Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Claims Anomaly Detection and Fraud Prediction

Abstract. Claims anomaly detection addresses the statistical problem of identifying unusual reported insurance events, claim transactions, documents, relationships, and temporal patterns. Fraud prediction addresses a different inference problem: estimating the probability of a defined fraudulent outcome conditional on information available at a particular time. The distinction becomes decisive when proven fraud is rare, uninvestigated claims are mislabeled as legitimate, fraud investigations are triggered by earlier models, and portfolio composition changes across time and jurisdictions. This monograph integrates statistical anomaly detection, supervised rare-event prediction, positive-and-unlabeled learning, calibration, sampling, operational decision theory, insurance conduct expectations, privacy, model monitoring, and reproducibility. It develops the consequences of selective verification algebraically and examines a clearly hypothetical claims portfolio to show why prevalence, alert capacity, and verification cost dominate the practical interpretation of classification metrics. It separates model scores, documented investigative findings, and legal determinations; compares sector-specific supervisory guidance without imposing a universal legal rule; and establishes a documented research design for prospective performance studies. The contribution is an original critical synthesis and worked methodological analysis, with no claimed discovery of a new theory and no asserted empirical performance on actual insurer claims.

Keywords: Insurance fraud; claims anomaly detection; predictive modeling; unsupervised learning; rare events; positive-and-unlabeled learning; selective labels; verification bias; probability calibration; precision-recall; false positives; risk-based investigation; cost-sensitive decisions; concept drift; fairness; model governance; insurance claims.

Introduction

Insurance claims arise from contracted promises, reported incidents, documentary records, and the decisions of multiple actors. An insurer may receive thousands of plausible claims that vary in amount, complexity, geography, duration, and supporting evidence. Statistical screening becomes useful when these variations exceed the capacity of uniformly intensive manual review. Its central task is to allocate attention to claims whose observable characteristics warrant further inquiry. The resulting alert is an invitation to verification, and every conclusion about fraud subsequently depends on admissible evidence, applicable contracts, and relevant law.

The historical statistical literature already recognized that fraud adapts to defensive measures. Bolton and Hand’s review (2002) relates predictive detection to changing transaction patterns, while Fawcett and Provost (1997) describe behavioral monitors that combine evidence to produce alerts. In insurance, the problem is amplified by infrequent definitive outcomes, lengthy settlement paths, claim re-openings, subcontracted providers, and event-related surges. A high anomaly score can indicate a catastrophic loss, a novel policy design, an administrative error, a rare but valid medical treatment, or intentional misrepresentation. The statistical mechanism identifies deviations; investigation decides their meaning.

This inquiry considers property, motor, health, casualty, and other insurance segments at a level where differences in claim processes can be articulated. The primary object is a data-generating and decision process: event occurrence; notification; registration; enrichment; preliminary scoring; possible referral; evidence collection; a recorded disposition; and later recovery or appeal. Each stage is timestamped, and each stage can affect the information subsequently available to researchers. Insurance fraud and financial crime studies properly elaborate legal elements, criminal schemes, investigatory powers, and sanctions. This Dossier concentrates on predictive inference, score validation, false-alert burdens, and decision thresholds.

Contemporary field-specific research underscores why this separation matters. Schrijver, Sarmah, and El-hajj (2024) examined 50 automobile insurance fraud studies from a defined 2019–2023 search window and documented limited public datasets, frequent use of supervised structured data, and comparatively little explicit attention to the different costs of classification errors. That evidence describes one body of scholarship, not all insurance lines or every jurisdiction. It motivates a systematic distinction between a benchmark score and a defensible production result.

A dated industry disclosure supplies an illustration of why denominators matter. The ABI October 2026 report described £1.34 billion in detected fraudulent general-insurance claims for 2025, representing 93,900 identified cases. The report concerns fraud detected by reporting insurers; it does not observe all hidden fraud and supplies no direct claim-level sensitivity or underlying prevalence rate. Detection totals may change because of reporting coverage, investigation intensity, or severity of claims, in addition to changes in underlying conduct.

The monograph uses published mathematical methods, official institutional texts, and algebraically reproducible hypothetical examples. No proprietary claims dataset was obtained, no insurer implementation was tested, and no actual portfolio detection rate is asserted. The analysis yields research criteria for observing data, evaluating models, designing verification, and governing automated triage. Its conclusions are justified by the reviewed sources, the stated assumptions, and the accompanying calculations rather than by invented field observations.

Research Object, Definitions, and Epistemic Scope

Let a claim be indexed by i, and let t denote a decision time. Define Xᵢ(t) as the information legally accessible, recorded, and available for that claim at that time. Define Yᵢ as a precisely specified target, for example a final, substantiated intentional fraudulent misrepresentation under an explicit organizational outcome taxonomy. Define Vᵢ(t) as the indicator that a claim has undergone a sufficiently complete verification by time t, and Lᵢ(t) as the label actually stored in the operational database. A model produces either an unsupervised anomaly function A(Xᵢ(t)) or a supervised score S(Xᵢ(t)) intended to estimate P(Yᵢ=1 | Xᵢ(t)) under a stated population, observation policy, and event definition.

These quantities have distinct meanings. Unusualness is low statistical compatibility with a reference distribution or neighborhood; it is relative to the modeled population. A fraud probability is a conditional estimate for a defined outcome. A referral decision is a policy action based on expected consequences. Verified fraud is an outcome of a specified evidentiary process. The last category varies with the investigation standard and source of authority. Treating these as identical collapses the fundamental inferential steps and invalidates precision calculations when the stored labels do not represent the target event.

The broad anomaly detection taxonomy (2009) distinguishes point, contextual, and collective anomalies. A modest claim may be anomalous only within a tiny vehicle or medical-service group; a sequence of normal-valued transactions may be anomalous because of timing or repetition; an unusually large claim may be entirely predictable after a major weather event. The reference population and conditioning variables therefore determine what an anomaly means.

