Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Individual Claim Reserving and Micro-Level Loss Development

Abstract. Individual-claim reserving estimates future insurance payments using the observable histories and characteristics of claims rather than relying exclusively on aggregate development triangles. This monograph examines the probability structure of occurrence, reporting, payment, case valuation, closure and reopening; the statistical consequences of observing claims at different maturities; the estimation of reported-but-not-settled and incurred-but-not-reported liabilities; and the reconciliation of claim-level forecasts with portfolio financial obligations. It synthesizes marked-point-process models, recurrent-event analysis, paid–incurred models, historical simulation, generalized regression, hierarchical methods and machine-learning approaches. Formal definitions separate the occurrence population from the claims currently visible to the insurer, clarify the difference between predictions of future cash flows and present accounting measurements, and show why development processes must be modeled on the information actually available at each valuation date. Two fully hypothetical numerical constructions illustrate expected outstanding amounts, accident-period reconciliation and predictive dependence. The analysis also develops a practical validation protocol based on historical evaluation dates, leakage-resistant features, competing aggregate benchmarks, calibration by maturity and subgroup, uncertainty decomposition and reproducible data lineage. The central finding is that granular data can improve conditional forecasting when reporting mechanisms, payment sequences, status changes and covariate histories are trustworthy and sufficiently stable. Information advantage does not guarantee predictive advantage: complexity, selection into the reported-claim population, case-estimate practices and structural changes can reverse the ranking of methods. A defensible valuation joins auditable event histories, probabilistic completeness, portfolio reconciliation and explicit judgment about model limits. The study is an original analytical synthesis of established scholarship and introduces no newly named theory, doctrine or proprietary statistical system.

Keywords: individual claims reserving; micro-level loss development; granular claim histories; marked point processes; reporting delay; settlement hazard; recurrent payments; reported but not settled; incurred but not reported; incurred but not enough reported; paid and incurred data; survival analysis; predictive distributions; case reserves; chain ladder; machine learning; actuarial data quality; historical backtesting; claims uncertainty.

Introduction

Every unpaid-claims estimate begins with a time boundary. A loss event has occurred, some information has reached the insurer and some transactions have been recorded, while subsequent payments remain uncertain. Traditional reserving summarizes this experience by origin and development periods. Individual-claim reserving begins from the same financial obligation but retains the histories that produced those summaries: the dates of occurrence and notice, multiple partial payments, interim case estimates, claim handling events, legal or medical milestones, recoveries and changes in settlement status. The scientific purpose is to project conditional future cash flows using that event information and then aggregate them into coherent estimates of the portfolio obligation.

The historical foundations are well established. Haastrup and Arjas (1996) described individual-claim development through a continuous-time Bayesian reserving approach, with reporting delays and partial payments represented as elements of a marked point process. Larsen (2007) studied an individual-claims reserving model that could accommodate settlement, business mix, seasonality and claim-size effects. Antonio and Plat (2014) examined micro-level stochastic loss reserving using reporting times, payment events and final settlement. These contributions established a statistical object richer than an array of cumulative paid amounts.

The literature subsequently extended the representation in several directions. Pigeon, Antonio and Denuit (2014) combined claim payments and incurred information in individual paid–incurred reserving. Godecharle and Antonio (2015) used claim markers and historical simulation of claim development. Wüthrich (2018) demonstrated machine learning on individual claims, while Gabrielli (2021) investigated a multi-period neural reserving model for already reported claims. Delonga, Lindholm and Wüthrich (2022) developed collective reserving with individual data to model both unreported and reported components. Differences in method, portfolio and validation data explain why conclusions about the superiority of one technique require qualification.

The object of this dossier is general-insurance claim development at a defined valuation date. It includes both reported claims and the claims that occurred but remain unreported. It addresses expected undiscounted future claim payments first, then explains how alternative purposes require changes in measurement. The principal related Science is Insurance Pricing and Reserving; adjacent areas include Actuarial Science, Claims Science, Insurance Statistics and Predictive Modeling, and Insurance Technology. Related issues such as solvency, accounting, reinsurance and claims governance enter where their assumptions materially affect the estimate. The methodology is a critical examination of source-supported models combined with explicit mathematical reasoning, constructed examples and a transparent implementation protocol. No insurer's confidential claim files were obtained, and the numerical illustrations are hypothetical.

