RESEARCH DOSSIER
Insurance Statistical Estimands and Prediction Targets
Abstract. Insurance prediction begins with the definition of the quantity that a statistical procedure is supposed to learn. A policyholder's probability of at least one covered event, the expected number of claim occurrences, the mean size of an individual claim, the expected aggregate insured loss, the ultimate cost of an accident year, and the cash paid by a valuation date represent distinct statistical objects. Their differences arise from the units under observation, contractual coverage, reporting processes, exposure duration, target populations, and information available when a decision is made. This monograph examines those differences as a foundational problem of statistical inference in insurance and risk science. It develops an explicit conceptual account of insurance estimands and prediction targets, formalizes frequency–severity and aggregate-loss relations, analyzes conditioning and selection effects, differentiates prospective, retrospective, developmental, and causal questions, and connects target definitions to validation procedures and decisions. The research draws on primary actuarial standards, insurance mathematics, probabilistic forecast evaluation, reserving research, financial reporting principles, and literature on calibration, data leakage, and causal identification. Original contributions consist of formal interpretive reconciliation, counterexamples, and a practical scientific verification procedure derived from established methods. Illustrations are synthetic and are checked by elementary probability identities; they are not empirical insurer findings. The central conclusion is that model validity is a relation among a precisely specified target, a compatible observation process, an admissible information set, an appropriate loss function, and the decision to be supported.
Keywords: Insurance estimands; prediction targets; statistical inference; insurance exposure; frequency; severity; pure premium; insurance contracts; claim reporting delay; target leakage; actuarial modeling; model calibration; claims reserving; causal inference; predictive validation.
Introduction
The statistical problem addressed in this Research Dossier is the specification of a population, an insured event, a contractual financial consequence, an observation interval, and a mathematical quantity before a model is estimated or a forecast is produced. Insurance data frequently occupy multiple systems and time scales. An application records an applicant; a policy records a contract; a risk register records an insured exposure; a claim file records an administrative case; a transaction system records payments; and an accounting system recognizes obligations under specified reporting rules. These units are related, but the relation must be reconstructed. A predictive algorithm can fit an available column exceptionally well while the column measures an administrative artifact instead of the financial or probabilistic question intended.
The CAS guide to insurance rating GLMs (Goldburd et al., 2025) presents statistical models in the context of developing practical rating plans. The United States ASOP No. 53 explicitly connects the intended measure of future cost to the purpose of the work, coverage, exposure bases, assumptions, and historical experience. These two authoritative materials converge on the same scientific requirement: the quantity to be predicted cannot be inferred from the available algorithm alone. An observed claim count, for instance, is partly a consequence of an insured event process and partly a consequence of reporting, policy terms, and file organization.
A useful precedent for precision in defining research targets is the ICH E9(R1) estimand guideline (2019). It distinguishes the scientific question, target estimand, estimation method, numerical estimate, and sensitivity analysis. Its binding subject is clinical-trial methodology, particularly treatment effects; the present study adopts the logical discipline of explicit target specification as an analogy for insurance science. Insurance prediction targets are frequently descriptive or prospective rather than causal treatment effects, so direct transfer of clinical estimand attributes would obscure the very difference this monograph seeks to clarify.
The research method combines a review of institutional sources and primary peer-reviewed literature with independent probability derivations and worked examples. Each mathematical statement below specifies its operative assumptions. Numerical cases use expressly hypothetical data so that arithmetic and conceptual distinctions can be independently reproduced. Cross-jurisdictional material identifies the reach of relevant standards rather than presuming one international legal regime. The object is foundational analytical synthesis; no newly named theory, causal effect estimate, insurer benchmark, or claim of independently verified novelty is advanced.
Insurance Estimands and Contracts
The statistical object before the estimator
An estimand is a precisely defined statistical quantity associated with a target population and specified outcome-generating conditions. In an insurance setting a population-level estimand may be a probability, expectation, rate, quantile, present value, difference, or distributional functional. A prediction target is an outcome or functional whose future value or conditional distribution must be forecast using information available at a specified time. A fitted model is an estimator or prediction procedure; its numerical output is an estimate or forecast. These definitions establish what is measured, how it is measured, and when a claim about it can be evaluated.
Let index an insured unit or contractual exposure, denote the prediction origin, the future horizon, a defined set of contract terms, and the population distribution of interest. Let denote underlying insured-event experience over the horizon. Let map those events and their economic consequences into a contractually defined payment or benefit. A conditional monetary target may be written as
This expression is a definition, not a claim that a particular predictive method identifies the expectation. The outcome must exist on the chosen probability space; the contractual response must be measurable and relevant; conditional expectations require the relevant integrability assumptions; and the conditioning variables must be specified as information genuinely available at the prediction origin. If future claims or payments remain unsettled at observation cutoff , a measurement process also enters the problem. The observed variable can be written , where represents claim reporting, case-reserve updates, transaction recognition, and data availability. The inferential target is generally a functional of or ; the training label may be a functional of . The gap is a scientific question requiring evidence or a sensitivity analysis.
