Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Underwriting Exposure Reconstruction under Sparse Information

Abstract. Underwriting exposure reconstruction estimates the relevant physical, economic, biometric, operational, and contractual characteristics of a proposed insured risk using partial submissions, external observations, historical information, and expert knowledge. This monograph examines the conditions under which such reconstruction supports scientifically defensible insurance decisions. It distinguishes the underlying exposure from its documentary representations, the loss process, the insurance contract, and the insurer's decision. Bayesian conditioning is analyzed alongside identification bounds, missing-data mechanisms, hierarchical pooling, expert elicitation, scenario enumeration, and predictive validation. A series of synthetic insurance examples demonstrates how coherent estimates can nevertheless produce unstable underwriting decisions when missingness, measurement error, severity, common-factor dependence, or contract terms remain uncertain. International supervisory material, American actuarial standards, New York regulatory expectations, Lloyd's market oversight, and European supervisory guidance supply a comparative institutional context. The principal analytical finding is that sufficiency attaches to the proposed decision and its material consequences: a reconstructed exposure may be adequate for preliminary classification yet inadequate for binding coverage or accumulation control. Referral becomes rational when feasible exposure states change the admissible action or materially affect retained loss. The study relies on accessible institutional texts, narrowly interpreted scholarly records, and reproducible hypothetical calculations. It does not estimate insurer-specific performance or claim a new statistical theory.

Keywords: Underwriting; exposure reconstruction; sparse information; missing data; partial identification; Bayesian inference; expert elicitation; proxy validity; selection bias; insurance risk; decision uncertainty; referral; exposure accumulation.

Introduction

The scientific object

An underwriting submission is a dated communication concerning a potential transfer of risk. Its contents may include the characteristics of the applicant, the item or activity proposed for coverage, past losses, estimates of future activity, and requested terms. The object of exposure reconstruction is the underlying configuration of risk-bearing quantities and characteristics that these communications imperfectly describe. Reconstruction connects a documentary record with a prospective insurance problem. It is a problem of inference because the relevant configuration is only partly observable at the decision date.

The term exposure has several operational meanings. In frequency modeling it can denote an amount of time or activity at risk. In property insurance it can denote an insured value, a site, or a concentration of susceptible assets. In liability underwriting it can denote activities or relationships capable of creating a future obligation. In life insurance it includes the amount and duration of benefits contingent on survival or death, together with relevant information about the applicable mortality process. These meanings share the idea of a risk-bearing relationship, although their units, mechanisms, and loss functions differ. Converting all of them into monetary amounts too early obscures the mechanisms by which the insurer becomes liable.

The reconstruction target is consequently a vector rather than a universal score. For a commercial location it might contain coordinates, floor area, replacement cost per unit, construction, occupancy, protection, downtime dependency, and a timestamp. A policy representation adds insured interests, attachment points, limits, exclusions, covered causes, and the relationship between the gross event loss and the insurer's payment. Some elements are directly observed, others are constrained, and still others are represented by distributions. A portfolio representation adds the identities of related insureds, geographical clusters, common-service dependencies, and the terms of reinsurance.

There is an important temporal direction to the analysis. A decision at time t concerns a coverage interval that may start later and span several months or years. The relevant exposure is E(t) or an exposure process over the interval, rather than merely the value shown in the last available source. A building can be expanded, a fleet can change utilization, a business can change suppliers, and a person's coverage may be renewed under a different institutional arrangement. The same recorded variable may therefore be suitable for a retrospective claim reconstruction and unsuitable for prospective selection.

The institutional relevance is visible in IAIS (2024): its ICP and ComFrame online text identifies underwriting acceptance procedures and contract conditions as possible components of underwriting policy under ICP 16.7.1. The technical material in this study develops the corresponding inferential question of what an insurer can reasonably know about a proposed risk at acceptance. The international supervisory framework is distinct from the domestic rules that determine the legal permissibility of particular decisions.

Research questions and method

Three questions govern the investigation. When does available information suffice for a particular underwriting decision? How can the effect of assumptions and scenarios on exposure be represented to the decision-maker? Which risk configurations require further investigation instead of routine acceptance based on a point estimate?

The methodology is a critical synthesis of missing-data inference, actuarial modeling, risk classification and the institutional governance of underwriting. The mathematical examples are deduced from stated assumptions. Where computation is useful, finite scenarios are enumerated exactly, enabling the reader to reproduce the results without access to private data. The examples are analytical experiments: all input values, rates, event probabilities, thresholds and costs are hypothetical. Alternative assumptions are tested within each example. The normative analysis is comparative rather than a comprehensive survey of national insurance law.

The relevant units of observation are explicitly distinguished. A proposal is a pre-contractual record. A decision is an underwriting action taken under a specific information set. A bound policy is an insurance arrangement conditional on acceptance and agreement. A policy-year describes time under risk, subject to coverage changes. An underlying person, asset or enterprise can appear in multiple proposals or policies. A claim is an observed report or contractual payment arising under the applicable coverage. A portfolio is a set of risk-bearing contracts, with dependence and duplication explicitly considered. Mixing these statistical units can produce an apparently large dataset whose effective volume of independent information is small.

The scope is exposure inference at underwriting and its consequences for decisions and validation. It connects to credibility theory, pricing, reserving, contract interpretation, catastrophe modeling and insurance conduct without treating any one of those fields as an interchangeable substitute. A premium is not a measurement of exposure; a reserve is not an observation of the proposed risk; and an applicant's assertion does not become independently verified merely because it has been normalized in an information system.

Sparse Submission Contexts

Why insurance information is sparse

A submission can be incomplete even when the applicant has answered every question asked. A form designed for standardized risks may fail to elicit the most important characteristic of a complex risk. An intermediary may know that a warehouse stores commodities without having reliable information on the building protection system. A business may provide financial statements from the prior year while seeking coverage for an expansion. A small insured may have no structured historical loss records because it previously retained losses. The informational gaps reflect measurement cost, institutional procedures, market competition, technology, privacy restrictions, and the novelty of the insured activity.

Absence, coarseness and inconsistency represent distinct problems. An absent field leaves its value unknown. A coarse field constrains a quantity to a category or interval. An inconsistent field provides conflicting information, potentially with different source quality or observation dates. An obsolete field has a value that may have been accurate in a former period. A censored report reveals only inequalities, such as a loss exceeding a retention. A structurally inapplicable field expresses a different data state from a nonresponse. A transformation that maps all these conditions to zero destroys information.

For an illustrative commercial submission, let turnover be reported as between $1 million and $2 million, employee count be absent, location be reliably known, and a safety certification be dated three years before inception. The location may be immediately usable for a hazard query. Turnover enters the analysis as an interval. The employee count requires a relationship with business activities and payroll, with uncertainty. The certification requires a decision about continuing validity. Treating all four variables as equally observed would misstate the condition of the evidence.