A research estimand should specify: the unit (claim, claim line, provider, person, incident, or network); the insured-event and contract scope; a fixed prediction horizon; whether the target is suspicion, confirmed fraud, adjudicated fraud, or prevented loss; how the target matures; the population to which the inference is intended to apply; and the operational point when the score is computed. For instance, predicting a documented fraud referral within 90 days differs from predicting confirmed fraud within 24 months. A model can perform well at predicting previous investigators’ decisions while failing to capture independent fraudulent conduct. This difference is measurable once outcome definitions and selection paths are recorded.

The term fraud prediction is used here only when the target is explicitly operationalized. It has no power to prove mental state or establish statutory elements. A posterior probability conditional on historical labels can inherit the characteristics of historical enforcement and investigation policies. A methodological audit must always ask which target the model actually learned.

Claims Data and the Observation Process

Claims databases are longitudinal operational records, while most benchmark datasets resemble fixed rows with a single label. The distinction affects virtually every modeling choice. Initial notification provides a limited set of facts. Subsequent stages add damage assessments, repair invoices, medical codes, adjuster notes, witness accounts, payment amounts, recoveries, and fraud-investigation records. Each variable has an origin, recording time, permissible use, missingness mechanism, and potential for retrospective correction. The version visible on the training extraction date is often richer than the version available at the intended prediction instant.

The primary unit of analysis requires explicit normalization. Multiple payment transactions may belong to a single claim; multiple claims may arise from one insured event; multiple affected claimants may belong to one collision; one service provider may treat many claimants. Splitting transaction rows randomly between training and test sets can place fragments of the same incident in both groups and produce inflated accuracy. A scientifically defensible split holds incidents, policies, households, organizations, or catastrophe clusters together whenever shared information would otherwise leak. It also respects chronological maturation.

Claims data contain several distinct zero and missing-value processes. A zero paid-to-date may mean a newly registered valid claim, not the ultimate amount. A missing provider identifier may mean that a service is not relevant, that a legacy system failed to capture it, or that information was intentionally withheld. Collapsing all three mechanisms into one missing flag may create a predictor of workflow rather than fraud. Missingness indicators can be informative, but their semantics must be documented and their stability tested.

Investigations and payments also change their own data environment. Once a claim is referred, documentation may become more extensive, payment timing may change, and additional linked claims may be inspected. A model trained on all these final-state variables can appear nearly perfect because it learns consequences of investigation. Feature availability should therefore be audited as of the scoring timestamp using an append-only, time-travel-capable event history. Claims with immature outcomes remain censored rather than mechanically assigned negative labels.

Domain-specific validation should compare incident-linked claim systems, policy periods, reporting channels, products, and catastrophe epochs. The automobile fraud literature review identifies limited and repeatedly reused public sources, complicating generalization. Reproducibility needs a data dictionary, transformations, inclusion criteria, time cutoffs, linkage rules, and documented access controls. A public dataset may support replication of code; it cannot, by itself, establish transferability to a particular insurer’s decision environment.

The causal order of observation is a methodological constraint. A variable created after referral is usually unsuitable for pre-referral prediction even if it correlates strongly with recorded fraud. A transaction attached to one of several linked records may accidentally transmit privileged future information into an apparently harmless input. The audit must operate on fields and relationships as they existed at the exact forecast time.

Fraud Labels, Investigation Selection, and Identifiability

The central statistical obstacle is selective verification. Suppose true fraud Y=1 exists in some share of claims, but a definitive positive outcome is observed mainly where an investigation V=1 occurs. Let the logged positive label be L=1 when fraud is confirmed and recorded. Under the simplifying assumption that recorded positives are correct, P(L=1 | X)=P(V=1 and L=1 | Y=1,X) P(Y=1 | X). If detection and recording conditional on true fraud have probability c(X), this reduces to P(L=1 | X)=c(X)P(Y=1 | X). Observed-positive risk is a product of fraud risk and verification/recording propensity. When c(X) varies with adjuster rules, amounts, geographical groups, or earlier scores, the model cannot identify P(Y=1 | X) from recorded positives alone.

This is an identification statement, not a new named theory. It follows from elementary conditional probability and clarifies the statistical significance of case selection. If some fraud is undetected, L=0 combines genuine claims, unexamined suspicious claims, and examined claims with insufficient evidence. If false positive labels also occur, the equation requires an additional term for erroneous positive recording. Models trained directly on L estimate a property of the entire enforcement-and-recording process.

A second selection mechanism operates through feedback. The existing alerting algorithm routes high scores to investigators. Investigations then produce more verified labels for high-score claims, making those features appear even more predictive in the next training cycle. Apparent improvement may reflect greater inspection. To estimate population performance, a validation study needs explicit selection probabilities and a controlled sample of claims outside the alert stream. A small random audit of non-alerted claims is often statistically valuable because it creates information in regions where the historical system rarely looks.

Positive-and-unlabeled methods are relevant because independently confirmed positives may coexist with a large unlabeled population. Elkan and Noto (2008) and the later positive-unlabeled learning survey explain how estimation depends on assumptions about why positive cases receive labels. Selected-completely-at-random labeling among positives is one possible assumption, not a realistic default for every claims program. Selected-at-random given observed covariates and explicit selection modeling require separate justification. When the true selection mechanism remains unknown, a point estimate of fraud probability may be unjustified; partial identification and sensitivity ranges are more appropriate.

Define a verification protocol before training: final disposition taxonomy; date of first suspicion; date of referral; evidence collected; investigator identity or team; reason codes; threshold for substantiation; appeal result; and reopening status. Auditable labeling protocols include reviewer disagreements and adjudication. Confidence in historical labels should derive from process review, not from the model agreement rate. Missing outcomes due to years-long litigation or subrogation need censoring rules and periodic re-extraction.

The practical consequence is clear: verified positives alone support discovery of patterns among investigated cases, while estimation of false-positive rates across all claims requires reliable observation among the non-referred population. Without that design, the confusion matrix is a description of inspection practice, not a complete portfolio performance measure.

Claim Patterns and Anomaly Scores

The univariate high-value claim is the simplest alerting example and often the least informative. Loss sizes vary with exposure, coverage, claim type, inflation, severity, catastrophe, and jurisdiction. A distributional anomaly score should be conditioned on materially relevant circumstances. A repair invoice that lies in the 99th percentile of one model-year and region may be ordinary for a different risk class. Robust transformations such as log amounts, exposure offsets, and date-specific distributions can improve comparability, but each can suppress or emphasize meaningful extremes.