For complete specification, the analyst must identify at least the target population; statistical unit; event or financial endpoint; covered transformation; exposure denominator; starting time and duration; information set; aggregation level; statistical functional; measurement and development rules; and intended decision. The list is an editorial synthesis of existing statistical and actuarial concerns, not a new named conceptual system. It resembles the ICH insistence on defining a target before selecting estimators, while including insurance-specific contractual and reporting dimensions.
Contractual responses and coverage
The economic severity of an incident and the amount payable by an insurer differ under deductibles, limits, exclusions, retentions, coinsurance, attachment points, aggregate limits, waiting periods, and settlement provisions. For a simple covered occurrence with ground-up economic loss , deductible , and payment limit , a schematic payment function is
The formula assumes eligible coverage, one occurrence, no additional conditions, and a limit measured above the deductible. It therefore functions as a controlled example rather than a complete insurance contract. Under and , an eligible loss of creates zero modeled insurer payment, while a $15{,}000$ loss yields $9{,}000$. A model of ground-up loss severity and a model of insurer payment severity consequently address different targets despite sharing the underlying incident.
A policy may have the same economic hazard under two contractual structures and different expected insurer costs. For prospective pricing, this distinction is essential: the risk-transfer object is the contractually borne amount, together with relevant expense components, rather than every economic consequence experienced by the insured. ASOP No. 53's intended-measure provisions require the practitioner to consider the applicable coverage and relevant elements of future cost. For life, annuity, disability, or health benefits the contractual transformation can depend on survival, benefit eligibility, waiting periods, medical definitions, renewals, or payment schedules; it therefore cannot generally be reduced to a property-loss deductible formula.
An insurer can also observe several monetary versions of the same claim: the first reported estimate, paid-to-date, current case reserve, incurred amount, ultimate modeled obligation, allocated expense, recoveries, and net-of-reinsurance amount. Each version creates a distinct response variable, with different maturity, measurement error, and decision relevance. A statistical claim that a model forecasts “the claim amount” is scientifically incomplete until the intended version is stated.
Estimation, forecasting, and decision use
A retrospective estimand might be the average ultimate insured cost for policies effective in a completed calendar year. A prospective target might be the expected covered cost for a new policy over the coming twelve months. A reserving target might be the remaining future payments for occurrences that have already happened at valuation date. Their apparent similarity is misleading: the temporal conditions, relevant exposure cohorts, and sources of uncertainty differ.
A probabilistic forecast provides , where denotes all admissible information at the decision date. A point forecast may use the mean, median, or a specified quantile of this distribution. The choice depends on the decision loss function: squared-error optimality generally selects the conditional mean; absolute-error optimality selects a conditional median; asymmetric linear loss selects a corresponding quantile. A model trained to minimize squared error cannot automatically be interpreted as an estimator of a tail quantile.
A decision can require a different object from a prediction. A risk-control choice depends on what would happen under alternative interventions, while an underwriting score may describe conditional associations under existing business operations. The distinction is developed in Hernán and Robins' causal-inference treatment (2020). The insurance application is direct in logic: a predicted high claims rate identifies a forecasting problem; it does not independently identify the reduction in claims attributable to changing behavior, premium, or policy terms.
Policy and Claim Statistical Units
Policies, insured units, occurrences, and claim records
Insurance observations have several levels: customer, household, enterprise, policy, coverage section, vehicle, building, location, employee, covered life, occurrence, claim, and payment. One enterprise can hold multiple policies; one policy can insure several objects; one catastrophe can generate many claims; one claim can contain several transactions. Consequently, a row of data is a database unit and may be different from the statistical unit of the estimand.
When predicting the probability that a policy generates at least one covered claim in a year, the target is a policy-level indicator . When predicting claim counts, the outcome is . For over a fixed exposure period, . Thus a claim frequency of per policy-year corresponds to a probability of at least one claim of approximately , not , under this model. The two quantities become close for small rates but remain mathematically distinct. Using a binary classifier's calibrated probability as a Poisson rate requires a justified transformation and suitable count-process assumptions.
At the next level, a single occurrence may generate multiple reported claim files or reopenings. Counting files rather than independent occurrences changes the claim-frequency estimand. Payments are even more granular: splitting an ultimate payment into several transactions should not mechanically create a higher occurrence frequency. Insurer operations must therefore establish a consistent entity-resolution and event-definition rule before modeling, especially when business mergers or system migrations alter identifiers.
