Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Paid versus Incurred Claims Development

Abstract. The comparison of paid and incurred claims development is an investigation of how an insurer's observed settlement cash flows and its evolving assessments of reported claim obligations convey information about ultimate claim cost. Paid losses accumulate through transactions. Incurred losses, as used in traditional reported-claim development triangles, combine cumulative payments with outstanding case estimates and can rise or fall as claim estimates change. Their respective triangles therefore encode distinct operational processes and respond differently to reporting lags, settlement acceleration, inflation, litigation, case-reserving conventions, portfolio mix, recovery recognition, and claim-handling interventions. This monograph establishes precise data identities, examines deterministic and stochastic estimation methods, compares independently developed ultimate-loss indications, and analyzes how paired information may be reconciled without assuming that agreement constitutes validation. A completely disclosed synthetic four-origin-year example yields paid-based ultimate claims of 1,568.11 and incurred-based ultimate claims of 1,829.17 in common hypothetical units, a divergence of 261.06. Sensitivity analysis shows why a case-reserve strengthening or an accelerated first payment can change development indications even with no demonstrated change in underlying claim severity. The investigation connects actuarial literature on chain-ladder uncertainty, Bornhuetter–Ferguson estimation, Munich Chain Ladder, and paid–incurred Bayesian methods with contemporary professional standards and international reporting distinctions. It concludes that paid–incurred divergence is evidence about the interaction of claims maturation, case estimation, and payment practices; a defensible ultimate estimate emerges from coherent definitions, independent diagnostics, explicit assumptions, tested alternative methods, and transparent communication of remaining uncertainty.

Keywords: Paid losses; incurred losses; reported claims; case reserves; claims development triangles; loss development factors; chain ladder; unpaid claims; IBNR; IBNER; Munich Chain Ladder; paid–incurred chain; stochastic reserving; settlement speed; reserve validation; actuarial estimation.

Introduction

An insurance claim can generate several sequences of observations. The loss event occurs, the claimant reports it, the insurer establishes an initial estimate of the unpaid amount, the claim passes through investigation or litigation, interim payments are made, and the final settlement is recorded. An actuarial valuation typically observes these processes at a limited accounting date. Ultimate claim cost, the total amount ultimately attributable to the selected cohort and valuation scope, must be estimated from that incomplete history. Cumulative paid losses and reported incurred losses provide two closely related but genuinely different observation channels.

The need to distinguish these channels is central to the Casualty Actuarial Society's reserving text, which consistently separates reported claims, paid claims, and estimated unpaid claims (Friedland, 2010). The distinction also matters institutionally: ASOP No. 43 identifies the estimand, the covered losses and expenses, gross or net treatment, and disclosure conditions before an unpaid-claim estimate is interpreted (Actuarial Standards Board [ASB], 2011). Accordingly, a numerically coherent exercise begins by specifying what each observation means and what the ultimate-loss estimate is intended to represent.

The scientific question developed here is how paid and reported-incurred views can produce different estimates of the same ultimate claim cost, how those differences should be diagnosed, and under what assumptions reconciliation becomes informative. The study treats claims-development triangles as observations of transactions and human or organizational estimates subject to change over time. It combines algebraic decomposition, a reproducible synthetic data experiment, literature-based method comparison, and a structured evaluation of model validation. All numerical examples are constructed for demonstration. They are not insurer data, calibrated loss forecasts, or evidence of market-wide performance.

The historical scholarship establishes several distinct approaches. Bornhuetter and Ferguson (1972) proposed an expected-loss framework addressing immature accident periods, as discussed in The Actuary and IBNR. Mack (1993) developed distribution-free prediction-error calculations for chain-ladder reserves, addressing uncertainty conditional on the method's structure in the original standard-error paper. Quarg and Mack's Munich Chain Ladder studies the joint paid–incurred information problem; Merz and Wüthrich's paid–incurred chain model later gives it a Bayesian stochastic treatment. These research lines illuminate different parts of a common estimation problem; their assumptions and output measures need to be compared before results are combined.

Case Reserve Changes

Movement in estimated obligations

Case outstanding is an evolving estimate of the remaining cost of claims already reported to the insurer. Its movement contains both information about expected severity and changes in administrative practice. An increase may reflect new medical evidence, an unfavorable court decision, a change in defense-cost expectations, a large-loss threshold being crossed, or a formal strengthening exercise. A decrease may follow a final settlement below an earlier estimate, recovery of an amount previously counted, reinsurance reassessment, or revised probability of an adverse outcome. Aggregate development combines these varied components and may obscure distinct trajectories in closed, open, and reopened claims.