The cost of information is itself part of the institutional environment. A low-value routine risk can be uneconomic to inspect individually. A single large location, however, may dominate a portfolio's marginal event loss. For it, even expensive verification can be justified. The resulting information systems are necessarily selective: what is measured depends on the expected economic and protective value of measurement. This endogeneity creates a connection between statistical missingness and underwriting procedure.

The Actuarial Standards Board's ASOP No. 23, Data Quality (2016) provides professional context for the assessment and use of data in actuarial work. The relevance to reconstruction is that source suitability depends on the intended analysis. A data warehouse can be operationally successful while its fields remain inadequate for a particular risk inference. The quality of an estimate must be assessed in relation to the decision quantity, observation process, and materiality of resulting errors.

Line-specific scarcity

In property insurance the principal uncertainty may concern the location, size, construction, use, condition, and economic replacement value of buildings and contents. A reliably geocoded address can be linked to a hazard layer but cannot by itself establish construction, basement elevation, fire protection, tenancy, or policy coverage. Even a precise structure footprint can mislead where gross floor area depends on several stories or includes space outside the relevant insured interest.

Business interruption introduces another exposure dimension. Revenue indicates a scale of activity; its relationship to the insured financial loss depends on variable costs, insured accounting definitions, interruption duration, machinery replacement time, alternative production, and supplier relationships. The reconstruction of property value and the reconstruction of interruption exposure are related tasks with different target quantities.

For automobile insurance a vehicle-year is one possible measure of exposure to frequency. Actual driving distance, time, route, driving environment and utilization can be relevant explanatory characteristics, subject to product and legal context. A vehicle may be registered yet driven infrequently. A vehicle may have multiple authorized users. A policy can terminate before the nominal year. A count of vehicles is a useful inventory measure but is insufficient for a precise mileage-based risk estimate.

For casualty insurance the underlying exposure may be an ongoing set of activities, contractual obligations, third-party interfaces or products in circulation. Payroll, sales and headcount each represent partial measurements. Changes in subcontracting can modify the risk configuration without a comparable movement in employee count. Claims arising under prior operations may be reported much later. A lack of reported losses during a short observation window supplies limited evidence for low future liability risk.

Life insurance offers a distinct reconstruction context involving personal information and mortality-related assessment. The NAIC's Accelerated Underwriting overview describes the adoption of guidance on external data and analytics in 2024 and subsequent oversight work. An accelerated process may economize on conventional evidence collection, but it does not eliminate the requirements of relevant data, regulatory oversight, and an appropriate path for cases that require more assessment. The analytical questions include the reliability of the alternative information and its ability to support the intended risk classification. They do not license direct inference of sensitive traits from arbitrary correlations.

Cyber insurance illustrates information that changes quickly. A point-in-time observation of an externally visible service can establish limited facts about visibility under the scanner's conditions. It cannot determine the entirety of an enterprise's internal configuration, recovery capability, privileged access, dependency architecture or future patching behavior. A cybersecurity questionnaire and an external observation may be useful together, provided their scope, time and error structure are documented. A missing signal can also reflect incomplete scanner coverage rather than the absence of exposure.

Provenance and observation states

An evidence record for a reconstructed feature should associate the source with the actual concept it measures. Consider the statement that a building has a value of $4 million. Is the value a purchase price, a depreciated accounting value, an appraisal, a replacement-cost estimate, a declared insured value or a limit? Each has a distinct interpretation. The unit, currency, valuation date, legal interest and source methodology determine whether the observation can enter a particular equation.

A practical ledger separates declared, externally observed, derived, inferred, conflicting and unavailable information. These labels describe evidence status rather than the inherent quality of the value. A declared recent payroll statement can be more reliable than an inferred value from an outdated third-party source; a direct physical inspection can still omit an inaccessible portion of the relevant premises. Provenance thus informs the error model and the scope of validation.

The missingness indicator is part of the record. A failed API lookup, a field withheld under applicable privacy rules, an unasked question, and a refusal to answer should be preserved as different mechanisms. In a subsequent audit these categories make it possible to identify whether sparse information is concentrated within a channel, location, product, cohort or application stage. They also permit interventions directed at the actual source of missing information, rather than indiscriminately requiring every applicant to provide more data.

Observation of losses

Past losses offer a further source of information but are often sparse in precisely the cases for which their statistical value would be most useful. Let N denote a count of reported events within a coverage interval and T its duration. Even a perfectly observed count cannot be interpreted without an exposure measure. If a report includes only payments greater than a deductible, it describes a truncated view of an underlying event process. If the reporting period is incomplete, recent experience is subject to lag.

Under a homogeneous Poisson illustration with event intensity lambda per exposure unit, the probability of observing no events over total exposure e is exp(-lambda e). At lambda = 0.1 and e = 1, the probability remains approximately 0.905. Over twenty equivalent independent exposure units it falls to approximately 0.135. The absence of reported events at one exposure unit therefore supplies little evidence of a vanishing intensity. Where the recorded event threshold or reporting probability changes, even this elementary calculation requires revision.

The proposal-level reconstruction must distinguish claim count and covered event count from the broader phenomenon of damage. Deductibles, noninsurance, retention, changes in reporting practices, and operational changes all affect the observed record. Treating a short clean history as a directly measured favorable hazard characteristic creates a false appearance of certainty. Conversely, a large historical event is evidence about risk but does not by itself establish a stationary future hazard.

Reconstruction Assumptions

A measurement model for the exposure

Let X be the vector of relevant exposure characteristics at the decision date; D denotes the observed submission; Z denotes external observations; M denotes the missingness pattern; and C denotes documented context, including class and time. A Bayesian reconstruction specifies a prior density p(X | C) and a measurement and observation model p(D, Z, M | X, C). Conditional on these elements,

p(X | D, Z, M, C) is proportional to p(D, Z, M | X, C) p(X | C).

The formula is ordinary Bayes conditioning. Its substantive content is the definition of X, the reliability of D and Z, and the missingness process M. If the same original declaration feeds three commercial databases, those observations cannot be treated as three independent measurements merely because they arrive in separate interfaces.

Under additional conditional independence assumptions, the likelihood can be factored into source-specific components. Such factorization is useful when data come from genuinely different measurement mechanisms. It becomes dangerous where sources are copied, share common geocoding errors, inherit the same classification rule, or use an identical upstream estimate. Dependence should be explicitly modeled using a shared latent error, or the apparently redundant sources should be collapsed into their common evidence origin.