Claim-Level Data Structures

The object of estimation and the information boundary

Let the valuation cutoff be T, the insured event time for claim i be O_i, the reporting time be R_i, and its jth signed payment transaction be (t_ij, Y_ij). A positive Y_ij denotes an outgoing indemnity or expense payment under a chosen sign convention, while a negative value may denote a recovery or reversal if recoveries are included in the same ledger. Let C_i(t) denote its cumulative paid amount through time t, and let S_i(t) denote its recorded claim state, including open, closed, reopened, disputed or other operational classifications. The stochastic target for a specified portfolio is the sum of signed payments after T attributable to events O_i ≤ T, under the intended definition of covered claims and expenses. Stating this boundary excludes post-T occurrences from the standard unpaid-claim target. A separate prospective cost estimate would require a different exposure horizon.

The information set F_T includes everything genuinely observable by T: reported claim identifiers, their transaction and status histories, known policy and exposure characteristics, reliable external covariates and perhaps a snapshot of the case reserve. Reported-but-not-settled (RBNS) claims satisfy R_i ≤ T and retain an uncertain future payment exposure. Incurred-but-not-reported (IBNR) claims satisfy O_i ≤ T but R_i > T or are never reported within the applicable reporting window. Incurred-but-not-enough-reported (IBNER) commonly denotes development of existing case assessments beyond amounts already carried; its usage requires an explicit operational definition. Not all closed claims have zero expected future payments because closure can be reversed or subsequent expenses paid.

For a gross undiscounted future-payment target, define X_i(T,H) as the sum of claim i's future payments during (T,H], where H may be a finite horizon or the end of the modeled runoff. The central conditional estimate is R(T,H) = E[sum_i X_i(T,H) | F_T]. This is a mathematical expectation for one carefully specified target, not by itself a booked reserve, risk margin, regulatory capital requirement or cash available for settlement. ASOP No. 43's guidance on unpaid claim estimates specifically requires the intended measure, gross or net basis, expenses, dates and scope to be identified for work within its US applicability.

Event histories, identifiers and valuation snapshots

A robust claim-level system uses at least four connected tables. The claim table records immutable claim identity, occurrence and report dates, product, coverage, origin period, claimant and exposure references. An event table records status transitions and their effective and recording times. A transaction table records actual cash movements, reserve adjustments, recoveries, reversals and relevant currency attributes. A periodic snapshot table records case estimates and other fields as they were known at each historical valuation date. Time-sensitive auxiliary tables may hold the evolution of litigation status, medical assessments, counsel assignment or policy interpretation. Each table must preserve both the economic event timestamp and the time at which the insurer first knew or recorded the event.

This dual-time design prevents hindsight in training. If a settlement was posted in March but backdated to an accident in January, a model trained at a February cutoff may use the accident date only when the event had already become known in February. The definitive future outcome and the historical feature view should come from separate reproducible extracts. A current database column labeled 'latest case estimate' cannot stand in for its historical value when backtesting past forecasts. Mature outcomes must be joined after the historical feature snapshot is frozen.

Claim IDs require careful deduplication. One occurrence may generate multiple coverage files, claimants, lawsuits, bodily injuries or insurer ledger identifiers. Conversely, one claim file may represent several insured events or layers of indemnity. The modeled unit therefore must be specified: event, claim file, claimant, occurrence-coverage pair or payment stream. A change in the operational unit without corresponding historical mapping can create artificial movements in frequency and severity. An insurer that migrates claims systems should reconcile split and merged records using immutable crosswalks before fitting any statistical model.