The algebra clarifies the relevant question. For any cohort, the difference between successive reported incurred observations is the sum of paid transactions and the change in case outstanding. A case reserve can be released when a claim settles even when its final settlement amount was already anticipated. Thus, the measurement of incurred development depends on both the insurer's estimate of remaining claim severity and its policy for moving amounts between paid and case outstanding. If claims administrators begin setting stronger initial reserves at the first report, early incurred factors can fall in a historical comparison; older cohorts with weak initial case reserves often required larger upward revisions at later ages. A mechanical application of the older factors to the newly strengthened cohorts would risk overstating ultimate cost.

The reverse phenomenon is equally important. If new claim handlers record systematically lower early case estimates, reported incurred may appear more favorable at a common maturity before its later catch-up. The apparent improvement may be a procedural delay in recognizing claim cost. ASOP No. 23 on data quality directs actuarial attention to data review, reasonableness and appropriate disclosure of data limitations; the standard's precise applicability depends on the professional engagement (ASB, 2016). A data-quality test should therefore include case-reserve policies and valuation continuity, not solely missing fields and arithmetic totals.

Case adequacy, claims practices, and induced bias

There is no generally valid inequality under which paid-based ultimates must exceed incurred-based ultimates, or vice versa. Paid methods may be distorted by slow or accelerated settlement, while incurred methods may be distorted by inadequate, excessive, or discontinuous case estimates. It is entirely possible for both methods to be biased in the same direction: a newly emerging class of expensive claims might be both unreported and unpaid, giving neither observation system a complete record. Such cases require external information from exposures, claim notifications, changes in policy terms, medical or legal cost trends, and comparable cohorts.

An operational review should identify the date and magnitude of any change in reserve-setting authority, payment automation, defense strategy, settlement negotiation policy, litigation backlog, claims-system migration, or reinsurance administration. Changes in the administrative environment deserve particular care when a development pattern seems to alter around a calendar date rather than consistently with accident or report age. The case study in Evaluating Reserves in a Changing Claims Environment develops this general concern in the actuarial literature (Gonwa, Weinstein, & Halpert, 2001).

One should also separate adjustments based on observed new facts from assumptions chosen to make paid and incurred estimates agree. An actuarial strengthening that arises from identifiable new case evidence is a substantive measurement change. An ad hoc case adjustment introduced merely to force two triangles into agreement is an artifact of calibration. Its use would erase the very discrepancy from which an analyst might otherwise learn.

Reserve development versus claims cost development

Development factors are historical ratios of observations made at different dates. For reported incurred, such ratios are affected by revisions of earlier estimates; they are not necessarily ratios of an underlying physical loss accumulation process. Consequently, a decline in a selected incurred age factor can reflect improved case-reserve accuracy, a changing proportion of already-reported claims, or systematic releases. It does not by itself imply a decline in ultimate accident cost.

For paid claims, development factors encode a payment schedule as well as loss emergence. Insurance with frequent rapid settlement may show high early paid emergence even when ultimate severity is stable. Long-tail coverage with late settlement may show persistent large payment factors despite a mature reported-claim count. Actuarial inference should distinguish the amount ultimately due from the timing with which that amount becomes paid. This separation is essential for comparing ultimate-loss methods with projected cash flows for liquidity or discounting purposes.

Development Triangle Comparisons

Matched triangles and chain-ladder indications

Arrange cohorts by row and development age by column, with observations available only where valuation history permits. Define paid age-to-age factors for each development interval using a volume-weighted estimator:

f^jP=∑i∈AjPi,j+1∑i∈AjPi,j,f^jJ=∑i∈AjJi,j+1∑i∈AjJi,j,\widehat f^{P}_j=\frac{\sum_{i\in A_j}P_{i,j+1}}{\sum_{i\in A_j}P_{i,j}}, \qquad \widehat f^{J}_j=\frac{\sum_{i\in A_j}J_{i,j+1}}{\sum_{i\in A_j}J_{i,j}},

where AjA_j includes only rows observed at both ages. This paired-column eligibility rule matters: calculating the numerator over more origin years than the denominator creates an invalid factor. The chosen estimators are conventional volume-weighted chain-ladder factors, not observed probabilities of payment or reporting. They need reasonably stable underlying development behavior to serve as predictive factors. Their implicit weighting can concentrate the estimate in a few large cohorts, especially when reported-incurred balances have been altered by a major case-reserve adjustment.

