RESEARCH DOSSIER
Health Insurance Incurred But Not Reported Reserve Estimation
Abstract. Health insurance incurred but not reported (IBNR) reserve estimation determines the value of covered medical events or services that have occurred before a valuation date but have not yet entered the insurer’s recognized claims records. Its operational context also requires estimating reported but unresolved amounts, subsequent revisions to recognized claims, and the total incurred-unpaid liability. This monograph examines the origins and measurement of claims reporting lags; service-date and payment-date development triangles; completion factors; Bornhuetter–Ferguson and expected-loss methods; frequency-severity and stochastic models; credibility; runout; data provenance; and reserve adequacy across insurance accounting and prudential regimes. An entirely synthetic, fully reconciled numerical study distinguishes total unpaid claims from pure IBNR and demonstrates structural sensitivity to immature cohorts, reporting delays, payment inflation and organizational change. Comparative analysis draws on U.S. health actuarial and statutory materials, IFRS 17, European Solvency II, Australian private health insurance prudential standards, and international health financing statistics. The principal conclusion is that reliable health IBNR estimates emerge from documented event boundaries and data transformations, statistically credible development projections, separately tested reporting and settlement processes, and transparent reconciliation to the governing liability basis. No patient-level claims data, insurer-specific actuarial certification or newly invented actuarial doctrine is asserted.
Keywords: health insurance; incurred but not reported; IBNR; unpaid claims liability; claim lag; claims development; completion factors; Bornhuetter–Ferguson; chain ladder; stochastic reserving; runout; credibility; IFRS 17; health actuarial science.
Introduction
A health insurer promises to finance medically defined benefits while information about the underlying care moves through several institutions. A patient receives a service; a provider codes and submits a bill; a clearinghouse or administrator validates it; the insurer applies contractual rates and coverage rules; payments, denials, appeals, coordination-of-benefits recoveries and subsequent adjustments follow. At a reporting date, a portion of the covered obligation has occurred without becoming visible in the data used for valuation. That information gap generates true incurred-but-not-reported claims. A second gap exists after claims have been reported but before the amount is final or paid. Both gaps must be addressed when estimating the incurred-unpaid obligation.
The U.S. ASOP 5 on incurred health claims expressly addresses estimation for financial reports, claims studies and other actuarial work. Its definitions include incurral date, exposure unit, development method, projection method, incurred claims and unpaid claims liability. These definitions give the health-specific anchor for the present inquiry. A mathematical projection is only as meaningful as the contract-defined incurrence event and the data process that maps that event into the claims ledger. The estimate therefore begins with semantic and institutional reconstruction before an analyst selects a development factor.
The research object is the measurement of unpaid obligations attributable to covered prevaluation health events, with strict analytical identification of the genuinely unreported component. The term reserve may refer to a management best estimate, a statutory liability, a prudential technical provision or an accountancy carrying amount; these quantities can differ in expenses, discounting, recoveries and risk margins. This study distinguishes the underlying cash-flow expectation from each reporting basis. It uses established actuarial methods and primary standards and develops an original comparative synthesis and worked illustration. The formulas below are identities or applications of published methods, not new proprietary theories.
The source hierarchy is explicit. Adopted standards and original official reporting templates establish what a regime requires; original scholarly work establishes model properties; government data caveats show the practical meaning of delays; and the synthetic example makes assumptions and arithmetic inspectable. Publication year and legal applicability are separately recorded. An official U.S. model regulation describes a regulatory template, while applicable enacted state rules govern particular insurers. A European directive and Australian prudential standard apply within their respective regulated domains. This distinction governs every comparative inference that follows.
Foundations and Source-Critical History
The object of estimation
For a valuation date , let be the ultimate covered cost of medical events incurred through , evaluated on a consistently specified benefit and currency basis; let be the corresponding paid amount. The total incurred-unpaid obligation before separately identified expenses, discounting and margins is . This accounting identity describes the economic quantity that a paid-claims development method seeks to estimate. It does not identify which unpaid items have already been reported.
A decomposition requires a reliable reporting status and an amount measure. Let denote the estimated remaining payment for events already registered in a claims inventory, inclusive of subsequent adjudication development under the chosen definition. Let denote the expected ultimate payment for covered events unreported at . Then when classification is exhaustive and the same settlement expenses and offsets are consistently treated. The pure IBNR component is . Claims conventionally called IBNER—incurred but not enough reported—belong to development of the registered population. Some carrier ledgers label the entire development provision “IBNR”; that operational abbreviation must be disclosed and translated before values are compared. Schnieper’s separation of true IBNR and IBNER supplies the classic conceptual research distinction.