Exposure and denominators
An actuarial claim rate ordinarily requires a denominator reflecting the time or quantity at risk. If each policy contributes exposure and claims, a pooled observed claim frequency is . This is an exposure-weighted empirical rate. The average of individual ratios, , answers a different question and may be unstable when is very small. The distinction matters when policies begin or terminate within the observation year or coverage is intermittently suspended.
A typical Poisson frequency model writes
The offset scales the exposure in a specified counting process. It is justified when claim occurrence intensity is proportional to the designated exposure under the model's assumptions. A year of vehicle insurance, one vehicle-kilometer, a payroll amount, a building-year, and a life-year are different exposure bases. Each has domain-specific appropriateness and may produce a different estimand. The CAS insurance GLM monograph documents practical rating-model construction, while the ASOP No. 53 exposure-base discussion addresses the relationship between exposure base and prospective cost.
Observing only policies that completed twelve months of coverage can produce an unintended survival or renewal-selected population. For instance, early cancellations can be associated with circumstances that also predict claims. Without a clear target population, dividing claims by full-year exposure for all policies can distort both denominators and conditional comparisons. Policy adjustments, changes in insured values, and shifting deductible structures also demand temporally aligned exposure records.
Selection, weighting, and changing portfolio composition
A claim-level dataset conditions on an occurrence having been reported. A policy-level population includes zero-claim policies. The expected severity among reported claims is therefore usually unrelated to the unconditional expected cost per policy without additional count and coverage information. Severity samples can also overrepresent portfolios that generate frequent claims. This is a form of size or event-frequency weighting, and it persists even if severity is measured without error.
Consider two policy types with identical exposure duration. Ten type-A policies each have expected claim count ; one type-B policy has expected count . Suppose every claim from either type has expected covered severity $1{,}0006$ and expected aggregate cost $6{,}000545.45$ per policy. An analysis of claim records alone represents type B heavily and cannot recover the policy-average annual cost without denominators and type-specific frequencies. This is not a defect in the claims sample when the estimand is genuinely claim-conditional severity; it is a mismatch when the business question concerns prospective policy cost.
The population itself may change as underwriting, marketing, premium levels, reinsurance participation, or consumer behavior changes. A conditional model can be stable while the portfolio average changes through the distribution . Conversely, a stable overall average may conceal offsetting shifts in risk groups. Therefore, a model's validation population needs to match the prospective decision population at the relevant date, or the difference must be represented through weighting, standardization, or explicit scenarios. These methods require overlap and assumptions; they are not repairs for arbitrary extrapolation into unsupported risk classes.
Frequency, Severity, and Conditional Targets
Frequency and aggregate cost identities
Let be the number of eligible covered occurrences in a future interval and the payment for occurrence under specified contract terms. Define aggregate covered loss as
with the empty sum equal to zero. The general conditional expectation identity is
This expression requires well-defined nonnegative or integrable losses. Under the further assumption that claim severities have a common conditional mean regardless of the realized count and that the relevant expectations are finite, it simplifies to
The familiar frequency-times-severity relationship is thus a conditional identity under stated assumptions, not a universally safe multiplication of independently estimated sample means. Severity may rise with catastrophe-driven counts, and claims occurring together may share inflation, claims-handling dynamics, or limits. Moreover, claim-level severity must use the same contractual payment definition as the aggregate target.
Expected cost per unit exposure, frequently called a pure premium or risk premium in non-life insurance contexts, can be represented by when is a specified fixed exposure and the ratio is appropriate to the exposure base. Aggregate expected total cost for a target portfolio is . Dividing by total exposure yields an exposure-weighted expected rate; averaging risk-specific per-exposure means without weighting yields a different functional.
Zero losses and compound distributions
Insurance aggregate losses frequently have a point mass at zero and a continuous positive component. A compound Poisson–gamma process is one way to represent this structure. Jørgensen and de Souza's insurance application (1994) models count and total claims cost with a compound Poisson process and gamma claim sizes; the family is related to Tweedie exponential-dispersion modeling. The later dispersion-modeling analysis (Smyth and Jørgensen, 2002) emphasizes that the variation of costs can require modeling separately from their mean. Dunn and Smyth's density calculations (2005) further show that likelihood computations for Tweedie families involve nontrivial numerical evaluation.