Distance-based methods evaluate deviations in a selected feature geometry. Their performance depends on scaling, missingness, and the meaning of proximity. Density-based methods can recognize observations that are unusual relative to a local neighborhood. Breunig and colleagues’ LOF (2000) compares the local density of a point with those of nearby observations. Its score is relative to the selected representation, and sparse legitimate subpopulations can receive high values. The investigator needs context about the reference population before interpreting a result.

Boundary or support-estimation methods instead seek a region containing most observations. The one-class support estimator (2001) supplies a principled method for novelty detection from unlabeled data, given assumptions about reference observations. An isolation-based approach, represented by isolation forest (2008), assigns higher anomaly tendency to cases that can be isolated with shorter random-partition paths. These approaches define suspicious statistical rarity through different mechanisms and require direct empirical comparison within a portfolio. No method is inherently a fraud-probability estimator without external validation and calibration.

Collective anomalies are especially important in insurance. Many claims can share an ordinary amount, provider, vehicle, telephone number, bank account, invoice pattern, or workshop while their aggregate timing reveals an implausible arrangement. Graph and sequence models may flag repeated co-occurrence, unnatural temporal synchronization, or high centrality. Yet an emergency medical network, flood, wildfire, mass-accident response, or large service provider can legitimately generate similar structures. Network scores require event-normalized baselines and cluster-level review.

Anomaly detection is best treated as a discovery system that supplements case-specific expert knowledge. It can support unsupervised exploration of previously unseen patterns when labels are unreliable, while supervised predictors can prioritize established fraud typologies. Comparative assessment should include score stability, operational explanations, computational cost, robustness to missingness, and referral yield. Only independently verified outcomes permit a conclusion about detection accuracy.

Supervised and Unsupervised Detection

Supervised fraud detection minimizes an objective function over labeled cases. Logistic regression yields interpretable conditional log odds under a specified feature set; random forests and boosted ensembles capture nonlinear interactions and heterogeneous effects. A forest can identify a high-risk pattern without yielding a calibrated prevalence estimate. The original Fraud Analytics volume (2015) describes the practical integration of predictive analysis, descriptive profiling, and social network methods rather than assuming one algorithm suffices for every task.

The supervised objective should be written precisely. Weighted cross-entropy is useful when training emphasizes rare positives, but weights alter the optimum and can damage deployment calibration unless the probability interpretation is restored. A training loss optimized for AUC, recall, or log loss is mathematically different from the decision-maker’s objective of minimizing costly inappropriate referrals and missed fraudulent payouts. A validated model can serve multiple decision policies only if its estimates, performance and uncertainty have the required properties for each.

Unsupervised detection is valuable when previously unseen modes may escape an outcome taxonomy and labels are scarce. Its statistical success cannot be assessed by simply checking overlap with earlier investigation queues, because that rewards replication of legacy rules. Evaluation requires analysts to review a probability sample of score ranges, including low-score cases, and to report the distribution of verified outcomes conditional on score after accounting for the verification policy. Benchmarking on an older, already triaged dataset supports only a narrower retrospective claim.

Semi-supervised designs combine a trusted reviewed-positive subset with unlabeled observations, or learn representations from large ordinary claim sequences before supervised fine-tuning. An ensemble of expert rules, trained classification, and anomaly detection can improve coverage of known and emerging patterns. Correlated alerts should be consolidated before case assignment; otherwise the system counts the same referral several times and overstates productivity. A claim and its connected entities should carry unique workflow keys so that an eventual investigation outcome can be linked to every contributing model without duplicating the measured benefit.

Model selection should begin with transparent baselines: a small set of prespecified business rules, adjusted logistic regression, and at least one anomaly score. More complex algorithms are justified by prospectively demonstrated gains in the relevant endpoint at an equal investigation capacity, under stable sample design and comparable information. Cross-study leaderboard differences are uninterpretable if one study labels suspicious referrals and another labels final adjudication, or if their unit of analysis differs. Benchmark sophistication is subordinate to target validity.

Rare Fraud Labels and Positive-Unlabeled Learning

When a provider or claimant is investigated after a preexisting suspicion, the probability of receiving a reliable fraud label depends on both underlying conduct and investigation. Positive-and-unlabeled learning offers a framework for observing known positives alongside a mixed population. Under selected-completely-at-random labeling, the chance of being recorded positive is constant among truly positive cases, and the observed label probability can be divided by the labeling propensity when that propensity is identified. The seminal positive-unlabeled classifier analysis derives estimation relationships under its specified sampling assumptions. In actual claims portfolios, investigator selection is often an explicit function of claim amount, referral source, geography, and earlier score, making homogeneous propensity a proposition to test rather than a presumption.

A useful representation is c(x)=P(L=1|Y=1,X=x), so that under one-sided error P(L=1|X=x)=c(x)P(Y=1|X=x). An observed score of 0.02 can correspond to a true conditional fraud probability of 0.02 if c(x)=1, or 0.10 if c(x)=0.2. These values are hypothetical and illustrate nonidentification. Without information about c(x), the same recorded fraction is compatible with materially different latent fraud rates. Calibration against recorded positive labels can therefore be excellent while calibration to true fraud is unknown. The PU-learning survey by Bekker and Davis situates selected-at-random and other labeling mechanisms within the larger methodological field.

Investigation intensity can be recorded as an explicit selection process: prior risk tier, sampling reason, investigator capacity, referral channel, and policy changes. A probability audit outside existing flagged cases supplies information about hidden positives, but random audits must be methodologically and legally appropriate, proportionate, and conducted with a known sampling frame. Some uninvestigated cases will have no final ground-truth determination even after audit; researchers should report incomplete verification and a range of possible rates. Weighting by known investigation probabilities may correct selection where positivity and measurement assumptions hold, yet extreme weights amplify variance and positivity failures prevent straightforward recovery.