Health products complicate this decomposition. A hospital admission may straddle the reporting date; an episode-based payment may be adjusted after discharge; a pharmacy event may appear with a rapid electronic response and later rebate; a disability income benefit may accrue over an extended period. The incurral boundary must follow relevant contract provisions and the purpose of valuation. The same event can produce several service lines, remittances or corrected claims. Consequently, a payment record is neither automatically a unique medical event nor automatically evidence that the entire related liability has matured.
A practical taxonomy separates: unsubmitted eligible services; submitted but unreceived transactions; received but not registered claims; registered but unadjudicated claims; approved but unpaid claims; adjudicated amounts subject to appeals or reversals; and administrative or claim-handling expenses. It also distinguishes recoveries and interinsurer coordination from gross benefit payments. Boundaries between categories are operational, but every category must be assigned to exactly one component at a chosen valuation snapshot. The NAIC health reserve model explicitly requires reserves for incurred-but-unpaid health claims and estimated claims settlement expenses, and calls for prior-period runoff testing.
Historical and methodological development
Early reserve methods extrapolated the observed emergence of claims into ultimate liabilities. Bornhuetter and Ferguson (1972) demonstrated that an independently constructed expected ultimate can provide a stabilizing starting point where available development experience is immature. Their original actuarial paper on IBNR arose in property and liability insurance, but its algebra applies to health cohorts when the emergence curve and exposure basis are meaningful. It must not be treated as an empirical discovery about medical-claims lags.
The modern stochastic literature clarifies what deterministic factors leave unmeasured. Mack’s distribution-free chain-ladder error analysis addresses reserve prediction uncertainty under its assumptions. Renshaw and Verrall’s generalized-linear interpretation of chain ladder connects incremental development with statistical modeling and notes handling of negative increments. England and Verrall’s stochastic reserving review extends the discussion to predictive distributions, simulation and model selection. Wüthrich and Merz’s stochastic claims reserving monograph develops reserving theory and prediction errors comprehensively.
The transfer of these methods to health insurance requires specific attention to electronic claims submission, capitation, negotiated provider rates, coverage churn, authorization rules and seasonal utilization. A robust method cannot simply borrow a casualty triangle without examining those features. The source-critical distinction here is between proof of a method’s mathematical properties and proof that its assumptions fit a particular health claims process. The former may be available from a journal article; the latter requires local, reproducible experience data and valuation-specific tests.
Claim Reporting Lags and Runout
From medical occurrence to administrative observation
A claim can be dated by service, admission, discharge, receipt, entry into an adjudication system, determination, payment or remittance. The delay between service and initial insurer receipt is a reporting lag. The delay between initial receipt and payment is a processing or settlement lag. Total paid emergence from service combines both, and that combined process is the object of a paid development triangle. A reported-claims triangle instead follows administrative visibility or incurred valuations; it is dependent on a stable case-estimation convention. Each clock must be defined before time-indexed factors are calculated.
Let be service date, initial report date and settlement date. Reporting delay is , adjudication/payment delay is , and total lag is , with for an unambiguous single-event claim history. Reopened claims, adjusted remittances and multi-line bills require event-level linkage to preserve the identity. A single payment date rarely identifies pure reporting lag. The correct data architecture retains both original claim and revised claim identifiers, effective dates, posting dates and sign conventions.
Electronic submission can greatly accelerate the earliest visibility of a routine claim, but institutional bottlenecks persist in specialty services, incomplete coding, provider contracts, cross-payer coordination and retroactive eligibility. An observed trend toward faster paid emergence may reflect better administrative throughput, earlier recognition, changes in the mix of medical procedures or increased denial rates. Causal interpretation requires event-type and coverage controls. A change in measured lag can occur without any change in the ultimate medical liability.
A useful official boundary example comes from the CMS historical Medicaid claims-lag warning, which described a period in which the T-MSIS transmission process could require nearly seven months before CMS received 90 percent of claims. The statement relates to Medicaid and CHIP data received by CMS, including state submission processes, and to the reported historical program period. It is not a seven-month lag assumption for private health insurance, nor a direct insurer-specific reserving factor. Its scientific value lies in showing that observation at a central data recipient may lag both medical activity and the payer’s local transaction records.
Runout as a longitudinal experiment
A valuation is followed by a stream of newly reported claims, payments, reversals and releases. A runout study compares what was estimated at time with actual subsequent emergence for the same pre- occurrence cohorts. The analyst must carry the cohort boundary forward while excluding post- medical events. Otherwise the runoff denominator silently changes. Late benefit adjustments and recoveries should be attributed consistently to the original cohort; material corrections require an explicit restatement protocol.
Runout validates several claims separately. A count-based check asks whether previously invisible services later appeared in expected volume. A monetary check asks whether eventual net payments matched the selected ultimate. A timing check asks whether cash requirements occurred in the projected period. An adjudication check examines the level and direction of revisions. Strong performance on one dimension cannot establish performance on all others: a correctly estimated ultimate can still generate a severe liquidity miss when claims are paid earlier than expected.