Measurement quality, provenance and exposure linkage

Actuarial data quality is inseparable from reserving validity. The ASOP No. 23 Data Quality standard directs covered practitioners to consider suitability, internal consistency, sufficiency, limitations, available alternatives and appropriate review of data. Its distinction between review and formal audit matters: checking patterns, missingness and reconciliation is necessary scientific work, while asserting that every source transaction has been independently audited requires different evidence. Claim-level modeling amplifies errors in statuses and event timestamps that aggregate triangles sometimes obscure.

Core checks compare the sum of transactions with general-ledger paid amounts, investigate negative payments and stale case estimates, detect dates outside feasible orderings, identify claims that were closed before their recorded first report and determine how reopened claims are coded. Temporal ordering exceptions are investigated rather than automatically deleted, since operational systems may contain retrospective corrections. Payment currencies and inflation need stable normalization. Changes to deductibles, policy limits, reinsurance attachment, claims-handling authority or settlement law should be represented as effective-dated data where material. The denominator of exposure by origin period must reconcile to underwriting records when estimating unreported frequency.

Aggregate triangles as transformations of claim histories

An origin–development paid triangle can be obtained from individual transactions by selecting origin year a and development age d, then summing amounts within the corresponding cumulative time window. Formally C_(a,d) = sum over i with origin a of sum over j with t_ij ≤ end(a,d) of Y_ij. The triangle is thus a projection of a richer event ledger. Under a common information cutoff, consistent payments and stable claim definitions, the latest observed accident-period paid amount must equal the corresponding aggregation of claim histories. This equality is an accounting control and does not depend on model choice.

Aggregation removes claim identity and many dependencies. Two portfolios with the same paid triangle can have very different distributions of open-claim sizes, reporting delays, settlement statuses or future payment possibilities. The granular approach can exploit this difference if the lost attributes are informative for future outcomes and measured consistently. Conversely, pooling across heterogeneous claims can sometimes stabilize predictions when there are few historical examples for a very narrow subtype. The classic aggregate comparison remains important. Mack's (1993) chain-ladder prediction error analysis gives an important reference for assessing uncertainty under particular distribution-free assumptions, but it does not require replacing all models with triangles or guarantee independence across granular claims.

Payment and Settlement Processes

Occurrence, reporting and observation

The reporting mechanism links the occurrence population, which is partially invisible at T, to the observed claim population. For each claim with covariates z, the reporting delay D = R − O has a conditional distribution. Its survival function S_D(u|z) = P(D>u|z) describes the probability that a claim remains unreported u units after occurrence. A hazard λ_D(u|z) expresses the instantaneous report rate conditional on no previous report. For a cohort of occurred claims with incidence process N, an estimate of the unreported count at T requires both a model of incidence and the probability that reporting has not occurred by T. Reporting delays alone, estimated only from already reported claims without accounting for truncation, can understate the slowest reports.

A simple discrete illustration uses an origin cohort whose occurrence count has conditional expectation 100, with 80% expected to have been reported by T. The expected remaining unreported count is 20, under the assumed cohort size and reporting probability. If the expected ultimate payment on a future reported claim is 3 units, a frequency–severity approximation produces 60 units of IBNR expected payments. This example presumes the claims that report late have the same conditional expected cost as those used to calibrate 3; if larger claims report more slowly, or some events never become compensable claims, that assumption must be revised. Nothing in observed claim counts alone identifies the invisible population without occurrence exposure and reporting assumptions.

Haastrup and Arjas's (1996) marked-point-process formulation provides a coherent framework for occurrence, reporting and subsequent development; Antonio and Plat (2014) modeled reporting and recurrent payment events in a liability setting. Their use of detailed event times is scientifically significant because grouping to annual development ages can conceal concentration around notification, legal milestones or seasonal closures. Calendar effects must nevertheless be distinguished from accident age and company process changes to avoid spurious hazard patterns.