An observation model can include both systematic and random errors. For a positive quantity such as replacement cost, a multiplicative error representation can be more plausible than an additive unconstrained Gaussian one. On the logarithmic scale a source bias b and noise epsilon may produce log Z = log X + b + epsilon. This expresses relative error, allows adjustment for systematic source differences, and avoids negative estimates after retransformation. The bias parameters require their own evidence and sensitivity analysis; writing them explicitly does not make them identified.

Ignorable missingness and informative omissions

Rubin's (1976) Inference and missing data establishes conditions, stated in its accessible abstract, under which the missingness process can be set aside in direct-likelihood or Bayesian inference. Missingness at random is a condition defined relative to the variables observed and the inferential model, accompanied by distinctness of the relevant parameters; it is not a general description of data that look irregular.

An insurance example clarifies the difference. Suppose a digital form randomly fails to record the last field because of a server fault independent of risk and observed features. Conditional on the relevant circumstances, there is a plausible ignorable mechanism. Suppose instead that the absence of a roof inspection arises because observed addresses in remote locations are expensive to visit. Conditional on variables capturing remoteness and inspection rules, the problem might be adequately represented as missing at random. Suppose finally that inspections are selectively suppressed where an unobserved defect is suspected but never recorded. The missingness depends on the unobserved characteristic; treating it as ignorable can produce biased reconstruction.

These scenarios are conceptual, not empirical claims about the prevalence of any mechanism. Insurance applications frequently combine administrative and behavioral sources of omission. The scientific response is to identify defensible mechanisms, investigate their observable signatures and assess alternative specifications. Missingness at random is an assumption for analysis. It cannot ordinarily be certified solely by observing the completed parts of the data.

Suppose theta is the prevalence of a risk-relevant condition among 100 comparable submissions. Sixty supply an observation, of which 12 exhibit the condition. Forty do not supply an observation. If the observed cases are representative for the purpose, the empirical prevalence among respondents is 0.2. If the missing cases have prevalence 0.5, the combined prevalence is 0.32. If their prevalence is 0.05, it is 0.14. The same observed records are compatible with all three configurations. The discrepancy affects portfolio inference even if the data-entry system has perfect accuracy for the completed fields.

Priors and exposure-specific credibility

A prior can express a population distribution conditional on product, measurement regime and risk class. It can also encode engineering constraints, documented class characteristics, or explicitly elicited judgments. A class distribution should have a meaningful relationship to the target risk, including geographical, temporal and contractual comparability. A prior informed by insured commercial buildings in one market may be inappropriate for informal structures or another jurisdiction.

Hierarchical pooling is useful where individual observations are sparse and related classes have data. For example, let the claim intensity of an operation have a Gamma distribution with shape a and rate b, conditional on class. With a Poisson count n observed over exposure e under a stationary model, the posterior Gamma distribution has shape a + n and rate b + e. Its mean is (a+n)/(b+e). This quantity can also be expressed as a weighted average of the prior mean a/b and an observed rate n/e for e > 0, with weights b/(b+e) and e/(b+e). The weight emerges from the model; it is not a universal credibility constant.

Take a = 2 and b = 20, so the prior mean is 0.10 per exposure unit. With n = 0 and e = 1, the posterior mean becomes 2/21, about 0.09524. With n = 0 and e = 20, it becomes 0.05. With n = 3 and e = 10, it becomes 5/30, about 0.16667. These values demonstrate the difference between an individual observed rate and a regularized estimate. They remain conditional on the exchangeability, stationarity and observation assumptions of the example.

For future exposure e*, the Gamma–Poisson predictive mean is e*(a+n)/(b+e). Its variance includes both expected Poisson process variance and uncertainty in the intensity. Under the same shape A = a+n and rate B = b+e, predictive variance is eA/B + e^2 A/B^2. For A = 2, B = 21, and e* = 10, the predictive mean is approximately 0.95238 and variance approximately 1.40590. Ignoring the second term would systematically omit parameter uncertainty within this hypothetical model.

The prior should not become a concealed underwriting rule. Its source, calibration date, effective information strength and sensitivity matter. When a posterior is strongly influenced by the prior, the output may remain useful for portfolio-level estimates or preliminary triage while requiring corroboration for a high-consequence individual decision. The appropriate interpretation depends on whether relevant uncertainty survives in the action.

Expert knowledge elicitation

Some exposure variables lack sufficient comparable records. Examples include prospective recovery duration after a novel equipment failure, an unusual supply-chain dependency, and the effectiveness of an uncommon protection configuration. Expert knowledge can constrain plausible values. O’Hagan (2019) describes, in his directly accessible author abstract, elicitation of subjective probability distributions through structured procedures. The methodological implication developed here is that an expert estimate should be represented in a manner that preserves the conditions and uncertainty under which it was made.

An elicitation question should refer to a defined quantity, population and time horizon. Asking how safe a factory is creates an undefined cognitive task. Asking for upper and lower plausible values of time required to replace a specified unavailable machine after a particular type of insured damage is more interpretable. If the expert's quantiles describe repair time conditional on parts availability, they cannot be transferred unchanged to a scenario in which the supplier has ceased operation.

The analyst can obtain a range, a central quantile, additional tail quantiles and an explanation of the relevant mechanisms. Internal coherence requires ordering of quantiles and consistency with physical constraints. It also requires distinguishing uncertainty about whether an event occurs from uncertainty about its consequence conditional on occurrence. A probability of 0.05 for an event and a conditional severe loss of $5 million describe different aspects of risk; combining them into a single vague estimate impairs both reconstruction and governance.

Multiple experts may disagree because they use different information or models. The disagreement itself is informative. A pooled distribution should not erase an alternative interpretation capable of changing the decision. If two mechanically competent experts interpret fire compartmentation differently, the useful next step may be to specify which architectural fact would resolve their disagreement. A numerical average of opinions cannot substitute for verification of a decisive physical condition.

Constraints and internal consistency

A reconstructed exposure vector must respect physical and contractual constraints. Building floor area cannot be negative. A location should be associated with the correct insured interest. A schedule of declared values should reconcile with its location-level components to the degree stated by the submission. A policy period cannot generate earned time outside its effective interval. A deductible and a policy limit must be represented consistently with the actual contractual definitions.

These constraints may identify values more effectively than a generic predictive model. If a verified total replacement value is $12 million, and the values of two of three locations are separately measured as $3 million and $4 million on the same basis and date, the third is constrained to $5 million before allowing measurement error. If the total includes business interruption and the components are property only, subtraction is inappropriate. Accounting consistency depends on concept consistency.

Consider two measurements of gross floor area, 2,000 and 2,400 square meters. If both are independent unbiased measurements with equal variance, averaging produces 2,200 square meters with a smaller measurement variance. If one is a building footprint and the other is gross floor area, no such averaging is warranted. A scientifically meaningful reconciliation may instead establish that some floors or outbuildings are omitted from the first source.