A direct aggregate-loss model and separate frequency and conditional-severity models may target the same expected cost when they use compatible exposure populations, coverage definitions, development adjustments, and assumptions. They need not give identical estimates with finite data, regularization, different distributional specifications, or misspecification. Goodness of fit to the observed distribution of paid loss may favor a different procedure from prediction of ultimate covered loss. Thus model-family comparison comes after the exact target and label construction are fixed.
A hurdle specification separates the presence of a positive payment from its conditional magnitude. Its zero mechanism may reflect the absence of any event, an uncovered event, a deductible, or a recording threshold. These mechanisms have distinct substantive meanings. A zero-inflated count model similarly requires a defensible source for additional zeros; high predictive accuracy alone cannot prove the structural interpretation. An event can be economically costly and yet produce a zero insurance payment, demonstrating again why target definition precedes distribution choice.
Claim severity, truncation, and line-specific targets
Severity conditioned on having a reported claim is a selected outcome. Incurred-but-not-reported occurrences are absent from immature files, and large complex claims may report or settle later than small claims. Consequently, a severity average measured in the first months of development may be affected by both event heterogeneity and differential reporting. Deductibles, policy limits, salvage, subrogation, and reinsurance recoveries also change the observed severity scale. A severity target should identify whether it is ground-up, gross insured, paid, incurred, ultimate, net of recovery, or discounted.
The arithmetic mean emphasizes expected cost, but solvency, capital, and large-loss management may require quantiles or tail-oriented distributional descriptions. A quantile forecast is not simply an uplifted mean. In heavy-tailed settings, empirical quantiles can be unstable and sensitive to sparse extremes; measurement and sampling decisions should be stated with the target itself. Comparing tail-risk forecasts on truncated policy-limit data requires caution about what lies beyond the contractual observation boundary.
Insurance-line variation is substantive. In annual motor insurance, the target may be covered accident cost per vehicle-year. In property insurance, occurrence and catastrophe aggregation can be essential. In life insurance, survival, lapse, cause of termination, benefit amount, and mortality timing determine the insured event and obligation. In disability, benefit onset, elimination period, duration, and recovery produce several endpoints. In health coverage, utilization and paid amounts depend on benefit design, provider reimbursement, and reporting arrangements. A universal predictive column called “risk” would compress these distinct financial and probabilistic objects.
Prediction Horizon and Observation Window
Information sets and event clocks
A prediction is meaningful only in relation to a time at which it could have been made. Let represent information observable by the insurer or authorized analyst at origin . A forecast should be measurable with respect to . Features created after may be useful in retrospective explanation or claims management at a later date, but their use changes the prediction task. To distinguish tasks, consider policy inception, first notice of loss, claims investigation, a midterm renewal, and a valuation date as separate prediction origins.
Each insurance record can carry multiple clocks: occurrence date, policy effective date, coverage attachment date, notification date, claim creation date, first reserve date, payment date, settlement date, reopening date, accounting recognition date, and database ingestion date. The observation process is not merely a technical field. It determines what the insurer could know when the model is intended to operate. An accident-year claim-cost prediction evaluated in a calendar-year transaction dataset needs an explicit reconciliation of these clocks.
The outcome window also matters. “A claim within twelve months” may mean occurrence within twelve months of inception, report within that period, settlement within that period, or a payment posted by the period end. These are four different event definitions. A model trained on reports within twelve months generally cannot be interpreted as a model of all occurrences within twelve months unless development is addressed. A forecast for the next month and a forecast for the next year similarly involve different event probabilities, seasonality, exposure and portfolio states.
Reporting delays, development, and ultimate outcomes
Let be the number of covered occurrences during accident interval , and let be the subset reported by cutoff . For , ordinarily if one reporting record is used per eligible occurrence. Let denote reporting lag. In a simplified model with independent reporting delays and common reporting probability for the relevant occurrence cohort,
A count divided by then offers a simple development adjustment. Its use requires knowledge or estimation of and appropriate assumptions about the reporting process; it can fail under heterogeneous claim types, nonstationary claims operations, catastrophe delays, fraud detection differences, or changing calendar-time conditions. Avanzi, Wong, and Yang's claim-reporting model (2016) explicitly incorporates risk exposure, overdispersion, and reporting delays when predicting incurred-but-not-reported counts. Its relevance is precisely to the gap between occurrences and visible claim records.
Suppose a fully mature synthetic accident cohort contains 100 eligible occurrences, 60 reported after three months, and 92 reported after twelve months. Three-month and twelve-month reported counts are 60 and 92; the ultimate occurrence count is 100. A model perfectly predicting 60 at three months may be excellent for short-horizon workload while underestimating ultimate counts by 40 percent if its output is treated as ultimate. Dividing 60 by recovers 100 in the constructed cohort only because the eventual count and reporting proportion were stipulated; an actual prospective estimate must infer these quantities, quantify uncertainty, and validate their stability.