Payments as marked events and settlement as a state process

Following report, the next payment date and amount constitute a marked event: its timestamp is the event, its monetary amount is the mark. One model uses a conditional payment counting process N_i(t) with intensity λ_i(t|H_i(t−),z_i), where H_i is the claim's observed history immediately before t. The corresponding payment-size distribution may be conditional on time, state, past amounts, coverage, severity or exposure. If m_i(t) is the conditional mean signed amount of a payment at t, a schematic expected future cash flow for an open claim is the integral from T to H of E[λ_i(t)m_i(t) | F_T]dt. The expectation must encompass uncertain future states and paths; simply multiplying current hazard and current mean can fail when their evolution is dependent.

A competing model discretizes monthly time into transitions between states: unreported, reported/open, closed, reopened and terminated. In a month, the claim can remain open, receive one or several payments, close or reopen after a prior closure. Transition probabilities can be estimated by generalized regression, survival analysis or neural networks. Status and cash flow should be modeled together when settlement changes the payment opportunity or a payment triggers closure. Lindholm and Zakrisson (2021) explicitly include claim closing and reopening dynamics in a collective reserving model, demonstrating why 'closed' should be understood as an operational state rather than necessarily an absorbing physical condition.

Payment timing matters independently of ultimate cost. A claim expected to pay 100 in one month differs in liquidity and present value from a claim expected to pay the same 100 over five years. The undiscounted expectation remains 100 in both cases, but a present-value estimate changes with an appropriate discount curve and relevant financial reporting basis. The actuarial calculations must distinguish nominal trend, expected calendar inflation and discounting. The current ASOP No. 20 framework concerns discounting of property/casualty claim estimates for covered actuarial work, whereas the exact accounting measurement depends on its own standard and jurisdiction.

Likelihood, heterogeneity and dependency

For independently observed conditional event processes with suitable regularity and noninformative censoring, a point-process log likelihood includes a sum of log intensities at observed events and subtracts the integrated event intensity over each at-risk interval. For a single event type, ℓ(θ) = sum_j log λ(t_j|H(t_j−);θ) − integral λ(t|H(t−);θ)dt, with additional terms for payment marks and observed transitions. Reporting and payment processes may require joint or staged estimation to reflect selection into the observed claim sample. Heterogeneity can be modeled with frailty terms, random effects or hierarchical parameters. Dependence arises when a large payment changes the chance of closure, when litigation prolongs development or when multiple claims respond to a shared shock.

Identifiability governs how far the model can be elaborated. Short records rarely identify separate late-reporting tails, very large future payments, reopening hazards and court-delay effects without structural assumptions or external observations. The choice between parametric and nonparametric forms trades interpretability and extrapolation against flexibility. High-dimensional conditional methods may produce apparently precise forecasts on common claims while remaining unstable for rare catastrophic or systemic patterns. Reporting the estimation sample, at-risk periods, censoring rules, model specification and extrapolation range is as essential as reporting fitted coefficients.

Reserve components and the scope of the target

Let P_T be paid amounts to T for pre-T occurrences, F_T^R future payments on already reported claims and F_T^U future payments on claims unreported at T. The ultimate amount for that fixed occurrence population satisfies U = P_T + F_T^R + F_T^U if the cash-flow ledger and scope are consistent. Conditional expectations give E[U|F_T] = P_T + E[F_T^R|F_T] + E[F_T^U|F_T]. Expected payment on reported-but-not-settled claims is only one component of unpaid liabilities. If some currently closed claims can reopen, their future payments belong in the reported component despite administrative closure. An incurred-paid comparison is useful but depends on whether 'incurred' means paid plus case estimate, an actuarial ultimate estimate or a financial-reporting carrying amount.

The scope must make explicit loss adjustment expenses, defense costs, salvage, subrogation, deductibles, allocated versus unallocated expenses and reinsurance recoverables. Forecasts based on net paid transactions cannot automatically be compared with gross triangles, and mixing signed recoveries with indemnity payments changes both the conditional means and the residual tails. ASOP No. 43's gross-versus-net and recoverable provisions illustrate the significance of defining this target before model fitting. Multiple target estimates may be necessary when accounting, claims management and capital modeling have different decision needs.