The ASOP No. 56 modeling standard (Actuarial Standards Board, 2019), particularly sections 3.1.5 and 3.1.6, addresses the suitability of data and assumptions for the model's purpose. Its relevance is particularly pronounced when the reconstruction is an intermediate model whose output feeds risk classification, pricing, or portfolio monitoring. Input ambiguity can be amplified through multiple downstream models.

External Proxies and Validity

The difference between target, proxy and decision variable

A proxy is an observed characteristic used to infer another characteristic or a decision quantity that is not directly available. Its validity is relational. A satellite image may inform roof geometry; a business registry may inform the classification of operations; a historical financial statement may inform the scale of activities. The proxy's value depends on the target, its measurement error, and the target population.

There are at least three questions. Does the proxy measure or predict the intended underlying characteristic? Does the inferred characteristic have material relevance for the expected outcome or risk-bearing contract? Can the insurer lawfully and appropriately use this particular information for the proposed purpose? A positive answer to one question does not resolve the other two.

The ASOP No. 12 risk classification standard (Actuarial Standards Board, 2005, sections 3.2.1–3.2.4) addresses outcome relationships, interplay among characteristics, objectivity, and practical considerations. This provides a professional frame for proxy assessment. A proxy that accurately identifies a characteristic can still be irrelevant for a particular coverage. Conversely, a variable with an apparent outcome relationship may have poor measurement stability or problematic legal applicability.

A source's precision should not be confused with validity. A geospatial service may supply coordinates to six decimal places even when the point denotes a postal centroid rather than the insured structure. A commercial database can return a detailed industry code derived from an outdated primary activity. A numerical field can be precise in its formatting while its interpretation remains coarse.

External observation and calibration

Consider a binary external observation Z indicating that a feature is present. Let X = 1 denote actual presence, with class prior probability p. If sensitivity is s and false-positive probability f, Bayes' rule gives

P(X = 1 | Z = 1) = sp / [sp + f(1-p)].

For p = 0.10, s = 0.80 and f = 0.20, the positive posterior probability is 0.08/0.26, approximately 0.30769. Despite sensitivity of 0.80, a positive indication does not prove the condition. With p = 0.50 and the same conditional error rates, the posterior is 0.80. A proxy's positive predictive value depends on the prevalence in the population to which it is applied.

A negative observation under the first hypothetical calibration yields P(X = 1 | Z = 0) = 0.02/(0.02+0.72), or about 0.02703. If sensitivity deteriorates to 0.40 because the observation technology covers only a portion of the relevant structures, the probability after a negative observation becomes 0.06/(0.06+0.72), about 0.07692. A negative signal is consequently informative only in conjunction with its coverage and measurement properties.

The calibration sample must also be evaluated. If validation uses only risks that were accepted after passing an earlier screening process, the prevalence and even conditional error rates can differ from the proposal population. The appropriate sample for reconstructing proposed exposures is the set of relevant proposals with credible verification, adjusted for the mechanism that selected them for verification.

Transfer and distribution shift

Proxy calibration is local to an observation regime. Let S denote the origin of calibration observations and T the target underwriting population. Equality of P(Z | X, S) and P(Z | X, T) is an assumption. Changes in building design, reporting convention, product mix or technology can violate it. Even where conditional measurement performance transfers, the prior P(X | T) may differ.

One method of evaluating transfer is a limited but deliberately designed audit in the target population. The audit should sample across geography, class, uncertainty state, channel and proposed action to the extent that the product and lawful data collection permit. Errors should be measured against a target definition fixed before reviewing the source. Validation solely against an earlier version of the same commercial dataset is circular when both versions inherit their values from a common source.

In insurance an incorrect entity match can have greater effect than a modest feature-estimation error. A corporate name may refer to a parent company, subsidiary, operating unit or trade name. A mailing address may differ from the risk location. A data match requires evidence about entity identity. Fuzzy matching should carry uncertainty; an unresolved match in a material feature may be grounds for referral.

Proxy multiplication and portfolio errors

An external characteristic may be individually imperfect but nevertheless useful for classification when combined with other measurements. The information gain of combining them must be evaluated relative to their dependence. Two independently obtained tests with likelihood ratios of four would multiply prior odds by sixteen after two positive results. If the second test merely copies the first, the appropriate combined likelihood ratio is four, not sixteen.

Such duplication has a portfolio effect. A data vendor's systematic misclassification of a roof type or operating activity may be propagated across thousands of policies. The errors are then correlated. Conventional averaging may appear to remove idiosyncratic uncertainty while leaving the shared bias untouched. The insurer's accumulated exposure must therefore account for both individual errors and source-level common errors.

Lloyd's Catastrophe Exposure principle identifies capture and management of exposure data and risk-based quantification and monitoring among its institutional expectations. In the context of this monograph the connection is that source lineage and proxy quality influence accumulation decisions, as well as individual assessment. The scientific relevance of an external data service extends beyond the average accuracy of each field.

Lawful and ethically relevant inference

External information can introduce privacy, access and fairness problems. The use of a variable requires an applicable legal and contractual basis, appropriate handling, and consideration of how it affects the persons or entities subject to the decision. Predictive utility alone cannot establish lawful use. The scientific analysis should distinguish error rates, discriminatory effects, access burdens and protection outcomes, without reducing all of them to a single optimization target.

New York's final Insurance Circular Letter No. 7 (2024) addresses proxy assessments and governance for external consumer data and AI within its specified regulatory context. Its discussion recognizes that ostensibly neutral variables may be associated with protected classes and specifies expectations concerning evaluation. The jurisdictional implication is that a model's permissible purpose and source use must be assessed under the applicable rules; the circular is not a worldwide standard or a general authorization to infer protected traits.

A variable may reflect access to a measurement rather than underlying risk. For example, applicants operating in areas with limited digital records can have more missing fields. If a model systematically classifies missingness as adverse, access to documentation itself becomes an underwriting determinant. The mechanism requires scrutiny even if the association improves a retrospective aggregate prediction. An audit should ask whether alternate evidence could reduce the uncertainty without unnecessary differential burden.

The response to sparse data may entail seeking a better measurement, preserving a broad uncertainty interval, restricting the decision to a preliminary stage, or applying a legally permissible alternative process. Scientific reconstruction is strongest when its outputs remain explicable and contestable. A decision record should identify observed facts, assumptions and sources of uncertainty, rather than attributing an adverse result to a reconstruction as though it were directly established fact.

