisson with mean λ and Xᵢ are independent and identically distributed positive severities, independent of N, with mean μ and variance σ². Then E[S]=λμ and Var(S)=λ(σ²+μ²). These identities derive from total expectation and total variance under the stated compound-Poisson assumptions. They are not general formulas for catastrophe-correlated claims, common inflation shocks or endogenous reporting.
For λ=100, μ=1,000 and σ=2,000 monetary units, E[S]=100,000 and Var(S)=500,000,000, so the process standard deviation is approximately 22,360.68. This calculation conditions on fixed λ, μ and σ². If λ is estimated rather than known, the law of total variance applied to parameter uncertainty produces an additional term Var(E[S|parameters]); changes in dependence or structural family can introduce further effects. Publishing 22,360.68 as the total forecast error would confuse conditional process volatility with full predictive uncertainty.
Uncertainty propagation and bias amplification
Loss estimates often feed a sequence of transformations: estimated frequency times severity, projected ultimate loss, discounting, risk margin, retained loss after reinsurance and a capital measure. Errors can amplify through nonlinear operations. If a capital quantile depends heavily on tail shape, a small mean error may be less consequential than a modest shift in tail parameter or dependence. The total sensitivity of a downstream quantity to upstream inputs can be investigated with derivatives, finite perturbations or full simulation under explicitly chosen scenarios.
For a simplified expected-cost estimate C=eλμ, relative small changes satisfy approximately dC/C = de/e + dλ/λ + dμ/μ, under a first-order differential approximation for positive quantities. This shows how independent sources of measurement and estimation bias enter the output, but it does not imply that their probabilities or errors are independent. Covariances and shared scenario drivers matter in a probabilistic analysis.
In claims inflation projection, suppose current average severity is 10,000 and a forecast projects three years of compounding. At a constant annual rate of 4%, projected severity equals 10,000(1.04)³=11,248.64. At 2%, it equals 10,000(1.02)³=10,612.08. The difference is 636.56 per projected claim, or 127,312 for 200 claims with identical assumed timing. The exercise isolates an assumption sensitivity; it says nothing about which inflation rate is actually correct, nor about variation across coverages, judicial outcomes and calendar periods.
Structural misspecification and false precision
Statistical models typically operate within families of possible distributions. A Poisson model enforces equality of mean and variance conditional on its rate, whereas heterogeneous exposure or event clustering may produce overdispersion. A thin-tailed severity family may fit numerous small losses but understate aggregate losses dominated by rare very large claims. An independence assumption among underwriting risks may be reasonable for routine incidents but untenable for a regional catastrophe or common legal ruling.
A parameter confidence interval conditional on one selected model does not reflect uncertainty about which family should have been used. Likewise, generating thousands of simulated results from one unjustified distribution improves computational resolution of a conditional calculation without establishing the distribution's validity. Material structural assumptions should be listed as hypotheses tested against data where possible and stress-tested where they cannot be empirically identified.
The actuarial significance is reflected in Solvency II's actuarial-technique rule, which explicitly discusses material model and estimation error in the EU internal-model context. The actuarial modeling standard requires attention to intended use, model structure and material assumptions for covered actuarial work. The scientific principle is more general: a precise output is credible only when the premises that make its precision meaningful are supported.
Model risk at the portfolio and institution level
Multiple models can share a latent exposure dataset, severity index, catastrophe vendor, scenario generator or economic assumption. Their errors may be correlated even if model owners and software platforms are independent. An insurer that uses one underestimated inflation series in pricing, claims reserving, capital and reinsurance purchase decisions has a concentrated assumption exposure. Each model could pass a narrow local test while the organization simultaneously understates its overall liability and risk transfer need.
A group-level model inventory should therefore document upstream data dependencies, shared economic assumptions, aggregation and reporting pathways, third-party dependencies, model versions and responsible decision makers. Model concentration can be analyzed with common shocks and joint sensitivity scenarios. This is one area where the Federal Reserve's revised 2026 guidance provides a useful comparative banking example of institution-wide model-risk management; it superseded the historical U.S. banking SR 11-7 on April 17, 2026 and is not an insurance supervision requirement.
A similar jurisdictional distinction applies to the PRA's SS1/23 banking statement, which concerns UK banks and specified investment firms rather than insurers or reinsurers. Its inventory, independent validation and model-risk-mitigation concepts may be studied comparatively, but citing it as direct insurance law would be incorrect. For insurers, applicable local insurance regulation, IAIS principles and professional standards supply the appropriate institutional foundation.