The parallel monetary distinction is paid-to-date versus ultimate covered cost. Mack's chain-ladder uncertainty analysis (1993) is concerned with reserve estimates and their standard errors, an inferential task different from predicting the immediate next claims payment. Even a valid aggregate development model may hide how much of uncertainty concerns event occurrence, claim severity, case-reserve updates, litigation duration, or payment scheduling. A claims reserving target therefore includes the cohort, valuation date, development or payment horizon, gross/net basis, and financial definition.
Financial reporting and actuarial forecasting
Financial accounting is a further measurement domain. The IFRS 17 insurance-contract framework uses specified measurement, recognition, and presentation principles for groups of insurance contracts. The IFRS Foundation's definitions of coverage and incurred-claim liabilities distinguish the liability for remaining coverage from the liability for incurred claims, and describe fulfillment cash flows. These accounting categories are neither synonyms for a machine-learning response column nor unrestricted descriptions of expected future accident losses. A probabilistic cash-flow model can inform financial measurement, but its target must be reconciled to the applicable accounting requirements and reporting perimeter.
An important international distinction follows: actuarial standards issued in the United States, European supervisory discussion, and IFRS financial reporting have different institutions, scopes, and legal implementations. An international scientific monograph can compare their treatment of intended measures and governance without claiming that a US ASOP or an IFRS standard governs every insurer in every jurisdiction.
Validation in time and across portfolios
Random allocation of rows to training and test datasets is often misleading when the same policy, claim, catastrophe, household, or development cohort can appear in both. Leakage also arises when a model uses outcomes that could not yet have matured at the intended forecast date. A credible temporal test reproduces the actual decision sequence: choose an origin, build features available by that origin, fit on earlier information, and evaluate later predictions only when their target outcomes can be observed or appropriately estimated.
Kapoor and Narayanan's leakage investigation (2023) shows why test-set construction and feature legitimacy are scientific issues rather than mere software implementation details. In insurance, validation also requires assigning all records from the same underlying occurrence consistently to training or evaluation, identifying claims development and closed/open status, and recognizing that the future target population may differ from the historical underwriting portfolio.
A useful validation record names the training and testing origination dates, the latest event and processing timestamps allowed for each feature, the maturity of each outcome, the group-split rule, the portfolio scope, and the measure of performance. Holdout predictions should remain reproducible from these declared inputs. When an insurer changes product terms, administrative systems, underwriting selection, or claim management, the change becomes part of the prospective transport problem.
Target Leakage and Misdefinition
Structural and temporal sources of leakage
Leakage occurs when the modeling or evaluation process uses information that would not legitimately be available for the specified task or when its test construction provides an unrealistically favorable assessment. A claim classifier built for first notice of loss cannot use final settlement code; a renewal score cannot use next year's premium transactions; an application-stage underwriting model cannot use a post-inception telematics measure. These features may be valid for separate later-stage predictions, but they correspond to different information sets.
The published taxonomy of leakage failures (Kapoor and Narayanan, 2023) provides general methods for examining such problems. Insurance adds a particular complication: repeated rows can share hidden event identities, and the information date of a field can differ from the event date it describes. A retrospective backfill may cause a past-looking feature to contain later information. To prevent this form of leakage, the analyst must reconstruct point-in-time availability, not merely compare calendar dates in the final database.
Target misdefinition can coexist with technically leakage-free validation. Suppose a model is evaluated fairly against paid amounts twelve months after occurrence, and performs well, but the decision requires ultimate insured losses. No improper feature was used; the target itself is misaligned. Conversely, a severity model trained only on settled claims may display low error for those claims while systematically excluding large pending liabilities. Its metric may be valid for a selected population and misleading for the intended portfolio.
Other structural hazards include zero-claim rows removed during data preparation, policy cancellation treated as a full-year exposure, claims divided by different indemnity definitions across years, losses classified by accounting date while denominators use underwriting date, and net-of-reinsurance payments compared with gross loss-based premiums. These are scientific mismatches before they are statistical fitting problems.
Predictive association and causal intervention
Forecasting describes an outcome under the data-generating and decision regime represented by the available information. Intervention asks how outcomes would change if a policy, price, loss-control treatment, or claim-management action were altered. The predictive probability need not equal either potential-outcome probability under intervention , and the difference between two observed risk groups is not generally the effect of assigning one intervention.
Hernán and Robins (2020) develop the role of explicit causal questions, exchangeability, positivity, and identification. When applying this reasoning to insurance, prospective behavior and insurer action can change the composition of insured risks: raising a premium might influence renewal selection; offering a deductible might affect insured reporting; risk controls may change event probability. Thus the prediction of future claims under one historical underwriting regime cannot by itself establish the consequences of another regime.