Uncertainty Ranges

Four sources of uncertainty

Exposure reconstruction requires distinguishing uncertainty in a proposed characteristic, in a shared model parameter, in the future loss process, and in the admissible model structure. A probability distribution for a building's replacement value concerns the first. Uncertainty in a common cost multiplier concerns the second. Variability in the occurrence and scale of future fire losses concerns the third. Uncertainty over whether the model includes a relevant dependency concerns the fourth.

These uncertainties interact. The same reconstruction error can have minimal effect on one portion of the loss distribution and a substantial effect on another. A missing protection feature, for example, may have modest effect on the expected number of minor losses yet change plausible event severity. Underwriting sufficiency must accordingly be evaluated for the risk quantity that informs the proposed action.

An uncertainty interval has meaning only in relation to its construction. A posterior credible interval is probabilistic conditional on the specified model and evidence. A confidence interval has coverage properties under a sampling procedure. An identification region expresses the values compatible with evidence and assumptions. A stress range represents deliberately specified conditions, potentially without a probability measure. Providing a single range without identifying which of these interpretations applies creates ambiguity rather than clarity.

Partial identification

Manski’s (2003) book on partial identification, as described in the publisher’s overview, examines missing outcomes and covariates and the way additional assumptions narrow the identification region. The mathematical argument below independently applies elementary bounds to exposure inference.

For a binary characteristic with 12 observed positive cases among 60 respondents and 40 absent records in a set of 100, the number of positives in the full set is between 12 and 52. The proportion is therefore in the interval 0.12 to 0.52. The bounds are sharp for this finite set absent further constraints: assigning every missing case a zero or one reaches the respective endpoints. This is an identification statement, not a 95% confidence interval. Sampling variation would enter if the objective were a larger population.

If credible evidence constrains prevalence among the forty missing cases to between 0.10 and 0.40, between four and sixteen additional positives are compatible with those aggregates, yielding 0.16 to 0.28. The narrower interval comes from added assumptions. It should be presented alongside their source and applicable scope.

The same logic applies to nonbinary exposure measures when bounds are available. Let observed site values sum to $8 million for four sites. Two additional sites each have plausible replacement-value bounds of $1 million to $3 million on a consistent basis. Total value is then between $10 million and $14 million. If the missing sites share an uncertain unit-cost multiplier, their joint feasible values may be more restricted than the Cartesian product of their individual intervals; if the assumption of a common multiplier is credible, it should be represented in the joint structure.

Where no finite upper bound can be justified, an apparently narrow upper estimate is an artifact of unrecorded assumptions. A policy payout limit may constrain a single insured payment under certain contract conditions, without bounding physical destruction, repeated loss, uncovered liability or all portfolio payments. Exposure identification and contractual loss bounding require separate analysis.

Propagating reconstruction through contract terms

Let G denote gross loss from an event. In a simplified per-event indemnity illustration with an ordinary deductible d and payout cap L, the insurer payment is Y = min(L, max(0, G-d)). This representation assumes that G is eligible covered damage under the contract, with no coinsurance, exclusions, aggregate cap or other modifying provisions. Actual policies may use different definitions. The simple expression illustrates why the reconstructed physical exposure is distinct from the retained financial exposure.

Suppose a property has uncertain replacement value V. Two feasible states are $2 million and $4 million, given equal scenario weight. An event causes eligible gross damage equal to 50% of replacement value; the deductible is $100,000 and the cap is $1.5 million. The payments are $900,000 and $1.5 million respectively. Their conditional average is $1.2 million. Substituting the average value of $3 million into the formula yields $1.4 million, not $1.2 million. The difference follows from the nonlinear cap. This example does not represent an empirical severity distribution.

If the cap were $3 million, both gross damages would be within the active linear region after the deductible, so the payment at mean replacement value would equal the average payment of $1.4 million. Nonlinearity is thus an identifiable reason for propagation of an entire distribution or scenario set. The appropriate degree of detail depends on the structure of the loss function. When attachment or exhaustion is plausible, the discrepancy can become material to the decision.

Conditional on an assumed annual event probability of 0.02 and the two independent value scenarios of equal weight, expected annual payment under this one-event illustration is 0.02 × $1.2 million = $24,000. Plugging the mean value into the model gives $28,000. Using the larger coverage cap yields $28,000 under either approach. The calculation demonstrates the effect of an averaging procedure; none of these values is a proposed commercial premium.

Common-factor dependence and portfolio accumulation

Uncertainty becomes more consequential when reconstructed values share common factors. Consider ten policies whose individual modeled event payments are either $100,000 or $300,000, each with marginal probability one-half conditional on a specified scenario-generating event. Under independent state uncertainties the portfolio payment is $1 million plus $200,000 times K, where K has a Binomial(10, 0.5) distribution. Expected payment is $2 million. Variance in dollars squared is (200,000)^2 × 2.5 = 100 billion, and standard deviation is roughly $316,228.

If a common source-level error makes all ten low or all ten high together, the portfolio instead pays $1 million or $3 million with equal probability. The expected payment remains $2 million, while the standard deviation becomes $1 million. Under independence, the probability of the maximum $3 million is 1/1024, approximately 0.000977. Under the common-factor case it is 0.5. Identical marginal estimates and expected portfolio cost can therefore imply radically different accumulation risk.

This finding is directly relevant to the underwriting use of third-party estimates. Correlated errors may arise from common valuation inputs, shared geocoding systems, synchronized coverage assumptions or vendor classification rules. Merely increasing portfolio size does not eliminate all uncertainty. Averaging works only for the portions of error that diversify under the actual dependency structure.

An insurer's retained event loss must additionally account for reinsurance on its actual terms. Applying a common retention to the sum of per-policy payments differs from applying it independently at each policy. An occurrence definition and reinstatement provision can materially affect the retained tail. The example stops at direct policy payments because treaty reconstruction is a separate analytical task requiring its own contract evidence.

Robustness and the feasibility of decisions

A decision quantity Q may be constructed from the joint exposure, event and contract structure. Let the feasible model set be H. Evaluating Q(h) for h in H yields a range rather than a unique answer. A decision is robust with respect to this set when all relevant admissible exposure estimates lead to the same action under independently specified decision conditions.

Suppose a proposed location can be accepted within delegated authority only when its insured event payout does not exceed $1.2 million in a stated deterministic stress. The two hypothetical property scenarios above yield $0.9 million and $1.5 million. The uncertainty crosses the delegation threshold. An average of $1.2 million cannot establish that both scenarios satisfy the authority rule. If the threshold were $1.6 million, both payments would be within that specific authority bound, although other required underwriting checks could remain unresolved.

Robustness with respect to a set of scenarios is always conditional on the adequacy of that set. If the scenarios omit an important peril or policy feature, agreement across them is weak evidence. The construction of H should document material assumptions, alternative mechanisms, expert disagreement and whether any upper bound is physically or contractually justified.