A useful diagnostic is the cumulative paid completion ratio , for occurrence cohort at development age . It is known only retrospectively after sufficient maturity and depends on whether ultimate includes late recoveries. When payments can reverse, a ratio can temporarily fall or exceed one. For operational use, an ultimate benchmark requires either a defensible tail convention or observation of an effective closed cohort. The 2025 NAIC health development exhibit demonstrates why paid and incurred development are both relevant, while the health statement filing instructions connect prior-year incurred and unpaid-liability fields to regulatory reconciliation.
Measurement of reporting distributions
A count-level reporting curve may be estimated from the empirical distribution of by product, provider, state or geography, claim class and month. If service dates are left-truncated or reports are right-censored at extraction, naive frequencies are biased toward short lags. A survival-analysis treatment explicitly recognizes that unreported events cannot yet appear in a database of reported claims; denominators require exposure or external completion information. Case-mix weighting matters when high-cost claims report more slowly than routine claims.
For service cohorts with verified membership exposure , separate the expected number of events from expected cost per event: in a simple frequency-severity representation. The parameters and are conditional on policy design, population and provider payment terms. Development observations can identify portions of this structure, but a reported-only sample cannot independently identify events that will never report. Alternative assumptions about provider behavior and missing events must therefore be made visible.
The estimation of a future unpaid balance also depends on a valuation cut-off. A dataset extracted ten days after financial close may contain both legitimate late postings and corrections to amounts previously recorded. A reproducible process records the closing snapshot, each subsequent extraction date, immutable transaction history or revision log, and the event-time assignment rule. These details turn runout from a managerial anecdote into an auditable predictive test.
Development Triangles and Credibility
Triangular data and mathematical structure
Let be cumulative net paid medical claims for service or incurrence cohort at development age , where age measures elapsed months from the chosen cohort origin. At valuation time only cells satisfying the observation boundary are available. A triangle may be organized by service month, service quarter or accident/incurrence year; the appropriate aggregation follows volume, tail length and reporting regime. Cells should express a consistent set of benefits, with stable reinsurance treatment, recoveries and expenses.
For adjacent ages, a volume-weighted link ratio is , calculated over a consistent matched set of observable cohorts. The age-to-ultimate factor and . The associated paid-basis total unpaid estimate is . The calculation estimates ultimate remaining payments under the factor assumptions. Calling the entire pure IBNR would be a category error whenever reported but unpaid claims exist.
The method assumes a form of exchangeability in development after normalization. Unstable claim settlement systems, sudden coverage changes, cyber incidents, medical-price shocks or calendar-period processing interventions can invalidate it. Link ratios formed from a single large claim or small immature cell have wide and asymmetric uncertainty. A selected factor should be accompanied by exposure volume, observation count, historical range and a validation rationale. The estimator’s numerical precision does not substitute for clinical or operational comparability.
An incurred triangle, if defined as paid plus reported case estimates, may mature differently from a paid triangle. For health benefits, case reserves can be sparse for many routine bills, and payment may be the first reliable amount record. Where case estimates do exist, their systematic revisions can create apparent development unrelated to new medical services. The modeller must distinguish true claims emergence from changes in inventory recognition and case adequacy. The classic IBNR-versus-IBNER distinction is therefore operationally consequential, especially when managers use paid-to-incurred comparisons to infer reporting adequacy.
Expected ultimate and Bornhuetter–Ferguson
For highly immature cohorts, a development factor can multiply an unreliable paid signal. An expected-cost projection built from exposure, covered morbidity, observed utilization, seasonality and provider prices provides an independent anchor. Let be that prior expected ultimate, and the expected paid completion proportion. The standard paid Bornhuetter–Ferguson form is , with remaining unpaid . Under this specification, the as-yet unpaid portion is driven by the prior, while actual paid experience is taken as observed.
The prior must be independently supportable. A premium-based expected loss ratio can conceal risk-selection, benefit-design or subsidization changes. An exposure-based expected cost can conceal utilization shifts and provider mix. Both require documented population denominators and benefit exclusions. The historic Bornhuetter–Ferguson original analysis establishes the method’s intellectual provenance; the selected prior for a contemporary medical portfolio remains a local empirical and professional judgment.
Blended estimators may use a credibility weight : . This is a transparent judgmental interpolation in the present illustration; there is no claim that a particular is statistically optimal. A rigorous credibility model must specify the loss function, uncertainty covariance and method-dependence. Chain-ladder and BF estimates share development-factor uncertainty, so treating them as independent is generally inappropriate.