Modeling strategies include transparent supervised estimation on reliably resolved cases with selection warnings; semi-supervised or PU algorithms; anomaly scoring; a verified positive case-control study; and a prospective two-phase sample including randomly chosen low-score claims. Every strategy must disclose which performance quantity it estimates and which population remains beyond observation. A plausible machine-learning output is not a substitute for the missing label mechanism. The research quality criterion is a credible observation design rather than the sophistication of the classifier.

Social Networks, Text, and Multimodal Evidence

Fraud investigations may examine connected entities rather than isolated claim records. A graph can represent claims, policyholders, repair firms, physicians, vehicles, addresses, phone numbers, devices, accounts, or incidents. Edges may represent shared identifiers, co-occurrence, sequential billing, service relationships, or ownership. An unusually dense subgraph or an implausible temporal pattern can guide scrutiny, particularly when no individual claim has an unusually high monetary value. The practical role of relational methods is well developed in Fraud Analytics (Baesens et al., 2015), where network analysis complements statistical profiling and supervised prediction.

Graph construction is an inferential choice. A shared address in an apartment complex, a common account belonging to a legitimate provider, or a common workshop serving an entire region can yield high degree without collusion. Identity resolution errors may connect innocent persons; false splits can conceal repeated conduct. Edge provenance, validity periods, confidence, and legal relevance need to be preserved. A graph built from retrospectively linked entities must not be used to assess an earlier model as if those links were already known at forecast time. Cross-validation should keep entire incident or entity clusters from leaking across folds.

Narrative notes, invoices, images, and other documents provide additional signals. Text classification can estimate the similarity of descriptions, measure phrase reuse, or flag inconsistent chronology; document forensics can indicate modification, duplication, or metadata mismatch. Such findings are investigation leads requiring original-file examination and context. Text produced by adjusters after a referral can encode suspicion directly, making a learned model a recognizer of human judgments. Large document models additionally introduce uncertainty around extraction, language, scanning artifacts, and overconfident generated summaries.

Multimodal fusion requires explicit treatment of absent modalities and correlated evidence. Two models derived from the same invoice should not be counted as independent witnesses. Claims photos and location records may be restricted by privacy or retention rules and should be used only with a valid legal basis. A useful system stores the original source, immutable digest where appropriate, access log, evidence chain, and a model-independent description of what each feature purports to capture. An anomaly in an image or network supports an audit question; the claim remains subject to ordinary investigation and applicable procedural rights.

Class Imbalance and Probability Sampling

Fraud is frequently uncommon among all reported claims, while historical training datasets may contain an artificially high share of confirmed cases because insurers curate investigations for study. That mismatch is central to estimating real portfolio probabilities. Suppose the deployment probability of fraud is π and the conditional feature likelihoods for fraudulent and legitimate claims are stable. The posterior odds satisfy: P(Y=1|x)/P(Y=0|x) = [f(x|Y=1)/f(x|Y=0)] × π/(1−π). Case-control sampling that changes the class mixture changes the prior odds even when each class-conditional distribution is unaltered. King and Zeng (2001) study rare-events estimation and its sensitivity to sampling and finite data.

If a training dataset has prevalence q and a model outputs a properly calibrated training posterior p_q(x), under prior-probability shift and otherwise identical conditional feature distributions, the deployment log odds are logit(p_π(x)) = logit(p_q(x)) + log[π/(1−π)] − log[q/(1−q)]. That algebra describes an idealized correction; it fails if investigator selection modifies the features within classes, labels are unreliable, new fraud methods emerge, or the deployment base rate is not known. calibration after undersampling (2015) investigates the related classification problem. Rebalanced training data may be useful for optimization but should not silently dictate the real-world intercept.

Synthetic oversampling can also distort relationships among amounts, categories, networks, and document fields. Interpolating two fraudulent claims may create a nonsensical synthetic claim. Sampling should occur strictly within training folds, with the untouched validation and test sets preserving the real decision population or using prespecified probability weights. The audit should record the number of incidents, rather than only rows, and should separate rare incidence from scarcity of verified labels. An unreported oversampling transformation can make precision and threshold studies essentially uninterpretable.

Historical prevalence measured from recorded fraud cases is itself conditional on detection practice. When population prevalence is uncertain, investigators should present scenario analyses across a plausible range, justified by independent audit or disclosed priors. The uncertainty range is part of the result. For example, with fixed sensitivity 80% and false-positive rate 1%, precision equals 0.8π/[0.8π+0.01(1−π)]: at π=0.1%, precision is about 7.4%; at π=1%, about 44.7%; and at π=5%, about 80.8%. The same error rates can describe radically different referral workloads and yields when base rates differ. This is a direct Bayesian consequence, not an empirical forecast.

Calibration, Prevalence Shift, and Uncertainty

Ranking a suspicious claim ahead of another and assigning a trustworthy fraud probability are distinct achievements. The receiver-operating-characteristic area measures pairwise discrimination; it does not establish that claims scored 0.20 are fraudulent in approximately one-fifth of comparable cases. Calibration evaluates agreement between predicted and observed frequencies for the specified target and population. Van Calster and colleagues (2019) emphasize calibration as a separate critical dimension of predictive assessment. Fraud screening that routes cases by expected benefit needs meaningful probability scales or a directly validated action-value function.

Probability calibration can use logistic recalibration, Platt scaling, isotonic regression, or temperature scaling in appropriate settings. The empirical work of Niculescu-Mizil and Caruana (2005) studies differing probability behavior among supervised learners, while Guo et al. (2017) demonstrate miscalibration in evaluated modern neural networks. Any calibrator must be fit on genuinely out-of-sample labeled observations without contamination by test outcomes. For rare fraud, unverified non-referrals cannot simply be counted as true negatives. A plotted calibration curve of recorded referral labels may assess adherence to historical investigations rather than fraud probability.

Prevalence and class-conditional feature distributions may drift when product mixes, deductibles, repair inflation, claims channels, or fraud techniques change. Recalibration can repair an intercept under an appropriate prior-shift assumption, but cannot solve conceptual target drift or systematic investigation selection. Prediction intervals or credible uncertainty for fraud risks must account for limited verified sample sizes, sampling weights, model variation, and temporal dependence. Resampling individual payment rows when claims share incidents understates uncertainty; resampling should follow independent clusters.