Accept/Refer Decisions and Error Audits

Three different action classes

The reconstruction supports a spectrum of underwriting actions. Routine acceptance concerns risks for which the evidence and decision conditions are sufficient within the authorized product and coverage. Referral obtains human or specialized assessment, supplementary evidence or revised authority. A decision to decline, restrict or alter terms is a distinct adverse or conditional outcome, governed by the applicable product, law and procedural requirements. Referral should not be silently implemented as an adverse decision, or the validation dataset will conflate uncertainty with inferred danger.

The action is also contingent on time. An estimate sufficient to produce an indicative quote may be insufficient to bind coverage. A short-lived source may need renewed verification at inception. A recommendation can be issued subject to acquisition of specified evidence before the risk is accepted. The scientific assessment of uncertainty and the procedural authority to act must be represented independently.

Lloyd's Underwriting Profitability principle identifies risk appetite, controls and challenge of underwriting assumptions among its expectations for managing agents. It illustrates how a reconstructed exposure can enter an actual decision environment whose authority, tolerances and review processes are specified separately from the inferential model.

Expected value of additional information

Consider a simplified decision problem with two states of a risk, ordinary and adverse. There are two possible immediate actions, accept and decline, and a third administrative option to acquire information before choosing between them. Use hypothetical institutional cost units: accepting an adverse exposure generates cost 100; declining an ordinary exposure generates cost 20; the two other combinations generate cost zero. This loss table is purely illustrative and abstracts from the rights of applicants, applicable legal requirements, contract design and broader protection effects.

If the posterior adverse probability is p, the expected loss of acceptance is 100p, while decline has expected loss 20(1-p). Equality occurs at p = 1/6. At p = 0.20, expected losses are 20 and 16. If a perfectly revealing observation costs 5 units, then with this simple loss model the information option has cost 5 because the correct state-contingent decision incurs no modeled error cost. Its value before the acquisition cost is 16; net improvement is 11. At p = 0.02, minimal immediate loss is 2, so the same perfect test costing 5 is not favored by this loss table.

The conclusion is not that risk thresholds should take these values. The illustration establishes the dependence of the value of information on uncertainty and decision consequences. Even high statistical uncertainty can have little practical value of resolution where every plausible state leads to the same lawful action; modest uncertainty can have great value near a consequential decision boundary.

Imperfect evidence is evaluated by averaging subsequent optimized losses over its possible outcomes. Suppose p = 0.20, sensitivity for the adverse state is 0.8 and false-positive probability is 0.1. The probability of a positive signal is 0.24. After a positive signal the posterior adverse probability is two-thirds. The negative probability is 0.76 and the adverse posterior is 0.04/0.76, approximately 0.05263. The best modeled choice is decline after positive and accept after negative. Joint expected error cost is 0.08 × 20 + 0.04 × 100 = 5.6. If the information costs 5, total is 10.6 and improvement against immediate minimal loss 16 is 5.4. No arbitrary threshold conversion from uncertainty into refusal is required.

In an actual insurer, the feasible actions and loss consequences involve coverage duties, authorized rates or terms, consumer protection, capital, reinsurance, administrative cost and competition. Decision-theoretic illustrations organize those considerations, but they neither determine legally permissible risk classification nor supply a universal monetary valuation of applicant welfare.

Selective labels and observation of counterfactual losses

Underwriting validation faces a distinctive problem: decisions influence which future outcomes are observed. The insurer obtains contractual claims experience primarily from bound risks. Rejected applications may have no corresponding claims record in its system. Withdrawals, cancellations and alternative coverage can further change observation. Comparing the risk estimates of accepted and declined cases using only accepted claims artificially limits the study population.

Lakkaraju et al. (2017) address selective labels in algorithmic evaluation and include insurance among the domains identified by their conference abstract. This establishes the relevance of the observation problem without furnishing a general solution for every insurance portfolio.

Let A = 1 denote binding of coverage and Y a prospective insurance outcome under a specified contract. Observations of Y among A = 1 need not identify the distribution of Y among all proposals. For rejected risks, even the hypothetical outcome under the proposed contract is generally unavailable and may depend on the coverage itself. An evaluation of selection performance is distinct from evaluating conditional forecasts within the accepted portfolio.

If a selection propensity pi(D) = P(A = 1 | D) is known or estimable and suitable conditional exchangeability and positivity assumptions hold, inverse-probability weighting can help estimate certain target-population quantities. Where pi is zero for a class, weighting cannot recover its missing outcomes. Where unrecorded underwriter knowledge affects both acceptance and later outcomes, conditional exchangeability can fail. Extreme weights also increase estimation variance. These are identification and precision constraints, not merely inconveniences solved by larger computing capacity.

An exposure-reconstruction audit can be conducted without always needing future insurance losses. It might verify location identity, building attributes or turnover against a more reliable measurement among sampled proposals. This evaluates the reconstruction target directly. A separate prospective validation assesses downstream loss predictions and decisions where the outcomes can be observed or credibly bounded. The separation permits improvement in measurement even when full evaluation of counterfactual insured losses is infeasible.

Calibration, scoring and monitoring

A probabilistic reconstruction should be evaluated as a probabilistic forecast where independent verification is possible. Suppose four cases receive predicted probabilities 0.1, 0.3, 0.7 and 0.9 for a binary exposure characteristic, and verification results are 0, 1, 1 and 1. The mean Brier loss is [(0.1)^2 + (0.3-1)^2 + (0.7-1)^2 + (0.9-1)^2]/4 = 0.15. A constant probability of 0.75 for the same four realized results gives mean loss 0.1875. This small synthetic comparison demonstrates the calculation only; it cannot support a statistically reliable performance ranking.

For a true event probability q, the expected squared probability loss of issuing p is q(1-p)^2 + (1-q)p^2, which simplifies to q(1-q) + (p-q)^2. The expected loss is minimized at p = q. This establishes the elementary propriety of the binary squared-error score. Gneiting and Raftery's (2007) proper scoring rules study places probabilistic forecast evaluation in a general framework, as established in the article’s abstract and introductory page.

Calibration examines whether probabilities correspond to realized frequencies within an appropriate sample and observation regime. An accurate mean score may conceal poor calibration near the referral threshold or within a particular measurement channel. A reconstruction can also be calibrated but insufficiently sharp to be helpful. Both calibration and discrimination or informative concentration must be examined in relation to the intended decision.

Interval estimates require validation of the stated type. If the model issues nominal 90% predictive intervals for independently verifiable quantities under a stable regime, the realized coverage fraction provides a diagnostic. It cannot in itself establish correct conditional coverage for every class. Scenario bounds, by contrast, should be checked against observed violations and assumptions rather than judged solely by nominal coverage.