Individual Claim Predictors

Statistical baselines and conditional regression

A granular analysis should begin with interpretable baselines. Conditional average future payment by line, development age and open/closed status often gives an informative starting point. A generalized linear or generalized additive model can relate expected future payment to claim age, prior paid amount, injury type, geography, limit, legal representation, case estimate and claims-handling state. The functional form and treatment of zero payments depend on the response: an event indicator can use a binary link, a positive amount can use an appropriate severity distribution, and an event count can use a Poisson or overdispersed count specification. A two-part model predicts whether a payment occurs and its positive size conditional on occurrence. This factorization has practical value because many claimant–month observations contain no cash flow.

Covariates must correspond to the prediction date. Some are determined at policy inception, such as coverage and deductible; some arise at loss occurrence; others become known only when reported; and some are time-varying claims-process signals. A model that includes settlement date or final negotiated value in its features will exhibit excellent in-sample fit while making an invalid historical prediction. Missingness indicators can be informative, but a feature missing because it was not yet known cannot be imputed from the future. The suitable response horizon also matters: modeling expected payment next month repeatedly requires coherent simulation of subsequent status and features, whereas modeling total outstanding directly may ignore timing but can be simpler and less vulnerable to compounding error.

Conditional regression does not imply causal interpretation. A positive association between lawyer representation and ultimate severity does not establish that legal representation itself caused the increment; the underlying injury may influence both. Claims routing, specialist allocation and litigation status are frequently responses to unobserved severity. Their inclusion can improve forecasts on stable operational regimes, yet a redesign of claims workflows may invalidate the learned relationship. Judgment about intervention-induced distribution changes belongs in every production reserving assessment.

Paid–incurred and case-development information

Paid losses record completed transactions; case estimates record contemporaneous judgments about remaining or ultimate obligations according to an insurer's operational convention. Their relationship may be informative, particularly for large open claims that have incurred substantial exposure but few disbursements. However, paid and case values are not independent measurements. Case estimates may be revised because of payments, reporting guidance, staff changes or legal evaluation; their process can be strongly dependent on future settlement behavior and business strategy.

The individual paid–incurred chain method of Pigeon, Antonio and Denuit (2014) explicitly incorporates payments and incurred information into individual reserving. Its reported favorable comparison concerns the studied general-liability data and model specification. It does not establish a universal superiority claim. At the aggregate level, Merz and Wüthrich (2010) presented a paid–incurred chain approach for combining paid and incurred triangles. Comparing these approaches illuminates the different roles of case estimates: the aggregate model uses their summarized development, while a granular model can condition on current reserve level, earlier reserve revisions and claim descriptors.

Consider a reported claim with 40 already paid and a case estimate recorded as 90 of remaining exposure. A purely case-based outstanding estimate is 90. If a statistical model predicts an expected 60 of future payments because a stable, well-documented claim subtype tends to settle below such case assessments, the model's estimate is 60. Neither number acquires authority simply by being labeled actuarial or operational. Both should be evaluated against comparable historical outcomes and the stated estimation target. The difference of 30 is a reserve reconciliation signal requiring explanation; it is not an extra liability to be added to the predicted 60. If the case field instead denotes ultimate incurred of 90 inclusive of the 40 paid, its implied remaining amount is 50, and treating it as 90 remaining would double count paid claims.

The status of case reserves as predictors requires careful historical reconstruction. Where claims administrators implement new case reserve guidelines, a step change in case levels can reflect procedure rather than changes in underlying losses. In such settings, an apparent predictive improvement may exploit the historical manual practice and fail after policy reform. A useful diagnostic is a model restricted to objective pre-valuation fields compared with a model that adds case estimates and assessor activity. An improvement should persist on genuine out-of-time valuation cuts and remain interpretable through the changes in claim handling.