Operational reporting should include discrimination, a calibration plot or calibration intercept and slope where estimable, calibration within consequential score regions, precision and recall under a documented verification design, uncertainty intervals, and score-distribution changes. It should also report whether observed changes reflect labels, sampling policy, or underlying behavior. A sharp decimal forecast from a model trained on selected and delayed investigations can communicate more precision than the data support. Conditional probabilities require both a defined event and a defensible observation mechanism.

Predictive Validation and Reproducible Design

A trustworthy study begins with a protocol defined before inspecting the held-out outcomes. The protocol specifies the insurer business line, contracts, claim unit, target event, prediction time, label maturity, admissible features, case linkage, time horizon, investigator selection process, validation population, metrics, and economic consequences. This protects against the common retrospective practice of choosing features and thresholds until test performance appears attractive. The automobile fraud review by Schrijver and colleagues (2024) gives a concrete reason for caution about comparisons among published studies that rely on different datasets and labeling assumptions.

A temporal design may train on claims first notified before cutoff T, tune on a later nonoverlapping period, and test prospectively on claims notified after a deployment candidate was frozen. Outcome maturity must be explicit: if confirmed fraud often requires long investigations, the test cohort cannot be labeled complete at an arbitrary short cutoff. A backtest should either wait for a predefined follow-up interval, perform appropriate survival analysis, or report mature and incomplete strata separately. A rolling-origin evaluation tests multiple training/test origins to expose period sensitivity.

Some models draw information from a claim network that evolves continuously. The evaluation must replay the graph as of the decision instant, including entity identities known then. Identifiers, repeated photographs, claims from one accident, major storms, duplicate payments, and household groups should be partitioned to prevent incident-level leakage. The feature store should preserve source timestamps and availability dates; data corrected later may be used only if the same correction would have existed then. Externally derived features need their own update timestamps and licensing records.

Model-development documentation should provide inclusion and exclusion criteria, population counts, missingness reasons, number of verified positive cases, audit probabilities, sampling weights, hyperparameter search space, preprocessing and data-version code, class-rebalancing method, confidence intervals, and a sealed comparison against baseline rules. The recorded positives should be stratified by method of ascertainment (random audit, manual referral, police report, court finding, internal resolution) rather than assumed equivalent. Independent case-review of a masked sample can assess label reliability.

Reproducibility means another authorized analyst can reconstruct the split and calculations from the frozen data extract and recorded code while respecting confidentiality. Open benchmark data may illustrate algorithms but cannot prove a private insurer’s performance. Claims-level personal data should remain protected; aggregate results, simulation recipes, synthetic non-identifying examples, and a verification manifest can still support scientific scrutiny. Externally peer-reviewed benchmarks and internal operational validation are complementary evidentiary forms with different inferential reach.

Ranking Metrics and Operating Capacity

The relevant unit of success for an investigative team is often the yield from a bounded queue. A model that perfectly separates half the positives but sends nearly every remaining claim to manual review may overwhelm legitimate settlement. Receiver-operating-characteristic curves plot sensitivity against false-positive rate; precision-recall curves plot the positive predictive value among alerts against sensitivity. Davis and Goadrich (2006) establish their mathematical relationship. Saito and Rehmsmeier (2015) explain why precision-recall visualization can be more informative for a skewed binary task.

Let π be genuine positive prevalence, r(t) sensitivity and f(t) the false-positive rate at threshold t. The alert rate is a(t)=πr(t)+(1−π)f(t), and precision is πr(t)/a(t) when a(t)>0. If capacity is K referrals for N incoming claims, threshold selection must satisfy N a(t)≤K, subject to operational constraints and uncertainty. A model whose observed precision looks excellent on a fifty-percent-positive test dataset may generate overwhelming false referrals if deployed where the true rate is one percent. Report precision-at-K, verified-positive yield per K audited claims, and comparison with the baseline queue at equal work.

The F1 score is 2·precision·recall/(precision+recall). It contains neither verification cost nor the time a wrong referral imposes on a legitimate claimant, and it assigns a specific symmetric balance that a real insurer may reject. The threshold analysis in Lipton, Elkan, and Naryanaswamy (2014) demonstrates that optimizing F1 has its own distinctive mathematical behavior. In insurance, the appropriate threshold depends on claim amounts, investigation burden, legal constraints, available specialists, expected recovery, and the seriousness of delay.

Statistical accuracy should be reported with uncertainty intervals that reflect case clusters, verification sampling and time variation. Performance should also be decomposed by claim type and severity: a queue concentrated entirely on inexpensive claims could increase detection counts while reducing financially recoverable fraud. Net financial benefit estimates require counterfactual reasoning about prevented payment and actual recovery, not simply summing amounts on flagged claims. A simple queue model can incorporate expected investigation duration and service-level obligations to ensure that score-based triage does not become an unbounded claims delay mechanism.

False Positives, Verification, and Decision Economics

A suspicious score generates costs even when no formal denial follows. A false positive can impose adjuster time, further evidence requests, settlement delay, policyholder distress, appeal work, and reputational harm. The cost depends on the claim circumstances: a minor duplicate-invoice check differs from an allegation interrupting medical treatment or a catastrophe-related emergency payment. Consequently a classifier threshold is a decision under uncertainty and must be evaluated against legitimate-claim service consequences, not solely against fraud counts.

Suppose investigating a claim costs c_I, creates an additional expected delay burden d_0 for a genuine claim and d_1 for a fraudulent claim, and permits an expected avoidable fraudulent loss B conditional on true fraud. For a claim with defensibly calibrated fraud probability p, the incremental expected value of investigating rather than ordinary processing is p(B−d_1)+(1−p)(−d_0)−c_I, subject to any separate legal or ethical constraints. If B>d_1 and costs are otherwise fixed, the expected-value threshold is p>(c_I+d_0)/(B−d_1+d_0). This expression is a transparent illustrative arithmetic rule; actual insurance decisions depend on recoverability, investigation effects, claim payment rights, resource constraints, and uncertainty about p.