For each cohort with latest age k(i)k(i), define age-to-ultimate factors F^kP=∏j=kω−1f^jP\widehat F^P_k=\prod_{j=k}^{\omega-1}\widehat f^P_j and F^kJ=∏j=kω−1f^jJ\widehat F^J_k=\prod_{j=k}^{\omega-1}\widehat f^J_j, including an explicitly selected tail factor if development continues beyond the last observed column. Corresponding indications are U^iP=Pi,k(i)F^k(i)P\widehat U_i^P=P_{i,k(i)}\widehat F^P_{k(i)} and U^iJ=Ji,k(i)F^k(i)J\widehat U_i^J=J_{i,k(i)}\widehat F^J_{k(i)}. Both are estimates of the same ultimate loss amount only if the definitions, cohort assignment, and completion conditions agree.

Reproducible numerical study

Consider the following entirely synthetic cumulative observations. Monetary quantities are abstract units. Years are relative origin cohorts, and ages are months since the origin convention. Blank cells are unobserved at the valuation date, not zeros. The small sample deliberately places downward reported-incurred development in the oldest cohort.

Origin cohort Paid 12 Paid 24 Paid 36 Paid 48
AY1 100 200 270 300
AY2 120 230 305 —
AY3 150 280 — —
AY4 180 — — —
Origin cohort Reported incurred 12 Reported incurred 24 Reported incurred 36 Reported incurred 48
AY1 250 310 320 300
AY2 300 360 365 —
AY3 380 490 — —
AY4 600 — — —

Every reported-incurred observation is at least as large as the corresponding payment, consistent with nonnegative case outstanding. On the latest diagonal, case outstanding equals 0, 60, 210, and 420 respectively. For AY1 at age 36, paid is 270 and case outstanding is 50; at age 48, paid rises by 30 and case outstanding falls by 50, so incurred falls by 20. This example directly illustrates ΔJ=ΔP+ΔC\Delta J=\Delta P+\Delta C, rather than presuming all incurred observations increase.

Using only matched development pairs, the paid age factors are

f^12:24P=710/370=1.918919,f^24:36P=575/430=1.337209,f^36:48P=300/270=1.111111.\widehat f^P_{12:24}=710/370=1.918919,\quad \widehat f^P_{24:36}=575/430=1.337209,\quad \widehat f^P_{36:48}=300/270=1.111111.

The corresponding reported-incurred factors are

f^12:24J=1160/930=1.247312,f^24:36J=685/670=1.022388,f^36:48J=300/320=0.937500.\widehat f^J_{12:24}=1160/930=1.247312,\quad \widehat f^J_{24:36}=685/670=1.022388,\quad \widehat f^J_{36:48}=300/320=0.937500.

For calculation only, assume that 48 months is fully developed, so both tail factors equal 1. The assumption is particularly weak because only AY1 has a 48-month observation, and it must be challenged in an actual analysis. The resulting point indications, rounded to two decimals, are:

Origin cohort Latest paid Latest incurred Paid-based ultimate Incurred-based ultimate Paid-based unpaid Incurred-based unpaid
AY1 300 300 300.00 300.00 0.00 0.00
AY2 305 365 338.89 342.19 33.89 37.19
AY3 280 490 416.02 469.66 136.02 189.66
AY4 180 600 513.20 717.32 333.20 537.32
Total 1,065 1,755 1,568.11 1,829.17 503.11 764.17

The difference in ultimate indications is 261.06 units, the same as the difference between the two aggregate unpaid indications because paid to date is the common 1,065. It is not a measurement of a newly discovered loss or evidence that one method is objectively correct. AY4 contributes approximately 204.12 units, or about 78% of the divergence. This concentration identifies the youngest, most case-reserve-dependent cohort as the central object of subsequent examination.

A comparison should also inspect the ratio Ri,j=Pi,j/Ji,jR_{i,j}=P_{i,j}/J_{i,j} whenever the incurred denominator is positive. In the example the latest ratios are 1.000, 0.836, 0.571, and 0.300 from AY1 to AY4. These are paid proportions of reported incurred at their different maturities, so they naturally vary with age. The more revealing test is a comparison of each ratio with historical cohorts at the same age, followed by an inquiry into any persistent calendar-time break. Multiplication of recent paid–incurred ratios into ultimate factors without a defined probabilistic or operational link invites spurious certainty.