Historical simulation and machine learning

Historical simulation seeks future paths from claims with similar markers and development histories. Godecharle and Antonio (2015) developed reserving by detailed conditioning based on claim-specific markers and simulated future payments. Its principal strength is the ability to construct a predictive distribution from observed development analogues rather than only a single point estimate. Selecting comparable claims still requires judgments about maturity, coverage, inflation, tail severity and the stability of the historical population. Simulation can only reproduce a tail that has adequate information or has been extended by explicit assumptions.

Decision trees, boosting and neural networks can learn nonlinear interactions among claim attributes and past payment histories. Wüthrich (2018) gave an explicit regression-tree reserving application. Gabrielli (2021) introduced a neural model for reported claims with special treatment of incomplete histories of past payments; the paper reports a proof of concept on synthetic and real portfolios and explicitly leaves a full uncertainty analysis outside its central treatment. The qualification matters: the reported empirical performance is evidence about those settings, not a substitute for estimation of predictive distributions in a new portfolio.

Delonga, Lindholm and Wüthrich (2022) describe six linked reserving networks for separate elements of IBNR reporting, RBNS payment and closure, recoveries and IBNR amounts. The architecture demonstrates how a flexible collective forecast can be built from interpretable stochastic tasks. Its modules must still align on shared state definitions, signs, exposure cohorts and future paths. Multiplying a count prediction by a severity mean, for instance, produces an expected amount only when the severity mean is conditional on the same selection mechanism and covariates as the count model.

The individual claims history simulation machine of Gabrielli and Wüthrich (2018) offers a research environment for synthetic claim paths calibrated to empirical characteristics. Such generators are valuable for implementation stress tests and for checking whether a method recovers a known simulated target. They cannot establish general empirical accuracy because a model may perform well on simulated histories that resemble its own assumptions. Validation should therefore distinguish controlled synthetic experiments, retrospective testing on real historical claims and genuine post-deployment performance.

Partial pooling, uncertainty and operational change

Individual claims are heterogeneous, but training a separate model for each narrow category is often statistically unstable. Hierarchical regression shares information across categories while permitting differences in level or development pattern. Pooling should follow scientifically defensible relationships: small claims in one coverage may share some characteristics with comparable claims in another, while catastrophic bodily injury settlements may require distinctive tail behavior. The model should disclose which parameters are global, line-specific, origin-specific or claim-specific and how uncertainty in those parameters enters the prediction.

Expected payment is not the whole predictive problem. A probability distribution over future paths is necessary for a range, percentile, stress scenario or reserve variability analysis. A full prediction combines process uncertainty, parameter uncertainty and relevant model uncertainty. Process risk remains even when every parameter is known; parameter uncertainty grows with thin and volatile data; and model risk concerns the consequences of an inadequate description of the process. Correlated inflation or court decisions can generate portfolio variability beyond the sum of independent claim variances. Models fitted to individual observations are not automatically independent-claim models.

Operational change must be treated as an observable part of the insurance process. Claims closure drives, outsourcing, digital notifications, pandemic restrictions, judicial reforms and changes in the mix of deductibles can change the hazards observed at a given development age. A model may require calendar effects or segmented periods; however, unconstrained calendar factors can absorb true economic trends and become unstable under extrapolation. Scenario analysis and deliberate comparison of mature and recent cohorts provide a more transparent account of such limitations than a single goodness-of-fit number.

Micro-to-Aggregate Reconciliation

Additivity of conditional expectation

Let R_i denote the future signed payment for claim i that was reported by valuation time T, and let R_U denote the total future payment attributable to events that occurred before T but remain unreported. A portfolio expected unpaid amount is

[ \widehat{R}T=\sum{i\in\mathcal{R}_T}\operatorname{E}(R_i\mid\mathcal{F}_T)+\operatorname{E}(R_U\mid\mathcal{F}_T). ]

This identity follows from the linearity of conditional expectation provided that the claim and unreported components partition the same payment universe. Dependence between claim payments does not invalidate the expected-value sum. Dependence is critical for predictive variance and tail quantiles, which generally do not equal the simple sum of marginal variances and quantiles. Reconciliation therefore begins at the expected-value ledger but extends to the dependence assumptions when reserves are expressed as distributions.