Count-severity, generalized linear and stochastic models
A triangle of incremental payments supports statistical models with cohort, development and calendar effects. A log-link model may take under appropriate positive-mean assumptions. Identification constraints are needed because cohort, development and calendar indices are algebraically dependent; model choice and regularization must be explicit. Negative net increments require a distributional treatment consistent with reversals and recoveries, not silent deletion. Renshaw and Verrall (1998) show that a GLM formulation of development can accommodate features absent from a narrow deterministic rule.
A complementary microlevel formulation models report counts conditional on exposure and delayed reporting probability, then models allowed cost and remittance amount conditional on claim class. Frequency-severity models can better isolate whether increased medical expenditure reflects more services, higher negotiated prices, more intensive coding or larger claims. However, inaccessible or poorly linked patient/provider data can make a simple well-validated triangle more reliable than a complex unstable model. The scientific comparison is between predictive evidence, uncertainty and explanatory fit, rather than algorithmic sophistication.
Mack’s chain-ladder standard-error result provides a principled error estimate under its own independence and variance structure. England and Verrall (2002) document predictive-reserve distributions and simulation, including bootstrap and Bayesian methods. Reporting a reserve range demands attention to process variation, parameter uncertainty, tail uncertainty, covariate shift and model selection. A nominal 95 percent interval from an invalid development model does not deliver 95 percent real-world coverage.
Credibility across heterogeneous health pools
Risk pooling increases statistical volume but can average away different reporting processes. Product, provider, service line, geography, enrollment channel and administrator may each define a credible subgroup. Partial pooling seeks a balance: common trend and development parameters may be shared while outlier groups receive separate parameters. For very small groups, external completion assumptions can serve as provisional benchmarks, provided their population and administrative origins are documented.
The international health-financing context makes this especially important. The OECD’s classification of health financing schemes distinguishes government schemes, compulsory insurance, voluntary insurance and household payments. These are different financial responsibility structures. A paid-government-account figure, an insurer-allowed-claim figure and a household-expenditure figure are not interchangeable observations. A credible IBNR triangle must measure an actual obligation of a specified risk-bearing entity, not aggregate national medical expenditure.
A final safeguard is to distinguish data fit from future validity. A model can reconstruct mature historical triangles while failing when contracts, claim intake or disease patterns change. Rolling-origin out-of-time tests, cohort calibration and explicit change-point analysis provide stronger evidence than a single fitted development curve. The role of actuarial judgment is to define the plausible data-generating mechanism and document the reasons for accepting or overriding estimates.
Reserve Adequacy and Cash-Flow Stress
Economic liability versus financial reporting measure
An undiscounted expected unpaid-claims estimate is an input to financial measurement, not automatically the recognized balance in every regime. Under IFRS 17 Insurance Contracts, an insurer measures insurance contract obligations in a reporting framework that incorporates current estimates of future cash flows and applicable risk adjustment. The IFRS Foundation’s liability for incurred claims definition captures obligations associated with services already provided, while remaining-coverage liability addresses insurance services to be provided. Thus an IBNR estimate can contribute to incurred-claims cash flows without exhausting the accounting measurement. The application depends on contract grouping, measurement approach, discounting requirements and the entity’s accounting policies.
European Solvency II places technical provisions on a best-estimate-plus-risk-margin basis under Article 77 on technical provisions. The best estimate is a probability-weighted discounted value of relevant future cash flows, with reinsurance recoverables calculated separately. In the United States, the NAIC 2026 statutory manual addresses health claim reserves and their relationship to premium and contract reserve adequacy. The jurisdiction and insurer category determine which model or reporting treatment actually applies. These regimes should be compared through a reconciliation bridge rather than by comparing balances labeled “IBNR” at face value.
The Australian Prudential Regulation Authority’s HPS 340 private health insurance valuation, in force from July 2023, directly requires valuation of outstanding claims incurred before the reporting date whether reported or not. Its components address claims, claim-handling expense, risk equalization and settled-but-unpaid amounts, with specified recoveries, central estimate and risk margin. The standard provides particularly direct evidence that health IBNR sits inside a broader prospective valuation. Specific requirements, including its sufficiency floor, belong to APRA-regulated private health insurers and must not be generalized as IFRS or U.S. rules.
Internationally, IAIS insurance valuation principles provide high-level supervisory context. Standards differ in the composition of risk adjustment, risk margin, expenses and reinsurance. A comparative research dossier therefore distinguishes (a) expected covered benefits; (b) estimated settlement costs; (c) time value of money; (d) risk adjustment or margin; (e) recoverables, offsets and default risk; and (f) independent capital requirements. A reserve estimate cannot be judged adequate merely by matching a numerator from one framework to a denominator from another.