Factor selection and finite-sample uncertainty

The factor estimates above rest on only three, two, and one matched cohort pairs at the successive intervals. The oldest age factor is entirely based on AY1; it can be dominated by a single atypical large-loss closure or adjustment. Moreover, paid and incurred data share the same cash transactions, so method agreement cannot be treated as two independent confirmations of ultimate cost. The same events affect both sides of the identity, while the case outstanding balance adds a second information channel.

Mack's chain-ladder standard-error analysis provides a formal route for studying prediction errors subject to its specified development assumptions (Mack, 1993). England and Verrall's stochastic reserving survey reviews methods that represent uncertainty using stochastic and generalized-linear-model perspectives (England & Verrall, 2002). Neither reference justifies attaching a credible standard error to this miniature example without evaluating the available residual structure. Tail-factor estimation, inflation, changes in book mix, covariance across cohorts, and model selection can dominate uncertainty left outside simple within-triangle calculations.

In practical analysis, inspect individual link ratios, the stability of volume and simple averages, development by calendar period, very large claims, and the sensitivity of the selected factor to excluding one cohort. Reconcile exposure changes and incurred-limit changes where possible. A large amount of immature business magnifies the consequences of a small factor difference. An aggregate triangle that passes a conventional reasonableness check may still contain a shift in settlement practice, reporting patterns, or excess-layer severity hidden within a mixed portfolio.

Payment and Reporting Speed

Timing mechanisms

Payment speed is the rate at which the eventual claim cost becomes a recorded payment. Reporting speed is the rate at which claim events enter the insurer's reported-claim system. Case recognition speed is the rate at which expected costs of reported claims enter outstanding case estimates. These processes can respond to different economic, legal and administrative forces. A settlement initiative may accelerate payments while leaving the expected ultimate loss unchanged. A new claim intake platform may accelerate reporting and initial case creation without changing ultimate insured events. Court delay may defer both settlement payments and the final reestimation of liability. The distinct channels matter especially when an origin-year cohort spans a significant operational change.

A useful diagnostic decomposition records the date of accident, first notice, first case reserve, each subsequent case change, each payment, claim closure, and any reopening. One can construct cohort-specific reporting-delay distributions, lag from report to first payment, payment completion curves, and distributions of open-case severity, with proper treatment of censored observations and incomplete older cohorts. For paid claims, early or late cash settlement changes age-to-age factor profiles. For incurred claims, recognition of liability and case reserve adequacy govern a separate part of development. The same ultimate loss can support quite different observed paths.

Sensitivity experiments

Return to synthetic AY4. Its latest paid amount is 180, reported incurred is 600, and case outstanding is 420. Suppose the claims department records an additional 100 of case outstanding, with no extra cash paid. Reported incurred becomes 700. Holding the historical comparison cohorts and their selected factors fixed, the AY4 incurred ultimate indication becomes 836.87 instead of 717.32, an increase of 119.55. This mechanical change follows from the incurred cumulative development factor of about 1.19553; it does not establish that the claim portfolio's actual future cost increased by 119.55. The alternative in which case outstanding falls by 100, producing incurred 500, gives an incurred indication of 597.77. These are stress tests on reported observations, not forecasts of what policyholders will receive.

Now consider a separate scenario in which AY4 has paid 40 earlier than it otherwise would, paid increasing from 180 to 220, while reported incurred remains 600 because case outstanding falls from 420 to 380. Using the identical historical paid factors, its paid-based ultimate indication increases from 513.20 to 627.24, by 114.04. In this counterfactual, accelerated settlement has changed the paid method's input and its predicted ultimate even though the stipulated underlying ultimate cost did not change. It is a direct illustration of model sensitivity to development speed. Because AY4 is the most recent row, altering its latest observation does not change the matched-pair factors estimated from older cohorts; a modification to an older row would require recomputation of the factor set.

The different responses emphasize that paired development methods are not mechanistic descriptions of a known ultimate loss process. Each projects historical relationships into the future. Changing the timing or the estimation convention can alter the projected amount even when the future cost remains fixed under a specified counterfactual. Conversely, emerging severity inflation can change actual ultimate cost before payments or case estimates fully reflect it. A well-designed analysis identifies which component is being shocked and holds the remaining components consistent.