Operational monitoring includes source availability, missingness patterns, error rates by source and class, time from observation to decision, incidence of overrides, referral rates and the results of inspections. Comparing these before and after a new source or threshold identifies possible changes to the observation process. Drift in the observed population requires an interpretive response: the model may have deteriorated, the risk environment may have changed, or the institution may have changed its measurement procedures.

Errors, audits and governance

Measurement error, inferential error and decision error should be traced separately. A location may have been correctly measured but incorrectly linked to the policy. The inference may be coherent given an obsolete prior but unsuitable for current conditions. A correct exposure estimate can be passed to an erroneous contract payment calculation. A valid risk assessment can be used outside the insurer's decision authority. Each failure requires a different correction.

An audit record should retain the input state, provenance and model version at the time of decision. Reconstructing an old decision using today's improved data produces retrospective information leakage. Where an override occurs, its explanation and new evidence should be retained. An override based solely on an unrecorded impression may improve some decisions while making subsequent validation opaque.

The documented guidance in ASOP No. 56, sections 3.6.1–3.6.5 (2019) includes testing, sensitivity analysis, output validation, governance and controls for actuarial modeling. In the present research the corresponding audit separates technical reproduction of a calculation from scientific suitability for the intended exposure and procedural authorization of its use.

Materiality should incorporate decision and portfolio consequences. A small mean error in thousands of routine cases may matter less than a systematic error in common-factor catastrophe accumulation. An infrequent misclassification may have large effects for a person whose access to coverage is affected. Monitoring therefore includes errors at a risk-relevant scale and their distribution across relevant lawful assessment groups. Materiality is not exhausted by aggregate monetary loss.

The institution can close the feedback loop by collecting the minimum appropriate evidence needed to resolve identified errors, updating the observation model, retesting affected decisions, and recording whether prior conclusions still apply. Revisions should be versioned. They should not erase the historical evidence of what was known at an earlier decision date.

Comparative Synthesis and Reproducibility

International and jurisdictional settings

The core inferential problem is shared across insurance jurisdictions: relevant potential exposure precedes full observation. Legal duties, professional standards, products and market institutions determine which measurements can be collected, which uncertainty may be accepted, and how decisions are communicated. A comparative research architecture therefore retains common statistical questions while explicitly distinguishing institutional sources.

The IAIS introduction and assessment methodology (2024 revisions) describes the ICP hierarchy and proportional application. Principles, standards and guidance have different formal roles. They provide an international supervisory reference, and implementation occurs through relevant legal and supervisory arrangements. The present monograph does not attribute direct legal force in an individual dispute to the ICP text.

In the United States, the actuarial standards discussed above operate within their professional scope, whereas insurance oversight is state-based. ASOP 12 informs relevant risk-classification work and ASOP 23 and 56 address data and modeling. The NAIC's accelerated life underwriting account provides an example of interjurisdictional regulatory coordination. Adoption of guidance within the NAIC is not evidence that every state enacted the same rule. Product- and state-specific rules require separate verification before operational use.

New York supplies a more specific institutional example through its 2024 final circular. Its focus on external information, AI, proxy assessments and governance illustrates that a reconstruction can be statistically informative and still require scrutiny of the admissibility and consequences of its use. An empirical or hypothetical relationship between variables must be distinguished from the applicable legal inquiry.

Lloyd's oversight principles are directed to managing agents in that market. Their treatment of underwriting controls, assumptions and catastrophe exposure provides a concrete institutional setting for portfolio-level materiality. They are not interchangeable with domestic statutes binding all insurers. A location-level estimate should connect to the relevant syndicate's aggregate exposure and authorized strategy within the applicable contractual and market structure.

In Europe, EIOPA's official publication description for its August 2025 AI governance and risk management opinion identifies national supervisors as addressees and a proportionate risk-based approach. This study uses that institutional description only; the full opinion and all applicable EU and national legislation would require separate analysis before reaching operational legal conclusions. The example shows the importance of distinguishing statistical model governance from the detailed legal conditions that attach to particular uses.

This comparison supports a bounded institutional inference. Explicit uncertainty, provenance, purpose-specific data assessment and documented controls are shared research concerns, expressed through distinct institutional instruments. A uniform model output does not imply uniform authority to use it.

Reproducible synthetic experiments

A reproducible experiment begins with the joint data-generating assumptions, observation process, reconstruction method and target decision. The examples below can be replicated by finite enumeration. They are not Monte Carlo estimates, fitted models or a substitute for empirical insurance data. Exact calculation is preferable where the state space is small.

The missing-data example enumerates the number of unobserved positives from zero through forty. For each possible count k, the reconstructed finite-set prevalence is (12+k)/100. The resulting set runs from 0.12 to 0.52 in increments of 0.01. The restricted scenario range assumes k from four through sixteen, producing 0.16 through 0.28. These values demonstrate identification, not confidence bounds.

The proxy example enumerates joint probabilities of X and Z. With prior probability 0.1, sensitivity 0.8 and false-positive probability 0.2, the four joint probabilities of (X,Z) = (1,1), (1,0), (0,1), (0,0) are 0.08, 0.02, 0.18 and 0.72. They sum to one. Dividing positive-state mass within each observation group recovers the posteriors 0.30769 and 0.02703. Changing only prior probability to 0.5 makes the positive posterior 0.8, isolating the effect of prevalence shift.

The contract example applies the specified deductible and cap to each possible gross loss before averaging. For V in {$2 million, $4 million}, damage ratio 0.5, deductible $100,000 and cap $1.5 million, enumerate the two payouts and their equal weights. The difference between scenario mean and payout at mean exposure is negative $200,000. Changing the cap to $3 million removes the difference for these two scenarios. The two runs isolate the contract nonlinearity.

The portfolio example enumerates the independent states for ten risks. There are 1,024 equiprobable state vectors. A vector with k high exposures produces payment $1 million + k × $200,000. The multiplicity of the amount is the binomial coefficient C(10,k). In the common-factor version there are only two equally likely vectors, all-low and all-high. Both distributions have the same mean; their standard deviations and probability of maximum loss differ as shown above.

Finally, the information-value experiment enumerates the two risk states and the two possible test results. With adverse probability 0.2, sensitivity 0.8 and false-positive probability 0.1, the joint masses for adverse-positive, adverse-negative, ordinary-positive and ordinary-negative are 0.16, 0.04, 0.08 and 0.72. The test-based actions commit errors only in adverse-negative and ordinary-positive states. Multiplying their masses by the respective monetary loss units gives 4 and 1.6. The resulting cost 5.6 plus 5 acquisition units is below 16. This result is reproducible from the loss table and does not assert observed profitability.