Reserve adequacy, scenario testing and cash timing
Adequacy has at least three empirical meanings. Ultimate-cost adequacy asks whether estimated total remaining claims are sufficient for the eventual cohort settlement. Timing adequacy asks whether the projected payment stream supports liquid claim settlement at expected dates. Reporting adequacy asks whether financial disclosures and valuation methods faithfully reflect the applicable framework. A method can meet one test yet miss another. For example, accelerating provider submission while leaving ultimate benefits unchanged increases near-term payouts, reduces observed paid development age factors and changes short-term liquidity exposure.
A runoff diagnostic compares a locked valuation estimate at with later actual payments plus any defensible residual ultimate for the same cohorts. Define development error at horizon . This is not a final error when the cohort remains open; the current ultimate itself has estimation uncertainty. Back-testing should report horizon, maturity, adjustments, seasonality, materiality and any change in claims definition. A favorable observed runoff can be produced by temporary payment delays and is not proof of adequate ultimate reserving.
Scenario analysis should stress separable drivers. An administrative delay shock moves reporting and payment distributions with little direct change to ultimate benefits. Medical inflation alters expected allowed costs, particularly for unresolved or unpriced episodes. A morbidity shock alters service incidence or severity; a benefit expansion changes contract coverage; a provider coding intervention changes billed and sometimes allowed amounts. The same nominal “+5 percent IBNR” perturbation is not a substitute for specifying which data-generating mechanism changed. Mathematically, changes to cost expectations and changes to observation delays operate through different terms in the reserve equation.
The IAA's ISAP 4 model standard addresses actuarial services connected with IFRS 17, while local standards and financial-reporting rules establish obligatory practice in individual jurisdictions. Stress documentation should identify responsible assumptions, their source and approval, sensitivity to timing and financial variables, and reconciliation to booked liabilities. Under the U.S. ASOP 5 disclosure regime, explicit adverse-deviation provisions and other relevant assumptions require suitable communication.
Governance of reserve decisions
An actuary’s selected estimate is mediated by finance, claims operations, data engineering, legal interpretation and governance bodies. Finance owns the booked figure and disclosures; claims operations explain processing events; administrators provide event-level records; the actuary measures future obligations using those facts. Responsibility should be documented for data cut-offs, assumption changes, reserve adjustments and final approval. Governance becomes especially material when a projection changes immediately after a system migration or provider contract renegotiation.
Reserve sufficiency is also sensitive to the ability to pay. A reported but unsettled claim might be economically certain yet fall due sooner than a newly incurred but unreported hospital episode. Aggregate ultimate estimates obscure cash-flow concentration in time. Management liquidity forecasts should use settlement timing distributions and plausible severe-runoff scenarios, with recoveries timed separately. In a comprehensive assessment, capital stress reflects variability around expected cash flows while the reserve addresses the expected liability and the framework-specific adjustment. Neither one can be substituted mechanically for the other.
Data, Measurement and Validation
Minimum reproducible data architecture
A valuation-grade claims file needs service/incurral date, insurer receipt date, first adjudication, effective and accounting posting dates, original and corrected claim identity, covered member exposure, provider and product classification, billed amount, allowed amount, member cost share, insurer paid amount, recoveries, denials, reversals, negotiated rate version and reporting snapshot. Additional tables establish contract benefits, enrollment eligibility, claim-handling expenses and reinsurance or risk-transfer arrangements. The actual fields available will vary; the audit asks whether the chosen estimator remains identifiable from the available subset.
Every field needs a data dictionary, source system, time zone, unit, net/gross sign convention, version and ownership. Administrative changes can mimic statistical development. For example, a newly adopted claims system may backdate service dates while posting new adjustment records in current months. If a historical extraction is overwritten, the resulting triangle can be impossible to reproduce. A versioned ledger or reliable as-of data snapshots support genuine reconstruction of each prior valuation.
The ASOP 23 data quality requirements govern actuarial selection, review and treatment of known deficiencies in U.S. practice. Its guidance addresses relevance, sufficiency, data provenance, questionable values and material limitations. A data review is distinct from a financial audit, and uncertainty induced by data limitations must be disclosed. The ASOP 56 modeling standard provides a parallel framework for model validation, documentation and use. These standards anchor the process without proving that a given dataset is clean.
A useful automated data control suite checks duplicate transactions; missing service dates; claims paid before their assigned service period; inconsistent first-report timestamps; eligibility before policy inception; negative payment adjustments without reference claims; unexplained changes in adjudication duration; unusually large remittances; and cohort aggregate reconciliation. Findings must be prioritized for materiality, not merely counted. Excluding suspicious observations can itself bias reserve estimation, especially when high-severity claims are more likely to require correction.
Selection, calibration and out-of-time tests
A primary model should be compared with at least one alternative sensitive to different weaknesses. Paid development estimates are checked against reported/case emergence where data support it. Chain-ladder results are compared with expected-cost or BF results for immature cohorts. Count-severity models can identify compositional shifts hidden in aggregate payment factors. For each comparison, the analyst records periods, sample maturity, weighting, selected development tail and any externally imposed assumptions.