For example, if expected avoidable fraudulent loss B=$6,000, investigation cost c_I=$120, delay burden to a legitimate claimant d_0=$80 and fraud-case delay burden d_1=$20, the threshold is 200/6060≈3.30%. Changing the legitimate delay cost to $300 raises the threshold to 420/6280≈6.69%. These hypothetical values demonstrate that identical scores can rationally lead to different review policies as consequences change. They neither constitute legally permitted delay tariffs nor claims handling advice. An insurer should not impose extra delay automatically where claim obligations or emergency needs require timely processing.

Some inquiries may have informational benefits extending beyond one claim, such as confirming a recurring provider pattern; others may have low recovery prospects despite high fraud probability. Those externalities may be explored in an explicit study design. Uncertain utility estimates should be subjected to sensitivity analysis, including ranges for the true fraud prevalence, investigation effectiveness, investigation costs, and legitimate-claim harm. A robust program evaluates realized outcomes prospectively and compares against a prespecified normal-handling policy, while protecting claimants from disproportionate consequences.

Detection rate, referral throughput, and net benefit should be reported separately. The monetary amount demanded in a suspicious claim is not the amount saved by flagging it; recovery and prevented overpayment must be evidenced and adjusted for counterfactual payment. Model governance must ensure that business benefit is not achieved by simply refusing or slowing valid claims. The central empirical question is the incremental verified benefit of informed verification at equal lawful service obligations.

Worked Hypothetical Insurance Portfolio

Consider a fully hypothetical portfolio with N=100,000 newly reported claims and a true fraud prevalence of 1%, treated as known for illustration. There are 1,000 fraudulent claims and 99,000 legitimate claims. Assume the outcome is observed for every claim under a complete, independent verification protocol, an assumption rarely satisfied in operational claims data. Suppose screening policy A has sensitivity 80% and a false-positive rate of 4%. Its confusion matrix contains 800 true positive referrals, 200 missed fraudulent claims, 3,960 false positive referrals, and 95,040 true negatives. It generates 4,760 referrals. Precision equals 800/4,760=16.81%, so approximately 83 of each 100 referred claims are legitimate under the hypothetical assumptions. High recall alone therefore gives an incomplete operational picture.

Now consider hypothetical policy B, with sensitivity 60% and false-positive rate 0.5% on the same independently verified population. Policy B generates 600 true positive referrals, 400 missed fraud cases, 495 false positives and 98,505 true negatives. Total referrals fall to 1,095, and precision rises to 600/1,095=54.79%. The greater proportion of confirmed fraud in B's queue comes with 200 fewer detected positives. A and B are illustrative joint operating characteristics of potentially different thresholds or models; their points need not lie on a common model's ROC curve. Without verification cost, value of a prevented payment, claimant burden and capacity, there is no generally optimal choice.

Assume an illustrative equal financial benefit of $1,000 for every correctly referred fraudulent claim, an investigation cost of $100 for every referred claim, and an additional $50 processing burden for each legitimate referred claim. These values are constructed arithmetic inputs, not measured insurer costs. Under policy A, the expected accounting illustration is 800×1,000−4,760×100−3,960×50=$126,000. Under policy B, the corresponding net is 600×1,000−1,095×100−495×50=$465,750. Even though B detects fewer positive claims, its hypothetical net is higher because it creates far fewer false referrals. Changing per-case benefit or missed-fraud consequences may reverse that comparison. The example also presumes that every confirmed detection generates the full stipulated recoverable benefit, an assumption that must be empirically tested before actual economic reporting.

The base-rate sensitivity is equally important. If sensitivity remains 80% and the false-positive rate remains 4%, but prevalence is only 0.1%, N=100,000 contains 100 fraudulent cases, 80 true positives and 3,996 false positives, yielding precision 80/4,076≈1.96%. If prevalence reaches 5%, it contains 5,000 fraudulent claims, 4,000 true positives and 3,800 false positives, yielding 4,000/7,800≈51.28%. These are mathematical scenarios, not prevalence estimates. They demonstrate that direct transportation of observed precision across portfolios is unjustified without accounting for different base rates and potential shifts in conditional error rates.

A researcher with only the investigated subset cannot compute these confusion matrices from historical outcomes alone. The artificial complete-verification assumption is precisely what reveals the limitation. A valid empirical analog would require known sampling probabilities, outcome maturation, and appropriate stratified estimates for non-referrals. A finished economic evaluation would additionally measure prevention and recovery, adjuster minutes, legitimate-claim impact, and whether referred and comparison groups had equivalent opportunities for confirmation.

Fairness, Proxy Variables, and Group Comparison

Insurance claims can reveal geography, disability-related needs, medical treatments, occupation, age, and other variables associated with protected or vulnerable groups. A fraud detector may rely on seemingly innocuous proxies that reflect differences in access to documentation, payment method, or prior investigation. If investigators historically audited one community more intensively, recorded fraudulent outcomes may be denser there even at comparable underlying conduct. Training on that label can perpetuate a selection pattern. Corrective analysis must therefore inspect both the score and the label-generation mechanism.

Fairness cannot be summarized by a single aggregate accuracy figure. Relevant questions include whether legitimate claimants experience different false-referral rates, whether investigation delay burdens fall unevenly, whether verified positives are ascertained comparably, and whether policyholders have meaningful access to review. Protected-group performance can be legally difficult to assess when sensitive attributes are unavailable or restricted. Removing a field does not guarantee independence from proxies such as location or service histories. Equally, group differences can reflect legitimate variations in covered exposures, contract terms, and incident types, requiring substantive contextual interpretation.

The official IAIS Application Paper (2025) addresses fairness and redress within a wider set of supervisory considerations for insurance AI. The United States NAIC model bulletin (2023) discusses the governance and oversight of insurer AI systems and third-party data; it is model guidance rather than a uniformly binding federal or state code. The relevant jurisdiction, insured product, and regulatory adoption therefore have to be documented before deriving compliance conclusions.

Fairness testing should compare claims with coherent eligibility and exposure definitions, and should evaluate the outcome measures that actually matter. If a subgroup is less frequently investigated, an observed low false-positive rate may be an artifact of missing verification; if claims from an event are investigated en masse, a temporary high rate may reflect an event rather than group targeting. Sampling across alert tiers, stratified audits and review of investigation decision records can help distinguish some mechanisms. Where exact group identifiers cannot legally be collected, researchers should explain the constraints and test defensible alternative process measures without inferring individual protected attributes from unverifiable surrogates.