Counterfactual stability and decision sensitivity
The materiality of model uncertainty depends on decisions, not just fit. Suppose two plausible severity specifications yield average cost estimates differing by 3%, but both place the same risk well within an insurer's underwriting limit. The resulting decision difference may be small. Suppose the same specifications imply materially different probabilities of breaching a reinsurance attachment point; the difference may be decisive despite nearly identical expected loss. The key calculation links uncertain model outputs to the action chosen and its plausible outcomes.
One can write expected decision loss under a state ω and action a as E[L(a,ω)|information]. If using an incorrect estimate changes the selected action from a* to a', the regret conditional on an assumed reference distribution is E[L(a',ω)−L(a*,ω)]. That expression is an ordinary decision-theoretic comparison. It must be accompanied by the assumption that defines the reference distribution, and cannot itself resolve ambiguity about the true data-generating mechanism.
This comparison makes the governance threshold rational: model errors that alter capital adequacy, reinsurance structure, consumer treatment or financial reporting require more intensive challenge than numerically similar errors in low-materiality internal dashboards. A governance program can prioritize attention by the severity of plausible decisions induced by uncertain estimates while preserving transparency about unquantifiable structural uncertainty.
Governance and Communication of Uncertainty
Internal control, actuarial independence and oversight
Insurance model governance translates epistemic limitations into accountable organizational processes. It establishes who owns the data, who specifies the estimation purpose, who develops the model, who challenges it independently, who approves its use and who acts when performance deteriorates. Governance requires meaningful technical competence and authority to alter a decision, not merely a signature following completion of a checklist.
ICP 8 and ICP 16 articulate risk-management and enterprise-risk-management principles for insurance supervision, including internal controls, actuarial functions and ORSA-related risk assessment. The IAA standards register includes international model standards for insurer enterprise risk models and enterprise risk management; these are model professional standards rather than directly binding rules in every country. Relevant domestic law and applicable professional standards determine the enforceable duty.
Within the European Union, Article 241 validation requirements require independence in internal-model validation as specified in that regime. For U.S. actuaries engaged in covered work, ASOP 56 sets actuarial modeling guidance; for technical-provision estimates under Solvency II, Article 265 documentation addresses the origins of data, assumptions, methods and validation. These frameworks overlap in subject matter but are not interchangeable legal instruments.
What independent challenge must examine
An independent model reviewer must be able to reconstruct the intended estimation problem before assessing software quality. The review begins with the insured risk, contractual coverage, estimand, valuation date, source-system extraction and population definition. It then asks whether selection, missingness and maturity are identified; whether external calibration data are comparable; whether training and validation are separated at the proper unit; and whether alternative assumptions could change material decisions.
A useful challenge record distinguishes detected defects, untested assumptions, limitations judged acceptable with compensating controls and issues requiring changes. It also identifies whether an issue concerns data correctness, estimand mismatch, estimation method, structural uncertainty, operational deployment or communication. Different problems require different remedies: a duplicate claim should be reconciled at source; a changed underwriting cohort may require reweighting or redesign; an unobserved catastrophe tail may require scenario limitation and explicit governance appetite.
Third-party models deserve identical scientific questions even where the source code is unavailable. Vendor assurance can provide a description of intended use, validation evidence, input requirements, empirical domain, update history and known limitations. It cannot erase uncertainty caused by undisclosed dependence structure or an unexplained training population. The NAIC's current AI work program describes regulatory attention to third-party data and models in U.S. state insurance supervision; implementation and obligations vary across jurisdictions.
AI systems as an extension of established model-risk concerns
AI-enabled insurance decision systems add complexity through learned representations, proxy variables, feedback loops, frequent retraining and limited interpretability. Underlying questions about data provenance, model selection, drift, lawful use and outcomes remain recognizable from established actuarial and statistical disciplines. The IAIS 2025 AI Application Paper emphasizes risk-based proportionality, accountability, robustness, transparency, explainability, fairness and redress in insurance supervision.
In the United States, the NAIC's 2023 model bulletin was adopted as a model bulletin, not a self-executing nationwide insurance law; state adoption and locally applicable provisions must be identified before claiming legal duties. The NIST AI Risk Management Framework is voluntary and cross-sectoral, providing complementary risk-management language without replacing insurance regulation. A research dossier focused on inference should not collapse actuarial risk into an AI-only topic: simple reserving spreadsheets and static underwriting tables can carry consequential structural errors.