Long-tail, short-tail, and heterogeneous portfolios

In short-tail property claims, payment may occur soon after claim notification, and case outstanding may run down rapidly. Recent catastrophe events can nevertheless create reporting, adjusting and rebuilding bottlenecks, while repair-price inflation can introduce further severity development. In bodily-injury liability, construction-defect liability or other litigation-intensive classes, disputes and delayed settlement can produce large reporting and payment lags; case reserving policies may exert a persistent influence on the incurred triangle. In workers' compensation, medical cost escalation, wage benefits, lump-sum settlements and statutory schedules may produce distinct payment trajectories. In proportional and excess-of-loss reinsurance, ceded paid and incurred patterns are influenced by reporting from underlying cedants, attachment points, treaty terms, reinstatements, recovery accounting, and reporting thresholds.

A high paid-to-incurred ratio is meaningful only in relation to a relevant maturity, claims mix, and case-reserve convention. A near-zero early paid-to-incurred ratio can be normal for slow-settling litigation and potentially anomalous for immediately settled consumer property claims. The analytical unit should be a reasonably homogeneous cell of business and observation process, rather than whatever aggregation was easiest to export. When data sparsity makes stratification costly, the loss of statistical precision must be weighed against the bias introduced by pooling dissimilar processes.

Reconciling Ultimate Estimates

Decision sequence and alternative estimators

Reconciliation begins with measurement, not averaging. First establish the common origin basis, recoveries, expenses, currency, exposure, cutoff dates, and reported-incurred definition. Second reproduce the algebraic identity J=P+CJ=P+C from transaction-level or frozen valuation data. Third diagnose which ages and cohorts generate the divergence. Fourth test the claims-process explanations before adjusting any ultimate estimate. Fifth construct alternative indications using different assumptions and compare their predictions with subsequent observed development.

An illustrative comparison set may include paid and reported-incurred chain ladder; an expected-loss or Bornhuetter–Ferguson estimate; an exposure-based expected loss ratio; a case-outstanding development analysis; and a frequency–severity analysis separated into report counts, closure rates, and outstanding-claim severity. The Bornhuetter–Ferguson IBNR paper remains a foundational source for combining a priori cost expectations with emerging experience (Bornhuetter & Ferguson, 1972). In an immature cohort, a credible independent premium- or exposure-based estimate may prevent an anomalously high early paid ratio from being projected without restraint. That expected estimate itself needs scrutiny for pricing-level adequacy, inflation, changes in cover, and claim frequency.

For a simple blended result U^=wU^P+(1−w)U^J\widehat U=w\widehat U^P+(1-w)\widehat U^J, the weight ww has substantive meaning only if it arises from a defined decision criterion, such as empirically evaluated predictive performance, case data reliability, or a reproducible uncertainty model. Choosing w=0.5w=0.5 because estimates differ is a presentation convention, not an independently supported statistical estimator. Moreover, paid and incurred errors are correlated. If VPV_P and VJV_J are the variances of the two unbiased indications and CPJC_{PJ} their covariance, the variance of a linear blend is

V(w)=w2VP+(1−w)2VJ+2w(1−w)CPJ.V(w)=w^2V_P+(1-w)^2V_J+2w(1-w)C_{PJ}.

Subject to the restrictive assumptions of unbiased forecasts, known second moments, and unconstrained linear weighting, the variance-minimizing paid weight is w∗=(VJ−CPJ)/(VP+VJ−2CPJ)w^*=(V_J-C_{PJ})/(V_P+V_J-2C_{PJ}) when the denominator is positive. This is elementary correlated-estimator algebra, not a new model for claims reserving. Its importance is interpretive: assigning equal weight without considering covariance or bias has no inherent optimality. If the weight is constrained to [0,1][0,1], the optimum is clipped to that interval. In actual reserving, squared-error optimality may not correspond to the financial reporting measure or risk appetite required by the engagement.

Joint paid–incurred methods and uncertainty

Joint approaches attempt to use relationships between paid and incurred observations rather than treat the two projections as independent forecasts. Quarg and Mack's Munich Chain Ladder study is a major example of explicitly relating the two development channels, with the 2008 Variance publication reprinting their 2004 work (Quarg & Mack, 2004/2008). Its title describes a reduction in the gap between projected amounts; reduction of a gap is not automatically evidence of reduced estimation error. A model must be judged by its assumptions, data suitability, and predictive behavior.