Explanations offered to an adjuster can support consistency but should not be mistaken for causal proof. Ribeiro, Singh, and Guestrin (2016) show local surrogate explanation methods; Lundberg and Lee (2017) provide additive model-attribution methods. Both describe model behavior under defined conditions. A feature contribution is not evidence that an insured person acted fraudulently. The evidentiary conclusion must rest on validated records and lawful procedure, with procedures for correcting source errors and contesting consequential decisions.

Investigative Workflow Boundaries

An anomaly score should enter an established claims administration process as a structured referral with a recorded reason, version, timestamp, and review consequence. The same score can support different forms of attention: ordinary document reconciliation, duplicate-claim clarification, a specialist fraud inquiry, or purely aggregate pattern monitoring. These actions have different evidentiary thresholds and burdens. A system that labels every high-score claim “fraudulent” prejudges the result and contaminates both investigator behavior and the next generation of training labels.

An effective referral note identifies the model and calibration version, decision-time data used, relevant observed pattern, uncertainty, possible legitimate explanations, and a clear next verification question. The note should distinguish source evidence from model inference. For a network alarm, the investigator may ask why two claims share an account; for an invoice anomaly, whether unusual pricing is explained by parts availability or catastrophe inflation; for a document similarity signal, whether the standardized template was used by a legitimate contractor. The process should allow a reviewer to override the referral, record the reason, and remove inaccurate features from future use.

Investigation assignments should incorporate conflicts of interest, confidentiality, and line-of-business expertise. Internal fraud teams, outsourced investigators, medical professionals, and law enforcement operate under different legal powers and reporting obligations. The statistical model cannot confer authority to access records or impose sanctions. Case records must preserve the type of final disposition, adjudication standard, and appeal result. A settled claim without prosecution or recovery does not universally mean fraud was absent, while an internal suspicion is not a conviction. End-state categories must be correctly represented in the training target.

Periodic process audits should test how often an alert changes the speed of legitimate claims, the burden of information requests, and the consistency of final decisions across reviewers. Audits should also record erroneous alerts identified before formal investigation, which can otherwise disappear from the database and make a screening system look artificially precise. Attention to overturned and corrected decisions improves accountability and future label quality. The statistical goal is better-directed verification without replacing the underlying contractual and legal procedures that determine an insured's entitlement.

Regulatory and International Governance

International insurance supervision offers principles for governance, proportionality, fairness, and redress while preserving differences among legal systems. The 2025 IAIS AI supervision paper discusses supervisory concerns across multiple jurisdictions and is intended to support application of insurance core principles rather than to enact a universally binding claims-fraud algorithm law. The US NAIC insurer AI model bulletin provides a supervisory governance template whose legal operation depends on state-level implementation and applicable law. The NIST AI Risk Management Framework is a voluntary general risk-management resource, not insurance fraud legislation. A research protocol should distinguish international supervisory guidance, model regulator instruments, binding national or subnational legal duties, and voluntary management frameworks.

Governance begins with an inventory of each system's intended use. An anomaly alert prepared only for internal triage differs from an automated denial, price increase, cancellation, or formal accusation. A use case may be high impact because it affects urgent medical payments or economically vulnerable consumers even when the model is technically limited to internal review. Risk assessment should account for actual downstream outcomes, outsourced data providers, data provenance, operational dependence, and staff incentives. Senior accountability, independent validation, incident management and vendor oversight remain relevant wherever a complex automated decision aid shapes claims outcomes.

The European legal landscape is similarly use-case-specific. The official European Commission AI Act examples distinguish certain high-risk life and health insurance risk assessment and pricing from other insurance operations, including some claims-management examples. Fraud screening is not automatically assigned a uniform AI Act category merely because it uses artificial intelligence; the precise intended function and legal annex determine classification. GDPR duties apply to relevant personal-data processing irrespective of the label put on the model. An organization must assess its own processing, legal basis, access controls, retention, transfer, and data subject rights, taking account of the jurisdiction and date of application.

A model record should therefore state: responsible legal entity, relevant territory, insurer and insurance line, vendor and hosting roles, training population, external information sources, permitted decision function, model validation date, significant harm hypotheses, change-control protocol, and human review path. General supervisory material supports governance design but does not replace a jurisdiction-specific compliance opinion. This distinction is particularly important for multinational insurers using a shared technical model across legally distinct claim-handling operations.

Privacy, Contestation, and Evidence Preservation

Fraud screening can combine policy records with medical information, images, device traces, geolocation, third-party databases, and network information. Statistical availability is not a sufficient reason to collect or deploy these features. The official General Data Protection Regulation sets out principles of lawfulness, fairness, purpose limitation, data minimization, accuracy, security, and relevant rights within its territorial and material scope. Article 22 provides a particular framework for qualifying solely automated decisions with legal or similarly significant effects, including applicable exceptions and safeguards. Whether a specific claims action falls within those conditions requires examination of the operational decision, meaningful human intervention and other statutory facts.

A claims model can expose private information through trained parameters, investigator interfaces, feature explanations, and graph links. Governance should restrict access to the minimum necessary roles, log access to sensitive evidence, keep source purpose and consent or other legal basis where applicable, and define retention by a defensible legal schedule. The preservation of an investigation record must be reconciled with rights to correction, data security obligations, and legitimate records retention. A model should not amplify data collected through incompatible purposes or infer sensitive characteristics merely because they improve apparent classification.

Contestation must operate on the actual decision path. A claimant who challenges an unusual invoice attribution needs a route for correcting the underlying data; a claimant affected by payment delay needs reasons connected to the permissible claim-handling action, access to review within legal rules, and assurance that an anomaly score is not treated as conclusive proof. Recording reviewer overrides and final corrections makes the monitoring dataset more informative. Automated scoring can be used to prioritize human inquiry; meaningful oversight requires that the reviewer has information, authority, time and incentives to examine the merits.