A direct comparison also clarifies the term estimand. The clinical ICH E9(R1) framework formally describes treatment-effect estimands under explicit alternatives and intercurrent-event strategies. Insurance studies may have causal estimands when comparing interventions, but claims frequency, expected loss, and individual predictive distributions are ordinary descriptive or predictive targets unless a causal question and identification design are explicitly supplied. Using the vocabulary precisely prevents an otherwise serious category error.
Data quality and professional-standard controls
Data quality is not exhausted by completeness or technical correctness of database fields. For statistical inference, it means that the selected records, definitions, and transformations are fit for the target and that their limitations are communicated. United States ASOP No. 23 addresses actuarial practice in selecting, reviewing, using, and relying on data, including appropriate disclosures. The official ASOP No. 56 adoption notice identifies professional modeling guidance covering model design, development, selection, modification, use, review, and evaluation. These provide jurisdiction-scoped professional reference points rather than worldwide guarantees of model validity.
Insurance governance additionally considers what a highly granular target accomplishes and whose data or circumstances it uses. The EIOPA report on insurance AI governance (2021) discusses applications in risk assessment, pricing, claims management, and fraud prevention. The statistical object must therefore be described together with its legitimate use, privacy context, discrimination and proxy concerns, and the distinction between decision support and autonomous action. A model can be statistically calibrated and still be subject to independent contractual, legal, and governance restrictions. These restrictions vary by jurisdiction and use case.
Estimand-to-Decision Verification
A complete research specification
Before comparing algorithms, an insurance statistical investigation should write a complete ordinary-language statement of its objective. For example: “For all private motor policies eligible under product version C and commencing during the next underwriting quarter, estimate at policy inception the distribution of gross covered claims occurrence costs over the following twelve months per earned vehicle-year, including eligible occurrences reported subsequently, under the current contract and claims-handling regime.” The statement is intentionally specific. It identifies population, contract, prediction origin, event window, basis of monetary measurement, denominator, maturity requirement, and operational regime. A different sentence is needed for cash paid within the year, immediate claims workload, loss ratio, insurer net retention, or the impact of a safety intervention.
The mathematical stage then assigns a well-defined random variable or functional to that statement. In case of policy-level ultimate cost , the target could be , or the full conditional distribution . For a portfolio with exposures , a premium-relevant target may be . The unit weights must match the decision. Claim-conditional severity requires a separate conditional distribution and a specified selection process.
The empirical stage documents the available observed label, the transformation from recorded data to intended endpoint, the maturity and exclusion rules, and any assumed missingness, truncation, or censoring mechanism. Where an identification gap remains, the study states what assumptions close it, what can be learned under weaker assumptions, and which sensitivity analyses change the target or merely change the estimator. This distinction is important: a sensitivity analysis that silently changes the outcome from ultimate to paid loss is a comparison of estimands, not robustness of one estimand.
Loss functions, calibration, and model comparison
A forecast should be evaluated against its stated target with an appropriate and justified criterion. For a conditional mean , mean squared error emphasizes larger mistakes, while mean absolute error elicits a median in the usual continuous setting. For an event probability , the Brier score and logarithmic score assess probability predictions. For a complete predictive distribution, the proper scoring-rule analysis of Gneiting and Raftery (2007) covers coherent probabilistic evaluation, including continuous ranked probability scores and interval scoring. The loss function must match the financial decision and the target type; a point-loss comparison alone cannot establish calibration or tail adequacy.
Calibration asks whether reported risks correspond to observed frequencies or distributions in an appropriate target population. Van Calster and colleagues' calibration analysis (2019) underscores the scientific importance of calibration for predictive models. In insurance, calibration may need to be examined at several levels: probability of a claim, expected claim count, expected aggregate cost, claim-size bands, contract classes, development maturity, and geography. Calibration of an overall portfolio total does not ensure correct allocation across materially different groups.
Event rarity complicates comparisons. The precision–recall analysis of Saito and Rehmsmeier (2015) illustrates the usefulness of precision–recall summaries for strongly imbalanced binary outcomes; precision itself varies with positive-class prevalence. Accordingly, a fraud or rare large-loss classifier's performance cannot be transferred to another portfolio without attention to base rates, thresholds, selection, and costs. This is a different decision task from pricing an expected claim amount, and its favorable discrimination does not prove a correct premium target.