Out-of-time validation freezes the historical information available at a prior valuation date , calculates the estimate using only that vintage, then compares with the subsequently observed cohort runout. The split must be by reporting date and not random claim row: random holdouts leak later development and create unrealistic performance metrics. The principal metrics can include weighted absolute error, reserve bias, squared prediction error, interval calibration and error conditional on cohort maturity. Reporting should preserve scale and direction; a pooled average absolute percentage error can be distorted when an individual cohort has a near-zero ultimate.
For stochastic predictions, evaluate whether stated coverage levels approximate observed proportions over many genuinely prospective or out-of-time cohorts. Examine process, parameter and tail uncertainty separately and consider correlations between cohorts due to calendar-year shocks. Mack’s standard error under chain ladder is valuable within its assumptions; if cohorts are dependent or claim inflation shifts the process, its nominal uncertainty may omit material risk. Bootstrap residual independence assumptions and distributional tails also require testing rather than default acceptance.
Model sensitivity is reported for alternative lag curves, selected tail, recent cohort exclusion, seasonality, large-claim treatment, paid-versus-allowed definitions, benefit changes and provider mix. A reasonable selected estimate should be accompanied by explanation of the plausible range and the decision significance of deviations. Repeated sensitivity calculations do not independently establish causal effects; they reveal dependence of estimates on assumptions.
Identification problems and research boundaries
Claims observation is a selection process. An insurer sees claims that providers or patients submit and its own decisions to accept or deny them. It does not directly observe all clinically delivered services. Historical paid records include decisions by payers and providers, so they do not provide an unbiased sample of medical need. A model seeking liability must follow covered benefits and enforceable obligations, not a counterfactual measure of total medically necessary care.
International data comparisons intensify this distinction. OECD (2025) documents different financing arrangements for care; reporting units can be national government programs, health insurers, hospitals or households. National expenditure datasets describe financing flows and are generally not service-cohort insurer payment triangles. Genuine cross-national comparison requires mapping institutional liabilities, benefit boundaries, paid and incurred definitions and administrative lag measures before interpreting ratios.
A final validity distinction concerns published standards versus full legal implementation. A model regulation, an international model standard or a conceptual actuarial paper does not by itself establish what a given insurer must book in a particular year. Actual local regulatory version, company legal form, contract portfolio and accounting standard must be documented. In this monograph the stated requirements are attributed to their official issuing bodies and their intended scopes; no unstated jurisdiction-wide applicability is inferred.
International Synthesis and Scientific Conclusions
Comparative institutional findings
The U.S. health-claims actuarial literature, NAIC health financial-statement architecture, IFRS 17, European Solvency II and Australian HPS 340 converge on a basic substantive problem: already incurred claims generate obligations whose full amount and timing remain uncertain at valuation. They specify different boundaries, measurement adjustments and governance duties. Their agreement is therefore in the need to estimate uncertain prevaluation obligations, while precise accountancy and prudential treatment is framework-specific.
U.S. ASOP 5 guidance centers methodological practice for estimating incurred health and disability claims, with explicit attention to development and projection methods. The NAIC health annual-statement development template demonstrates the institutional relevance of both paid and incurred emergence. IFRS 17’s incurred-claims liability distinction frames the cash-flow obligation within insurance contract measurement. Solvency II’s probability-weighted best estimate and APRA’s specific private health liabilities supply different prudential structures. The substantive comparative conclusion is that the same insurer’s modeled future payments can be mapped into different required reporting quantities without changing the underlying medical event history.
Three causal mechanisms distinguish health claim reserving from a generic payment-tail exercise. First, utilization and illness incidence determine the insured services generating covered events. Second, provider behavior, coding, claims submission and insurer adjudication determine visibility and payment delay. Third, contract pricing and provider reimbursement determine the monetary value of those events, sometimes after initial reporting. A model that combines all three into historical link ratios may predict acceptably under stability, but it cannot reliably attribute a sudden movement in the result. A decomposed model requires richer information while improving diagnosis of why the reserve changed.
Original analytical synthesis
The evidence supports an ordered estimation workflow. Establish contract-defined incurrence and the population bearing risk; reconcile event, report and payment records; estimate total ultimate on consistent cohorts; separate already reported unpaid obligations where event information permits; characterize uncertainty; transform expected cash flows to the relevant regulatory or accounting measurement; and validate against independent subsequent runout. This is an analytical synthesis of established scientific practices, not the announcement of a new theory, principle or doctrine.
The central interpretive claim is that reported claims are observations generated by an administrative process. Estimation error is produced both by stochastic medical costs and by the institutional mechanism through which those costs become visible. An administrative delay change that alters triangle completion factors without altering total medical services is a particularly clear example. Conversely, a provider price increase can raise ultimate claims while reporting speed is stable. These mechanisms require separately documented evidence to support a causal attribution.