Merz and Wüthrich's paid–incurred chain method introduces a Bayesian stochastic construction that combines claims payments and incurred information and derives a predictive distribution for outstanding claim liabilities (Merz & Wüthrich, 2010). The one-year claims development extension addresses uncertainty in the annual claims development result within that framework (Happ, Merz, & Wüthrich, 2012). These results depend on the specified model structures, lognormal relationships, prior assumptions, and the quality of the claims histories. They do not imply that every paid/incurred dataset should be fit to the same joint model.

A further copula-based Bayesian extension by Peters, Dong, and Kohn (2014) examines more flexible dependence structures within paid–incurred claims models (Insurance: Mathematics and Economics, 2014). It illustrates that dependence between paid and incurred observations can be modeled rather than ignored. The relevant statistical burden increases: credible estimation now requires sufficient historical information to identify the dependence structure and validate predictive distributions. With a four-row example such as the one in this monograph, sophisticated parameter fitting would be unsupported.

The ChainLadder package vignette (Carrato et al., 2026) supplies an accessible technical description of established actuarial development methods, including paid–incurred techniques. A software implementation is useful for reproducibility after the practitioner has verified the observations, mathematical conventions and selected assumptions. This monograph's numerical tables were calculated directly from the displayed data and equations; no fitted implementation of the Munich, Bayesian PIC, or package-specific model is claimed.

Reconciliation with financial reporting

An undiscounted actuarial unpaid-claim indication is not automatically an amount suitable for any particular balance-sheet standard. The intended measure may be central expected value, a selected percentile, a booked amount subject to management judgment, or a discounted and risk-adjusted liability. ASOP No. 43's scope guidance requires that the measure and scope of an actuarial unpaid-claim estimate be identified; it distinguishes that exercise from the amount booked as reserves (ASB, 2011). Where applicable to professional work, the revised ASOP No. 20 on cash flows addresses the analysis of property/casualty cash flows and discounting, effective for covered work products issued on or after June 1, 2026 (ASB, 2025).

Within US statutory property/casualty reporting, NAIC annual statement instructions provide prescribed categories and reconciliation requirements for the reporting schedules; the 2026 edition carries its own schedule definitions and must be used for the relevant reporting period (NAIC, 2026). Historical reserve data can therefore be analyzed in a reported-incurred form while regulatory schedules separately require detail for paid amounts and other components. Under IFRS 17 terminology, insurance liabilities include measurement features not contained in P+CP+C, while Solvency II valuation rules add their own present-value and risk-margin framework. A cross-jurisdictional comparison requires mapping each measure explicitly to its legal and accounting basis. A reported-incurred triangle may inform valuation without equaling the final statutory or IFRS liability.

Claims Process Bias and Validation

Retrospective validation and diagnostic design

Validation is the test of whether an estimation process predicted future observations under the circumstances to which it was applied. A useful backtest chooses a historical valuation date, removes all information that was not then observable, reconstructs both triangles as they were at that date, estimates an ultimate amount using a documented procedure, and compares forecasts with subsequent development. If the full ultimate is still unknown, evaluate observable subsequent payments, claim-count development, settlement rates, and reserve movements alongside any later ultimate estimate. Using a later revised triangle while claiming it represents the earlier valuation introduces look-ahead bias.

For multiple historical valuations tt, an observed-versus-predicted development error can be defined for the amount paid over a held-out period hh:

ei,t,hP=(Pi,t+h−Pi,t)−ΔP^i,t,h.e^{P}_{i,t,h} = \bigl(P_{i,t+h}-P_{i,t}\bigr)-\widehat{\Delta P}_{i,t,h}.

A parallel incurred-developing-error measure compares the realized change Ji,t+h−Ji,tJ_{i,t+h}-J_{i,t} to its earlier prediction. These two errors measure different future observables and should be analyzed in their original units before aggregation into ultimate losses. Bias can be assessed by mean signed error conditional on age or business line; accuracy by absolute or squared errors, with care for scale. A calibration test for a stochastic method checks whether empirical held-out outcomes fall into forecast intervals at plausible frequencies. A model may predict unpaid cash flows acceptably while overstating reported incurred, or conversely. The validation target must match the business decision.

Diagnostic scrutiny should include triangular residual plots by origin, development and calendar period; link-ratio changes around known claims-system conversions; outlier impacts; case-reserve-to-paid ratios at constant maturity; number of open cases; time-to-report and time-to-close distributions; reopenings; frequency and severity; inflation and legal developments; gross and net reconciliation; and sensitivity of selected factors to large claims and tail assumptions. This should be accompanied by a written ledger of observed source data, exclusions, adjustments and supervisory assumptions. Both the absence of external validation data and a sparse triangle materially constrain conclusions.