Fairness is both a governance and inference issue. A policy variable may correlate with protected status through neighborhood, income, credit characteristics or claims-management history. Errors in data labels can be concentrated in disadvantaged groups. Overall calibration can coexist with distributionally unequal error rates or contested allocation of risk-pooling costs. Scientific validation should identify group-specific evidence where lawful and feasible; legal analysis must separately determine the applicable anti-discrimination rules and permissible rating factors.
The uncertainty statement as a scientific output
A competent uncertainty statement names the quantity estimated, period, coverage definition, population, method, principal data and assumptions, process variability, parameter uncertainty, structural limitations, plausible alternative scenarios and known conditions under which the estimate would cease to be appropriate. It also names the intended decision and the person or function responsible for review. These elements allow readers to distinguish a conditional probabilistic interval from a scenario range or a judgmental margin.
For covered U.S. actuarial communications, ASOP 41 communication standards address reporting assumptions and methods. ASOP 43 unpaid-claim disclosures separately addresses constraints, methods and the basis of reserve measures in its scope. A headline such as 'reserve 100 million ± 5 million' has insufficient meaning until one knows whether the interval captures a selected quantile of process variability, conditional parameter error, several independent model specifications or a range of judgmental choices.
There is a material difference between uncertainty communication and ex post explanation. Ex post explanations may identify which features moved a model's score, while the statement of uncertainty explains whether the score is reliable for its intended prospective use. Both can be required. A detailed local explanation cannot repair an invalid data pipeline, unsupported extrapolation or unstable conditional distribution.
Monitoring, interventions and retirement
Model approval is a time-bounded decision. Monitoring should trigger review when the relationship between data and the intended decision changes, when observed claims development breaches pre-specified tolerances, when the source schema changes, when a model is applied to a new portfolio or when relevant law and contract terms change. A documented response may include continued use with controls, recalibration, restriction of use, revalidation, fallback to a simpler estimate or model retirement.
Automatic retraining can suppress some drift in measured fit while obscuring changes in data labels, selection and discrimination. Change control should therefore record the trigger, revised assumptions, validation design, independent approval and impact on downstream reports. The concept drift literature provides methods for recognizing evolving prediction relationships, but the choice of intervention remains a risk and governance judgment under the insurer's obligations.
Analytical Case Studies and Decision Sensitivity
Case study: underwriting selection and data representativeness
The two-group example above reveals a fundamental representativeness failure. The target portfolio has 1,000 exposures and expected claims of 75, whereas the nonrepresentative selected group yields a rate of 34.5/190≈0.18158. Suppose an analyst reports a rate with a narrow Poisson standard error computed only from the selected count. Even with a large sample, that interval would quantify conditional sampling variation around the selected mix, not the 0.075 whole-book target.
The analytical remedy is to document and, where possible, estimate the inclusion mechanism, preserve valid exposure denominators, stratify the target portfolio and perform scenario sensitivity for excluded or underrepresented groups. An approved forecast for a new portfolio must specify whether risk composition is expected to reproduce the historical book. This inference is consistent with selection-bias research and the model-purpose discipline of actuarial modeling guidance; it is an application of established principles, not a new selection theory.
Case study: reversal from unequal claims reporting maturity
Consider two hypothetical accident cohorts each with 100 ultimately reportable claims. Cohort A has 60% reported by the evaluation date, yielding 60 observed claims; cohort B has 80% reported, yielding 80 observed claims. If A had 100 exposures and B had 120, the observed rates are 0.60 and 0.667, suggesting B has a higher incidence. The ultimate rates are 1.00 and 0.833, reversing that ordering. All counts are stylized and deliberately high for arithmetic clarity.
The reversal arises from different maturity and exposure denominators, not an unusual machine-learning failure. More sophisticated fitting of the incomplete observations can strengthen the wrong ordering if the reporting mechanism is not identified. In a claims triangle or individual event-history design, reporting delay and its changes should be studied before direct comparison. Stochastic claims reserving research provides context for uncertain development, while the EU data-quality rule shows its institutional relevance where applicable.
Case study: inflation uncertainty and linked decisions
With hypothetical severity 10,000, 200 expected claims and a three-year horizon, the 4% versus 2% inflation assumption difference creates a projected total-cost difference of 127,312. A simplified reserve adjustment could match that amount if all other assumptions were fixed. A real insurer would also face timing dispersion, changing claim counts, social and legal inflation, discounting, coverage changes and possible reinsurance recoveries.
If the same inflation forecast is used in pricing, reserving and capital, the error's organizational significance cannot be found by simply adding three model standard errors that share the same assumption. Joint scenario analysis exposes the correlated dependency. The institutional response is a source-traceable common-assumption review, independent challenge and consistent communication of downstream exposure. Technical-provision lineage requirements and insurance ERM principles support the need for such traceability in their respective contexts.