A transparent model report should show performance against a simple baseline, temporal out-of-sample tests, appropriate uncertainty, groupwise calibration, and the decision-weighted consequences of errors. TRIPOD+AI's prediction-reporting guidance (2024) is written for clinical prediction studies; its systematic reporting discipline is useful as an external methodological comparison, not as an insurance-specific regulatory requirement. Reporting should make the target recoverable from the published text without relying on the private variable names of a dataset.
Worked case: identical pure premium, different risk structures
Consider two synthetic annual insurance portfolios with identical covered-loss definitions and no expense or discount adjustments.
| Quantity | Portfolio A | Portfolio B |
|---|---|---|
| Earned policy-years | 100 | 200 |
| Covered claim occurrences | 20 | 20 |
| Aggregate covered payment | $100,000 | $200,000 |
| Observed occurrences per policy-year | 0.20 | 0.10 |
| Observed mean covered severity | $5,000 | $10,000 |
| Observed covered payment per policy-year | $1,000 | $1,000 |
Every entry follows by direct arithmetic: $20/100=0.2, \20/200=0.1, \100{,}000/20=5{,}000, \200{,}000/20=10{,}000, and both aggregate costs divided by earned exposure equal \1{,}000$. These are sample statistics, not inferred population parameters. The example demonstrates that equal empirical pure premiums can arise from different count and severity combinations. A procedure that is adequate for comparing mean cost may be inadequate for estimating claims workload, typical claim size, or the probability of a positive payment. The equality also gives no information about variance or upper-tail risk without further assumptions.
Worked case: paid-to-date as a selected forecast target
Take a second synthetic cohort with ultimate covered cost $1{,}000{,}000. Assume cumulative paid costs of \400{,}000$ at one year and $850{,}000$ at three years. These are stipulated hypothetical facts. A model that perfectly forecasts the one-year payment meets a liquidity-timing target of $400{,}000. If that output is labeled as ultimate covered loss, the understatement is \600{,}000$ or 60 percent of the stipulated ultimate. Likewise, at three years, $850{,}000$ still understates ultimate by $150{,}000.
The case isolates target mismatch from statistical estimation error. A reserve may be concerned with unpaid future cost at a valuation date; a cash forecast concerns transactions during a future calendar interval; and ultimate loss concerns cumulative covered obligation on a specified basis. This is why reserving literature, including Mack's reserve standard-error work, is connected to but conceptually distinct from claims payment forecasting.
Worked case: a prediction threshold and an intervention
Suppose a purely illustrative action costs and would prevent an otherwise covered $5{,}000$ loss whenever that loss occurs. Under risk neutrality, an assumed complete preventive effect, and no other consequences, the expected net benefit is $5{,}000p-$200, where is the event probability in the relevant population. The threshold for positive benefit is . This calculation is arithmetic, not a recommended insurance practice or an empirically supported preventive effect. If the action only prevents 50 percent of the loss, the threshold becomes under otherwise identical assumptions. A model's event probability may help estimate the relevant baseline risk, but the assumed effect of action requires separate causal evidence.
The threshold demonstrates the general dependency of insurance decisions on target and action definitions. If a model predicts claims reporting rather than covered occurrences, or reflects a selection-changing underwriting regime, its output can be calibrated to its original labels and still misinform the specified intervention. The appropriate investigation estimates both the predictive target and the counterfactual effect needed for the decision, rather than presenting their product as a directly observed actuarial result.
International applications and research priorities
International comparability requires mapping local insurance products, claim rules, fiscal periods, currency and discount bases, exposure measures, administrative procedures, and regulatory contexts. A motor-claims frequency model from one country cannot be transported by translating the predictor names alone; claim-reporting incentives, court systems, policy limits, covered risks, and insurer selection may change the underlying response. Life and health portfolios add competing events, service definitions, and benefit eligibility. Catastrophe and cyber exposure add correlated occurrences and delayed information. These variations require stated transport assumptions and targeted external validation.
At a minimum, future empirical investigations should test whether point-in-time feature reconstruction changes performance; whether incurred and paid outcome definitions lead to different portfolio rankings; how heterogeneous reporting delays influence calibration; whether differences in underwriting population alter exposure-weighted targets; and how much of a model comparison is explained by outcome-definition differences rather than algorithmic differences. Each question can be investigated with historical snapshots, matched exposure cohorts, transparent outcome definitions, and sufficiently mature evaluation periods. Sensitivity to missing information, claims reclassification, and contractual revision deserves explicit reporting.
The original synthesis of this dossier is an interpretive result: the target's semantic identity is maintained only when the population, event, contract, temporal origin, observation process, functional, and decision remain aligned. None of these elements is reducible to a generic machine-learning metric. An evaluation that holds the algorithm fixed while changing any of them is a new scientific question. Conversely, comparing multiple estimators for one unchanged target can reveal genuine differences in predictive accuracy and uncertainty when their validation conditions are the same.