Credibility should therefore be interpreted conditionally. Larger aggregated volumes reduce sampling volatility for a stable mechanism, while increased aggregation can conceal heterogeneity in provider reporting or benefit design. Hierarchical statistical approaches can share information across groups when exchangeability is credible. Where heterogeneity is pronounced, narrower segmentation and explicit expected-cost priors can improve interpretability, although they may increase estimation variance. A final estimate needs a defensible compromise supported by out-of-time evidence.
Limits, open scientific questions and warranted implications
The monograph establishes mathematical relationships and interprets verified source texts; it does not estimate an actual insurer’s reserve, test a real health plan database, or infer a globally representative claims-lag duration. A synthetic demonstration checks arithmetic and exposes assumptions, while empirical calibration and real-world predictive success remain distinct. Specific model performances can only be claimed after historical-vintage tests with available portfolio data.
Several continuing research questions follow directly: how to estimate the invisible tail when provider behavior changes at the same time as medical costs; how to propagate dependence between service incidence, severity, reporting and settlement; how to identify coding-intensity shocks from price changes; and how to validate insurer-level models when privacy or administrative constraints restrict event-level data. Further international work requires unit-consistent repositories and careful identification of the legal bearer of claims obligations.
The supported practical implication is a demand for transparent reconciliation and repeated retrospective testing. A financial statement's IBNR caption, a payer's service-cohort payment triangle and a prudential technical provision have different definitions. A scientifically adequate analysis provides a documented bridge between them, rather than assuming a common label proves a common measurement.
Conclusion
Health insurance IBNR estimation measures the expected future cost of covered medical events incurred before valuation but not yet visible in claims reporting. Its correct institutional use also requires consistent assessment of reported-unpaid and subsequently developing claims, because the broader incurred-unpaid obligation encompasses all expected remaining payments for the same events. Total unpaid claims estimated from paid development cannot be equated automatically with pure IBNR.
The essential scientific results are fourfold. Incurrence boundaries and transaction chronology determine what is being measured. Development triangles and expected-cost methods can recover ultimate values under explicit, testable stability assumptions, with the Bornhuetter–Ferguson formulation providing a controlled prior for immature cohorts. Stochastic analysis and runoff validation are needed to quantify prediction uncertainty and detect systematic reserve errors. Finally, internationally comparable economic cash flows must be reconciled separately to the governing accounting and prudential valuation framework.
The numerical appendix makes these relationships reproducible using fully synthetic data. Its purpose is methodological transparency. The resulting international synthesis integrates original historical and regulatory source reading with the established reserving literature while avoiding unsupported empirical extrapolation and any claim of a newly created actuarial theory. The research identifies a disciplined basis for actuarial measurement, supervisory examination and further empirical investigation.
Technical Appendix: Reproducible Synthetic Estimation
Synthetic cohort design and selected factors
All values here are invented educational monetary units and are not insurer data. Three service-month cohorts are observed at the same valuation cut-off. Their most recent cumulative paid values are July: at age 3; August: at age 2; September: at age 1. The illustration assumes preselected age-to-age factors , and , with the last factor representing all tail development. These are assumed inputs and are not statistically fitted from the three displayed balances.
| Service cohort | Latest age | Paid observed | Age-to-ultimate factor | Projected ultimate | Projected total unpaid |
|---|---|---|---|---|---|
| July | 3 | 920.00 | 1.050 | 966.00 | 46.00 |
| August | 2 | 720.00 | 1.260 | 907.20 | 187.20 |
| September | 1 | 410.00 | 2.205 | 904.05 | 494.05 |
| Total | — | 2,050.00 | — | 2,777.25 | 727.25 |
The independent check is ; observed cumulative paid is ; hence total projected unpaid is . Every row satisfies ultimate minus paid equals unpaid. This is a paid-runoff projection. Pure IBNR cannot be separated from reported-unpaid from these three paid values alone.
Alternative expected-ultimate calculation
Assume an independently selected prior ultimate for August of and September of . The paid completion proportions are and . The Bornhuetter–Ferguson formula gives August unpaid , ultimate ; September unpaid , ultimate . Rounded to two decimals, the two estimates are and , compared with paid-chain-ladder estimates and .
The difference arises because BF assigns the un-emerged proportion to an independent prior rather than multiplying the small initial payment. The example does not demonstrate that one method has superior real-world accuracy. It demonstrates precisely which assumption drives the result. Selection between them requires past valuation vintages, credible exposures and validation against actual future payments.