Case study: a correct ranking that makes a wrong decision
Imagine that a model assigns every customer a predicted claim probability half the actual probability, while preserving the exact rank ordering of risks. Its ranking statistic could remain unchanged. If an underwriting decision uses absolute expected loss to set premium or retention, however, the systematic calibration error is material. Under a cost-sensitive decision rule, the threshold-crossing customers and thus the chosen book can differ from those selected by a calibrated model.
The Brier example above demonstrates a local score improvement from 0.17 to 0.16 when a constant predicted probability moves from 0.10 to the true population frequency 0.20. This provides a statistical metric; decision-specific evaluation would additionally quantify changes in premium adequacy, coverage availability or expected portfolio loss. Proper scoring provides disciplined incentives for probabilistic forecasts (proper scoring rule theory), while the choice and social consequences of insurance decisions require additional evidence and applicable regulation.
Case study: an opaque risk concentration across models
Consider three independently governed model teams using one purchased geospatial exposure file. A systematic geocoding error moves industrial properties away from a high-hazard zone in that source. Pricing, catastrophe accumulation and reinsurance purchase models may all understate exposure for the same underlying reason. Independent validation of the software and arithmetic of each model does not demonstrate independence of error in their upstream data.
A group-level challenge would reconcile the vendor file with independent address or property records, map the shared dependency, quantify plausible reclassification stress and disclose which outputs and decisions are affected. This exemplifies the integration of data lineage with institutional model risk and the relevance of third-party model controls. The U.S. regulatory discussion of third-party data and the IAIS AI supervisory paper supply insurance-sector context without implying that every purchased geospatial tool is necessarily an AI system.
Synthesis and Research Implications
The material surveyed supports a consistent analytical conclusion: estimation quality is jointly determined by the truth concept encoded in the insurance contract, the population to which the conclusion applies, the observation mechanism that generates data, the identification and modeling assumptions, and the decision for which the estimate is used. This is a comparative synthesis of recognized statistical and actuarial findings rather than the definition of a new named model or independent doctrine.
Five implications follow. First, a data-quality audit has to address the target of inference, not only the absence of missing fields. Second, selection, censoring and temporal maturation should be treated as structural features of insurance records rather than incidental preprocessing inconveniences. Third, a model's conditional fit and nominal confidence interval must be read alongside uncertainty about the assumptions and population on which they depend. Fourth, forecasting diagnostics should be designed from the decision horizon and financial materiality backward to the validation sample. Fifth, risk governance must follow shared sources and assumptions across the insurer's models, because many failures arise from a common error rather than from an individual algorithm.
These implications also establish a disciplined boundary between scientific and legal claims. The IAIS principles are international insurance supervisory principles; Solvency II internal-model validation has an EU statutory framework; ASOP 23 applies to actuarial services within its professional scope; the NAIC model bulletin requires jurisdiction-specific adoption analysis; and SR 26-2 governs banking organizations, not insurance firms as such. Treating them as universally interchangeable would produce an institutional version of the very model-specification error this dossier studies.
Open research questions include how to validate catastrophe and systemic tail risks with severely limited independent realizations; how to distinguish a genuine hazard change from revisions in reporting and underwriting selection; how to quantify uncertainty where vendor inputs are only partially auditable; how to reconcile individual predictive precision with legally required risk pooling and fairness; and how to avoid shared assumption failure across financial, underwriting and capital processes. Empirical progress requires versioned longitudinal insurance data, public benchmark definitions, explicit evaluation populations, and transparent documentation of assumptions that remain untestable.
Conclusion
Insurance Data, Estimation, and Model Risk establishes that the reliability of insurance inference depends on the integrity of the entire path from contracted exposure to recorded outcome, estimated parameter, predicted distribution and consequential decision. The same statistical estimator can be appropriate in one portfolio and misleading in another because risk selection, claims maturation, contract terms or future use differ. The numerical case studies demonstrate how exposure definitions, sample selection, reporting delay, inflation assumptions and calibration can change the sign or magnitude of a result without any programming defect.
Estimation uncertainty cannot be exhausted by the variability calculated within one model. Insurance scientific judgment must separately examine sampling, missingness, structural assumptions, dependency, distribution shift and operational feedback. Suitable validation combines prospective testing, financial decision alignment, probabilistic calibration, subgroup review, scenario analysis and independent challenge. International supervisory and actuarial sources converge on the need for controlled methods, data quality, transparent assumptions and accountable use while retaining distinct legal and professional scopes.