An evidentiary chain should preserve original documents, provenance metadata, extraction procedures, relevant software versions, and interpretation. Cryptographic hashes can support integrity checks but do not independently establish whether the underlying content is truthful. Explanations should identify which data were observed at what time and which inference was generated from them. For comparative research, de-identified aggregate outputs and synthetic examples can permit methodological scrutiny without public release of individual claims.

Drift, Adversarial Adaptation, and Model Stewardship

Claims patterns evolve with repair costs, health service billing, climate events, vehicle technologies, economic conditions, enforcement, and intentional adaptation. Gama and colleagues’ concept-drift survey studies learning under changing data distributions and supplies vocabulary for distinguishing changes in covariates, labels and predictive relationships. An alerting rule can degrade because fraudsters change techniques, but it can also change because a major hailstorm creates thousands of legitimate claims with previously unusual features. Drift detection must therefore incorporate business-event context rather than automatically interpret distributional shifts as new misconduct.

Ongoing monitoring should include incoming claim volumes, missing fields, duplicate entities, class and product mixes, score quantiles, referral proportions, investigation completion times, appeal rates, claim payment times, and verified outcomes on an independently sampled subset. A sudden improvement in apparent precision after management narrows investigation criteria can be a selection artifact. Recalibration should not be undertaken merely to align scores with the biased subset. Model retraining requires a frozen comparison with the existing system, investigation-selection diagnostics, and a rollback plan. New fraud typologies should be described in behavioral terms, with evidence rather than labels derived solely from the detector.

Adversarial adaptation adds a strategic layer: a claimant or organized group may vary claim timing, amounts, identities or document formats after learning screening patterns. The classical fraud-detection literature reviewed by Bolton and Hand and adaptive monitoring research by Fawcett and Provost situate this issue historically. A responsible insurer balances confidential operational criteria with the policyholder's legitimate entitlement to understandable decisions and challenges. Excessive secrecy can hide errors; complete disclosure of manipulable feature thresholds can reduce detection effectiveness. Information disclosure should be purpose- and rights-sensitive.

Model stewardship also requires retirement. A detector whose data cannot be lawfully maintained, whose outcome labels are irreparably contaminated, or whose false positives impose unacceptable harm may need to be withdrawn. Replacement should require an independently verified improvement in the specified operational objective. Model lifecycle records should preserve prior versions, alert policies, documented incidents, and known limitations. The purpose is durable inference and proportionate investigative support across changing portfolios.

Research Agenda, Limitations, and Synthesis

The empirical research gap is not simply the absence of a more complex algorithm. It is the shortage of studies that join an auditable fraud target, prospective feature availability, an unbiased or explicitly modeled verification sample, realistic investigation capacity, valid probability calibration, and independently quantified claimant impacts. Many benchmark publications emphasize retrospective classification under curated labels. The automobile fraud systematic review (2024) illustrates this concern in a defined sample of automobile studies, while general anomaly literature develops methods applicable to other domains. Cross-disciplinary synthesis points toward stronger research designs rather than a claim of a new statistical discovery.

A rigorous future study would preregister an operational target and a mature follow-up horizon; preserve exact claim-time features; use clustered temporal splits; record every referral decision; stratify randomly audited non-referrals; model or bound selective-label bias; fit calibration on a truly independent sample; compare to an ordinary-rule baseline at equal investigation capacity; and report legitimate-claim delay and appeal outcomes alongside confirmed detections. If random audits cannot be lawfully or ethically undertaken, the study should state that population recall and false-positive rates remain partially unidentified and use transparent sensitivity analysis. Selection assumptions must be tested where possible and communicated wherever they control the conclusion.

A second research program should measure financial benefit by independently evaluating actual avoided payments and recovery, including investigation costs and harm to legitimate claimants. A third should assess network algorithms on incident-held-out and catastrophe-held-out test samples to distinguish collusion signals from shared-event effects. A fourth should investigate how explanations influence reviewer behavior, whether investigators over-trust numerical scores, and how data correction alters both referral fairness and future label quality. These are research directions derived from the literature and analysis, not newly proposed proprietary conceptual systems.

This Dossier is bounded by secondary literature, official institutional materials and transparent hypothetical arithmetic. It does not report a new labeled claims dataset, field experiment, claims prevalence estimate, tested insurer process, exhaustive global insurance regulation survey, or external peer review. Some peer-reviewed works were accessible through publisher abstracts and indexed metadata rather than complete subscription texts. Institutional figures report detected fraud outcomes according to the issuing organization, not total latent incidence. The comparative synthesis is designed to make those evidentiary boundaries explicit rather than to transfer conclusions beyond their basis.

Conclusion

Claims anomaly detection identifies statistical departures from a defined reference process, while fraud prediction estimates a specific conditional outcome under a documented observation mechanism. A model can identify unusual behavior without detecting fraud, and a model can predict historical investigations without estimating underlying fraudulent conduct. Distinguishing those targets is the foundation of a defensible insurance application.

The mathematical centrality of selective verification follows directly from the fact that verified labels are generated by investigation decisions. Observed fraud indicators are jointly shaped by actual conduct and the opportunity to uncover and record it. Positive-and-unlabeled learning, selection modeling, controlled auditing and sensitivity analysis each address portions of this difficulty under explicit assumptions. Unidentified labels do not become ground truth through optimization, and impressive retrospective accuracy can coexist with unknown population performance.

The hypothetical insurance portfolio confirms that base rates and false-positive rates determine much of an alert queue's practical meaning. A high-recall model can create thousands of legitimate referrals, and a lower-recall policy can deliver a larger verified yield per unit of investigative capacity under specific costs. Probability calibration, time-valid features, event-cluster separation, mature outcome follow-up and economic counterfactuals are therefore integral to scientific evaluation. Fraud scoring should be judged as an intervention in a claims process that also owes contractual service to legitimate policyholders.

Across jurisdictions, insurance supervisory guidance, privacy law, model-governance instruments and consumer safeguards require attention to the actual use of the system and the actual effect of its decisions. Statistical flags support lawful investigation; final claim treatment follows verified evidence and the applicable legal framework. The strongest future research combines transparent probability inference, representative outcome verification, operational economics and measured claimant impacts. That synthesis defines the present limits of warranted claims about fraud prediction and the evidence needed to move beyond them.

References