Synthetic sensitivity and classification controls
If the aggregate projected unpaid of is assumed to rise uniformly by 6 percent solely due to unanticipated claim settlement costs, the sensitivity is , giving before rounding. This is a scenario, not an observed inflation effect or a substitute for a cost-by-service-class projection. If payment timing changes without ultimate amounts changing, the undiscounted total remains while short-term cash-flow needs and possibly present value change.
Suppose a separate, verified open-claims inventory yields of expected unpaid for already reported prevaluation events, with no overlapping entries, on the same benefit and gross/net basis. Then pure unreported . This identity is conditional on the inventory's reliability and completeness. A booked case reserve of that excludes IBNER would not by itself justify the same classification. The example therefore separates mathematical reconciliation from empirical identification.
A rigorous independent recomputation should verify all displayed factors, row-level ultimate differences, prior completion ratios, rounding conventions and aggregate cross-footing. For a real portfolio, the same tests also require deterministic data extraction, scenario versioning, out-of-time replay and source-system reconciliation.
References
Actuarial Standards Board (2017). Actuarial Standard of Practice No. 5: Incurred Health and Disability Claims. Revised March 2017; effective September 1, 2017. Actuarial Standards Board (2017)
National Association of Insurance Commissioners (2017). Health Insurance Reserves Model Regulation, Model No. 10, Section 2, Claim Reserves. NAIC Model Laws, Regulations, Guidelines and Other Resources. National Association of Insurance Commissioners (2017)
National Association of Insurance Commissioners (2025). 2025 Annual Statement Blank—Health, Underwriting and Investment Exhibit, Part 2C, Development of Paid and Incurred Health Claims. National Association of Insurance Commissioners (2025)
National Association of Insurance Commissioners (2025). 2025 Health Annual Statement Instructions, Unpaid Claims Analysis (U & I Exhibit). National Association of Insurance Commissioners (2025)
Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality. Effective April 30, 2017. Actuarial Standards Board (2016)
Actuarial Standards Board (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Actuarial Standards Board (2019)
Bornhuetter, R. L., and Ferguson, R. E. (1972). The Actuary and IBNR. Proceedings of the Casualty Actuarial Society 59, 181–195. Bornhuetter, R
Mack, T. (1993). Distribution-free Calculation of the Standard Error of Chain Ladder Reserve Estimates. ASTIN Bulletin 23(2), 213–225. DOI: 10.2143/AST.23.2.2005092. Mack, T
Schnieper, R. (1991). Separating True IBNR and IBNER Claims. ASTIN Bulletin 21(1), 111–127. DOI: 10.2143/AST.21.1.2005404. Schnieper, R
Renshaw, A. E., and Verrall, R. J. (1998). A Stochastic Model Underlying the Chain-Ladder Technique. British Actuarial Journal 4(4), 903–923. Renshaw, A
England, P. D., and Verrall, R. J. (2002). Stochastic Claims Reserving in General Insurance. British Actuarial Journal 8(3), 443–518. DOI: 10.1017/S1357321700003809. England, P
Wüthrich, M. V., and Merz, M. (2008). Stochastic Claims Reserving Methods in Insurance. Wiley Finance. ISBN 9780470723463. Wüthrich, M
International Accounting Standards Board / IFRS Foundation (2023 effective). IFRS 17 Insurance Contracts. Issued standard overview. International Accounting Standards Board / IFRS Foundation (2023 effective)
IFRS Foundation (n.d.). IFRS 17 Insurance Contracts—Key Terms. Supporting implementation materials. IFRS Foundation (n.d.)
European Parliament and Council (2009, as amended). Directive 2009/138/EC (Solvency II), Article 77, as displayed in the EIOPA Single Rulebook. European Parliament and Council (2009, as amended)
Australian Prudential Regulation Authority (2023). Prudential Standard HPS 340: Insurance Liability Valuation. In force July 1, 2023. Australian Prudential Regulation Authority (2023)
Centers for Medicare & Medicaid Services (2020). Fact Sheet: Service Use among Medicaid & CHIP Beneficiaries Age 18 and Under during COVID-19. Centers for Medicare & Medicaid Services (2020)
Organisation for Economic Co-operation and Development (2025). Health at a Glance 2025: OECD Indicators. Chapter: Health Expenditure by Financing Scheme. OECD Publishing. DOI: 10.1787/8f9e3f98-en. Organisation for Economic Co-operation and Development (2025)
International Actuarial Association (2019). International Standard of Actuarial Practice 4: IFRS 17 Insurance Contracts. Approved November 21, 2019. International Actuarial Association (2019)
International Association of Insurance Supervisors (current online edition). Insurance Core Principles and ComFrame, particularly ICP 14 Valuation. International Association of Insurance Supervisors (current online edition)
National Association of Insurance Commissioners (2026). Accounting Practices and Procedures Manual, as of March 2026, Appendix A, Health Insurance Reserves. National Association of Insurance Commissioners (2026)