An explicit process-change flag can be incorporated in analysis without treating it as proof of a change in ultimate losses. Consider a group of origin cohorts whose 24-month paid-to-incurred ratio becomes higher after implementation of automatic small-claim payments. That difference is consistent with accelerated cash settlement. It must be tested against the claim-size distribution, total reported claim count, and case balance changes; a simultaneous portfolio shift could generate a similar aggregate pattern. Historical comparison to cohorts under the old process cannot settle the issue on its own.

International comparability and governance

Actuarial practice is international, while statutory definitions, claim settlement institutions, reinsurance arrangements, and financial reporting bases differ. An insurer may report accident-year net paid amounts in one jurisdiction and calendar-year gross paid amounts in another. A case reserve may include defense costs in one system and exclude them in another. Claims arising under claims-made coverage may be grouped by notice date, and policy limits may be represented differently in claim systems. These distinctions can alter the shape and meaning of each triangle before any statistical method is applied.

For a professional engagement within its scope, the US data-quality standard and ASOP No. 43 provide an explicit framework for reviewing data limitations and communicating uncertainty. The actuarial-opinion standard imposes additional considerations for covered statements of opinion (ASB, 2024). Outside these US scope conditions, the relevant local professional, statutory and contractual requirements govern. Comparative scientific analysis can use these standards as clear benchmarks while preserving jurisdiction-specific applicability.

An international review file should preserve, for every valuation, the data dictionary, extract date, versioned transactions, currency and exchange-rate conventions, coverage segmentation, court and claims-process events, the exact code or formula used for each result, reviewer challenges, and the selected outcome with alternatives. Where a model uses licensed or restricted data, the audit trail may retain hashes, transformation steps and summary statistics without publicly exposing confidential claim details. Source provenance supports reproduction of results as authorized by the data's legal and contractual terms.

Research synthesis and authorial conclusions

Four substantive conclusions emerge from the combined mathematical, institutional and actuarial analysis.

First, reported-incurred development is inseparable from the insurer's historical case-reserving practice. The equation ΔJ=ΔP+ΔC\Delta J=\Delta P+\Delta C makes this unavoidable. Incurred triangles can contain early information about expected claim severity that paid amounts have yet to reveal, but that information is mediated by case estimates, assumptions and decision policies. Their value depends on how well those assessments predict subsequent settlement and how consistently they were made through time.

Second, payment development is simultaneously a signal about eventual loss and an operational schedule. Faster claims settlement can increase observed cumulative paid amounts at a given age and raise a chain-ladder ultimate indication without increasing the final payable loss. This occurs because factors estimated under one payment regime are being applied to observations from another regime. Payment-based estimates can be superior when case reserves are unreliable; they can be inferior when settlement speeds change sharply. The ranking follows evidence and circumstances rather than a permanent preference for one triangle.

Third, divergence between paid and incurred indications is most informative when localized to cohort, age, and claims-process mechanism. The synthetic study locates nearly four-fifths of the total divergence in AY4, whose latest outstanding case balance is also largest. That result suggests a targeted review of AY4 maturity, initial case reserve adequacy, expected reporting, and selected early development factors. Averaging away the divergence would discard the diagnostic information. The observed difference is a reason to investigate, not itself a selection rule.

Fourth, reconciliation becomes scientifically persuasive when competing methods make distinct, testable claims about future observed development and their common ultimate target. Professional judgment can weigh evidence from alternative methods, external cost expectations and case-process diagnostics. More complex Bayesian or multivariate methods add value when data and validation justify their extra assumptions. The core requirement remains unchanged: a forecast should identify its observation base, future estimand, uncertainty, and the conditions under which it would need revision. This analytical synthesis uses existing actuarial methods and publicly documented standards; it does not establish a new proprietary theory or conceptual system.

Conclusion

Paid and incurred claims development are complementary records of one insurance obligation viewed through transactions and changing estimates. Cumulative paid amounts observe settlement; reported incurred additionally incorporates case outstanding. The distinction creates a tractable accounting identity, but it also creates two different sources of bias when historical development relationships are projected into the future. Case-reserve strengthening, claims-system interventions, settlement acceleration, changing reporting lags, litigation, recoveries, and portfolio shifts can change observed development factors independently of the underlying ultimate claim cost.