The original contribution of this monograph is its integrated, cross-disciplinary interpretation of how errors in the data-generation and estimation chain become financial and institutional risks, together with explicit derived examples that make those transitions auditable. The manuscript does not claim to have established a new theory, solved identification problems lacking the necessary data, replicated restricted empirical studies or undergone external peer review. Its findings provide a research basis for more precise actuarial estimation and model-risk assessment in the international science of insurance.
References
Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. https://www.actuarialstandardsboard.org/asops/modeling-3/
Actuarial Standards Board (2010). Actuarial Standard of Practice No. 41: Actuarial Communications. https://www.actuarialstandardsboard.org/asops/actuarial-communications/
Actuarial Standards Board (2011 revision). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. https://www.actuarialstandardsboard.org/asops/propertycasualty-unpaid-claim-estimates/
International Actuarial Association (2018–2025). International Standards of Actuarial Practice, including ISAP 1, ISAP 5 and ISAP 6: official standards register. https://actuaries.org/publications/international-standards-of-actuarial-practices/
International Association of Insurance Supervisors (current online collection, accessed 2026). Insurance Core Principles and ComFrame, especially ICP 8 and ICP 16. https://www.iais.org/icp-online-tool/
International Association of Insurance Supervisors (2025). Application Paper on the Supervision of Artificial Intelligence. https://www.iais.org/2025/07/the-iais-publishes-application-paper-on-the-supervision-of-artificial-intelligence/
European Union (2009, consolidated rulebook). Directive 2009/138/EC, Solvency II, Article 82: Data quality for technical provisions; EIOPA Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2170_en
European Union (2009, consolidated rulebook). Directive 2009/138/EC, Solvency II, Article 121: Statistical quality standards. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2208_en
European Union (2009, consolidated rulebook). Directive 2009/138/EC, Solvency II, Article 124: Validation standards. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2211_en
European Commission (2015, consolidated rulebook). Delegated Regulation (EU) 2015/35, Article 229: Adequate, applicable and relevant actuarial techniques. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5886_en
European Commission (2015, consolidated rulebook). Delegated Regulation (EU) 2015/35, Article 241: Model validation process. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5900_en
European Commission (2015, consolidated rulebook). Delegated Regulation (EU) 2015/35, Article 265: Technical provisions documentation. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5928_en
European Union (2009, consolidated rulebook). Directive 2009/138/EC, Solvency II, Article 122: Calibration standards. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2209_en
International Accounting Standards Board (2023 effective date). IFRS 17: Insurance Contracts. IFRS Foundation standard overview. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
National Association of Insurance Commissioners (2023). NAIC Members Approve Model Bulletin on Use of AI by Insurers. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers
National Association of Insurance Commissioners (2026). Artificial Intelligence: insurance regulatory topic page. https://content.naic.org/insurance-topics/artificial-intelligence
Board of Governors of the Federal Reserve System (2026). SR 26-2: Revised Guidance on Model Risk Management. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm
Prudential Regulation Authority (2026 current version). SS1/23: Model Risk Management Principles for Banks. https://www.bankofengland.co.uk/prudential-regulation/publication/2023/may/model-risk-management-principles-for-banks-ss
Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. doi:10.1093/biomet/63.3.581. https://academic.oup.com/biomet/article-abstract/63/3/581/270932
Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica, 47(1), 153–161. doi:10.2307/1912352. https://ideas.repec.org/a/ecm/emetrp/v47y1979i1p153-61.html
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://www.casact.org/abstract/distribution-free-calculation-standard-error-chain-ladder-reserve-estimates
England, P. D., and Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. doi:10.1017/S1357321700003809. https://www.cambridge.org/core/journals/british-actuarial-journal/article/abs/stochastic-claims-reserving-in-general-insurance/60026990B6A88E8A6DDEECABD6506C65
Gneiting, T., and Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. doi:10.1198/016214506000001437. https://doi.org/10.1198/016214506000001437
Gneiting, T., and Katzfuss, M. (2014). Probabilistic forecasting. Annual Review of Statistics and Its Application, 1, 125–151. doi:10.1146/annurev-statistics-062713-085831. https://www.annualreviews.org/content/journals/10.1146/annurev-statistics-062713-085831
Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of IJCAI-95, 1137–1144. https://ijcai.org/proceedings/1995/147
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., and Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4), Article 44. doi:10.1145/2523813. https://doi.org/10.1145/2523813
Hernán, M. A., and Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC; author-hosted book. https://miguelhernan.org/whatifbook