The following pseudocode specifies the shared computational logic without introducing an empirical dataset.

prevalence = [(12+k)/100 for k in range(41)]
positive = .8*.1 / (.8*.1 + .2*.9)
negative = .2*.1 / (.2*.1 + .8*.9)
payout(v) = min(1500000, max(0, .5*v-100000))
average_payout = (payout(2000000)+payout(4000000))/2
center_payout = payout(3000000)
independent_losses = [(1000000+200000*k, choose(10,k)/1024)
                      for k in range(11)]
common_losses = [(1000000,.5),(3000000,.5)]
test_error = .04*100 + .08*20
net_test_value = min(.2*100,.8*20) - (test_error+5)

For larger correlated exposure sets, stochastic simulation can approximate the distributions considered here. A scientifically usable simulation records the random generator and seed, parameter scenario, sampling unit, dependence structure, number of draws, estimator and Monte Carlo uncertainty. More draws reduce numerical sampling error conditional on the model; they do not reduce uncertainty over the model or fundamental identification limitations. The exact examples supply small benchmarks for testing such implementations before empirical application.

Designing an empirical study

An empirical extension would define a proposal cohort by product, target geography and underwriting interval. Each observation would have a proposal identifier and an insured entity or location identifier, enabling repeated submissions and shared exposures to be recognized. The study would preserve the exact information available at the time of the original decision. It would record the source versions and time of all later verification.

The first endpoint is exposure reconstruction error against independent measurement. Depending on the quantity, absolute, relative, categorical or probabilistic errors may be appropriate. Sampling should include missing and disputed values and cases near a decision boundary. Where verification is more likely among referred risks, the study should document the sampling design and its impact on generalization.

The second endpoint is decision stability across defensible alternative reconstructions. Holding the applicable underwriting rules fixed, investigators would compare resulting actions. The analysis would estimate how often different reconstructions cross authorized decision conditions, and investigate whether the changes follow source errors, population differences or model assumptions. A comparison of this type is possible without presuming that every declined risk has an observable loss outcome.

The third endpoint is prospective predictive performance among bound risks and any additional lawfully observable comparison populations. The evaluation must distinguish underwriting year, event year, earned exposure, reporting lag and claims development. For long-tailed products, immature experience should not be interpreted as completed losses. Contract terms and changes in the coverage population must be retained.

Time-based testing can reduce leakage and assess transfer across observation regimes. Group-based partitioning should keep repeated proposals from the same underlying entity from appearing on both sides of training and test data where that would distort the intended new-risk task. A separate audit of source changes can evaluate whether a model remains calibrated after a vendor modifies its coverage or method.

A study investigating the outcome of additional information collection could exploit randomized or designed audit sampling where lawful and operationally appropriate. It might compare the knowledge gained from a supplementary verification procedure at different uncertainty levels. Such a study would require privacy safeguards, suitable permissions, and protection against unjustified consumer burdens; no such data collection was undertaken for this monograph.

Rival explanations and limitations

A favorable observed loss experience after adopting external data might reflect better reconstruction, altered selection, changing premiums, different retention, a shift in business mix or simple sampling variability. Without a design that distinguishes these pathways, improvement cannot be attributed to the reconstruction alone. Likewise, a high referral rate might reflect a cautious process, a poorly calibrated model, deficient source coverage, unusually complex risks, or changes in delegated authority.

The appropriate empirical comparisons must therefore hold relevant contractual and portfolio definitions stable or explicitly model their changes. Technical success in reproducing the algorithm does not demonstrate causal improvement in risk selection. A cross-sectional comparison of accepted portfolios is particularly vulnerable to differences in the risks selected.

The study's inferential demonstrations have deliberately simple functional forms. They isolate individual mechanisms: missing-case ambiguity, proxy uncertainty, credibility, contract nonlinearity, shared error and value of information. Real-world modeling may require multivariate measurements, nonstationary severity, dependence across insured parties, nonlinear contractual conditions and uncertainty about future decisions. Those extensions require evidence suited to their actual application.

Source access is also bounded. The Rubin argument is used at the level supported by its publisher abstract. The Manski discussion is restricted to the publisher’s book overview and O’Hagan to his author abstract. Gneiting and Raftery are used within the directly inspected abstract and introductory page of the author-hosted published article. The selective-label discussion draws on the conference abstract; no reported empirical effect size is imported. The institutional sources are used within the sections or descriptions inspected. External peer review, private-data analysis and empirical insurer-specific calibration have not been performed.

These limitations specify the scientific status of the findings. The examples are mathematically testable and the institutional interpretations narrow. They support conclusions about decision sufficiency and methodological requirements under stated conditions. They do not establish universal numerical thresholds, prevalence, insurer performance or legal advice.

Conclusion

Exposure reconstruction is the inferential conversion of partial evidence into a decision-relevant description of a proposed insured risk. Its central scientific requirement is preservation of the distinction between underlying exposure, observations, inferred quantities, prospective events, policy payments and institutional decisions. Each belongs to a different stage of the insurance problem.

Evidence suffices for an underwriting action when the material target quantities are adequately supported, the admissible assumptions and alternatives are explicitly represented, and the intended decision remains defensible within the resulting uncertainty. A precise point estimate alone cannot guarantee sufficiency. The partial-identification example shows that the same observations can support several prevalence levels. The proxy example shows that identical measurement performance can yield different posterior assessments under different base rates. The credibility example shows how limited observations interact with prior information.

Uncertainty should be propagated through the actual decision function. The contract example demonstrates the discrepancy between average scenario payment and payment calculated at average exposure where a cap changes the shape of the function. The portfolio example establishes the material consequences of common-factor errors even when expected exposure is unchanged. A reconstruction process must consequently preserve dependencies and contractual conditions relevant to the action.

Referral is appropriate when feasible exposure states change the authorized decision, when unresolved information can materially alter consequences, or when source suitability and applicable institutional requirements cannot be established. The value-of-information example shows that the usefulness of further evidence depends on its expected effect on the decision and its acquisition cost. Uncertainty should not be automatically equated with adverse risk.

Validation requires direct assessment of exposure measurement, testing of the inferential process, and separate evaluation of downstream actions and outcomes. The selective observation of insured losses limits the conclusions available from accepted portfolios. A credible audit preserves the original information set, uses independent verification where feasible, records observation mechanisms and reports uncertainty at the relevant class and portfolio levels.

The contribution of this monograph is an integrated analysis of established inference, actuarial, contractual and governance methods within the particular scientific problem of sparse underwriting information. The analysis supplies conditional mathematical demonstrations and a comparative interpretation of the relevant institutions. It does not propose a new theory or an empirically calibrated underwriting rule. Its findings establish why scientifically defensible reconstruction must be evaluated according to the purposes, evidence and consequences of the particular insurance decision.

References