The transparent four-cohort experiment demonstrates a 261.06-unit gap between independently selected paid and incurred chain-ladder ultimates, with most of the discrepancy arising in the youngest cohort. The case-strengthening and payment-speed counterfactuals show that a model indication can respond sharply to the timing or administrative measurement of a claim even when ultimate cost is not stipulated to change. These numerical results express the behavior of the stated synthetic dataset and selected assumptions; they cannot be generalized into empirical claims about insurer profitability or reserve adequacy.

A defensible scientific and actuarial resolution begins with common definitions and as-of data, traces the divergence to measurable cohorts and process changes, evaluates expected-loss and other alternative information, tests model predictions against later observed development, and presents the sensitivity and uncertainty of the adopted estimate. Models designed to combine paid and incurred information provide important further tools, including explicit stochastic dependence, but their reconciliation is informative only when supported by credible data, valid assumptions and a clearly defined estimand. The resulting decision remains tied to the actual insurance portfolio, applicable reporting framework and purpose of the valuation.

References

  • Actuarial Standards Board. (2011). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates (revised May 2011; effective September 1, 2007). Actuarial Standards Board. Official standard.

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revised December 2016). Actuarial Standards Board. Official standard.

  • Actuarial Standards Board. (2024). Actuarial Standard of Practice No. 36: Statements of Actuarial Opinion Regarding Property/Casualty Loss, Loss Adjustment Expense, or Other Reserves (revised March 2024). Actuarial Standards Board. Official standard.

  • Actuarial Standards Board. (2025). Actuarial Standard of Practice No. 20: Analysis of Property/Casualty Cash Flows, Including Discounting (revised December 2025; effective June 1, 2026 for covered work products). Actuarial Standards Board. Official standard.

  • Bornhuetter, R. L., & Ferguson, R. E. (1972). The actuary and IBNR. Proceedings of the Casualty Actuarial Society, 59, 181–195. CAS publication record.

  • Carrato, A., Concina, F., Gesmann, M., Murphy, D., Wüthrich, M., & Zhang, W. (2026). ChainLadder: Claims reserving with R [package vignette, version 0.2.21, dated February 18, 2026]. Comprehensive R Archive Network. Official package vignette.

  • England, P. D., & Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. Journal DOI.

  • European Union. (2009, as amended). Directive 2009/138/EC: Solvency II, Article 77 (calculation of technical provisions). European Insurance and Occupational Pensions Authority consolidated rulebook. Article 77.

  • Friedland, J. F. (2010). Estimating unpaid claims using basic techniques (CAS educational study note; publication record dated 2009, syllabus 2010). Casualty Actuarial Society. Official CAS study note.

  • Gonwa, C., Weinstein, S., & Halpert, A. (2001). Evaluating reserves in a changing claims environment. Casualty Actuarial Society Forum, Fall 2001. CAS Forum article.

  • Happ, S., Merz, M., & Wüthrich, M. V. (2012). Claims development result in the paid-incurred chain reserving method. Insurance: Mathematics and Economics, 51(1), 66–72. Journal DOI.

  • IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts: Key terms. IFRS Foundation. IFRS 17 terminology.

  • Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. Journal DOI.

  • Merz, M., & Wüthrich, M. V. (2010). Paid–incurred chain claims reserving method. Insurance: Mathematics and Economics, 46(3), 568–579. Journal DOI.

  • National Association of Insurance Commissioners. (2026). 2026 Property/Casualty Annual Statement Instructions (adopted June 2026). NAIC. Official annual instructions.

  • Quarg, G., & Mack, T. (2004/2008). Munich Chain Ladder: A reserving method that reduces the gap between IBNR projections based on paid losses and IBNR projections based on incurred losses. Variance, 2(2), 266–299 (2008 reprint of Blätter der Deutschen Gesellschaft für Versicherungs- und Finanzmathematik, 26(4), 597–630, 2004). Reprint and bibliographic record.

  • Schnieper, R. (1991). Separating true IBNR and IBNER claims. ASTIN Bulletin, 21(1), 111–127. CAS bibliographic record.

  • Peters, G. W., Dong, A. X. D., & Kohn, R. (2014). A copula based Bayesian approach for paid–incurred claims models for non-life insurance reserving. Insurance: Mathematics and Economics, 59, 258–278. Publisher article.