Conclusion
Insurance statistical estimands and prediction targets determine which economic, contractual, probabilistic, and administrative quantity a model is authorized to estimate. A policy-level probability, claim count, claim-conditioned severity, pure premium, ultimate loss, liability-related cash-flow projection, and payment at a fixed maturity each have different meanings and identifying assumptions. The same raw insurer database can support several such targets; the same numerical model output may be correct for one and inappropriate for another.
The mathematical foundations establish three especially consequential results. First, frequency multiplied by severity is an expected aggregate-cost identity only under a compatible conditional construction, with explicit assumptions regarding counts and severities. Second, insured-event occurrence, reported claims, and paid transactions are different observables whose divergence may persist long after the occurrence date. Third, statistically valid prediction under an existing policy regime does not, by itself, identify the consequence of intervening on that regime. Each result is grounded in established probability, actuarial research, or causal inference, and the novel work here lies in their joint analytical reconciliation for insurance targets.
The corresponding scientific practice begins with an explicit target sentence and a matching mathematical functional. It then reconstructs contractual coverage, exposure, information timing, administrative measurement, and target-population selection. Validation must use admissible features and outcomes with sufficient maturity, choose proper criteria, examine calibration and distributional risks, and state the intended decision and institutional scope. Accurate algorithms become scientifically meaningful when their predicted object remains stable throughout this chain.
This monograph provides a completed conceptual, methodological, and mathematical research assessment of that chain. Its examples are self-contained demonstrations, and its sources are transparent. Specific policy pricing, regulated reserving, financial reporting, and intervention decisions remain subject to their respective legal contexts, technical assumptions, and actual data.
References
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revised December 2016; effective April 30, 2017). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention (effective August 1, 2018). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Actuarial Standards Board. (2019). ASB Adopts ASOP No. 56: Modeling (official notice; standard effective for work commenced on or after October 1, 2020). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asb-adopts-asop-no-56/
Avanzi, B., Wong, B., & Yang, X. (2016). A micro-level claim count model with overdispersion and reporting delays. Insurance: Mathematics and Economics, 71, 1–14. https://www.sciencedirect.com/science/article/pii/S0167668716300208
Dunn, P. K., & Smyth, G. K. (2005). Series evaluation of Tweedie exponential dispersion model densities. Statistics and Computing, 15(4), 267–280. https://doi.org/10.1007/s11222-005-4070-y
European Insurance and Occupational Pensions Authority. (2021). Artificial intelligence governance principles: Towards ethical and trustworthy artificial intelligence in the European insurance sector. EIOPA. https://www.eiopa.europa.eu/publications/artificial-intelligence-governance-principles-towards-ethical-and-trustworthy-artificial_en?prefLang=ro
Goldburd, M., Khare, A., Tevet, D., & Guller, D. (2025). Generalized Linear Models for Insurance Rating (2nd ed., 2025 revision; CAS Monograph No. 5). Casualty Actuarial Society. ISBN 978-1-7333294-4-6. https://www.casact.org/monograph/cas-monograph-no-5
Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437
Hernán, M. A., & Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC; freely maintained author edition. https://miguelhernan.org/whatifbook
IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts. IFRS Foundation, standard and implementation overview (standard effective for annual reporting periods beginning on or after January 1, 2023). https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts—Key Terms. IFRS Foundation, implementation resource. https://www.ifrs.org/supporting-implementation/supporting-materials-by-ifrs-standards/ifrs-17/key-terms/
International Council for Harmonisation. (2019). Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials: ICH E9(R1) (final version adopted November 20, 2019). ICH. https://database.ich.org/sites/default/files/E9-R1_Step4_Guideline_2019_1203.pdf
Jørgensen, B., & de Souza, M. C. P. (1994). Fitting Tweedie's compound Poisson model to insurance claims data. Scandinavian Actuarial Journal, 1994(1), 69–93. https://doi.org/10.1080/03461238.1994.10413930
Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. https://doi.org/10.1016/j.patter.2023.100804
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://doi.org/10.2143/AST.23.2.2005092
Saito, T., & Rehmsmeier, M. (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE, 10(3), e0118432. https://doi.org/10.1371/journal.pone.0118432
Smyth, G. K., & Jørgensen, B. (2002). Fitting Tweedie's compound Poisson model to insurance claims data: Dispersion modelling. ASTIN Bulletin, 32(1), 143–157. 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
Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L., & Steyerberg, E. W. (2019). Calibration: The Achilles heel of predictive analytics. BMC Medicine, 17, 230. https://doi.org/10.1186/s12916-019-1466-7
Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378