At the origin-period level, the sum of future expected payments on claims originating in period a gives the forecast outstanding amount R_a. Adding actual cumulative paid to date P_a gives projected ultimate U_a = P_a + R_a on a consistent gross or net, nominal or discounted basis. An aggregate chain-ladder ultimate can then be compared with the micro-level ultimate for the same occurrence population and cutoff. Differences should be decomposed by origin period, development age, line, severity band and claim status. They should not be removed by mechanically forcing each individual payment to reproduce a chosen aggregate estimate without disclosing the calibration.

Synthetic claim-level calculation

Consider a fully hypothetical cohort at a valuation date, with amounts in thousands of monetary units. Four claims have been reported. Claim A has 40 paid and 60 of expected future payments. Claim B has 10 paid and 25 expected future. Claim C is closed, has 50 paid and zero expected future payment under the deliberately simplified assumption of no reopening. Claim D has no payment and 90 expected future payment. These estimates are statistical conditional means, not separately added case reserves.

Reported claim Paid to date Expected future paid Expected ultimate
A 40 60 100
B 10 25 35
C 50 0 50
D 0 90 90
Reported subtotal 100 175 275

Suppose a separately fitted reporting-and-severity component predicts 50 of future payments for unreported claims from the same occurrence population. The combined unpaid amount is 175 + 50 = 225, and the cohort's expected ultimate is 100 + 225 = 325. If an aggregate benchmark predicts an ultimate of 310, the difference is 15 and must be explained through reporting counts, severity, status transitions, development factors or data definition. It is inappropriate to describe that difference as observed model superiority, since this example contains invented inputs and no empirical testing.

Now assume an identified portion of claim D's future gross indemnity will be recovered from a reinsurer. If future ceded recovery has a conditional expectation of 30 under a specified treaty and collectibility assumption, the net outstanding expected amount becomes 225 − 30 = 195. The gross ultimate remains 325, and the net future paid reserve is 195. Whether previous payments already included recoveries affects the reconciliation of net ultimate; it cannot be inferred from the four-claim table. A reconciliation schedule must separately show gross losses, expected ceded recoveries, other salvage/subrogation, remaining insurer payments and the treatment of expenses.

Unreported claims and predictive dependence

For a simple compound unreported-claim model, let N_U have conditional Poisson mean λ and let future amount Y per unreported claim have conditional mean μ and second moment E(Y²). Under independent severities conditional on the parameters and independent N_U, E(R_U)=λμ and Var(R_U)=λE(Y²). If λ is itself uncertain, total predictive variance also includes variation of λμ across posterior or parameter scenarios. Correlated severity shocks, clustering of loss events and common inflation factors add further covariance terms. Reported claims can also be dependent upon the unreported process through shared calendar conditions, which a coherent portfolio simulation may need to reproduce.

A small calculation illustrates the difference between conditional mean and reserve variability. Suppose λ=2, every unreported claim has a future amount of 25 and the only stochastic element is Poisson count. E(R_U)=50, E(Y²)=625 and Var(R_U)=1,250, giving a standard deviation of approximately 35.36. This is a simple process-risk illustration under fixed assumptions. It does not include uncertainty in λ, unknown claim sizes, common shocks or heavy tails. If the original four reported claims also have random future outcomes, the total variance must incorporate their joint uncertainty and any covariance with unreported payments.

The expected-value reconciliation 225 remains valid even when individual liabilities are dependent. What changes is the risk of substantial adverse development. A reserve at a high percentile cannot generally be computed by adding the corresponding percentile of each claim: joint scenarios and dependence determine the aggregate distribution. A realistic program may simulate occurrence counts, reporting delay, claim states, payment times and amounts simultaneously or use a defensible approximation whose omitted dependencies are documented.

Paid, incurred, gross and net reconciliation

A complete reconciliation involves three independently traceable layers. The transaction layer links claim-level paid cash flows to aggregate paid triangles and ledgers. The projection layer links predicted future cash flows to aggregate unpaid estimates, with matched dates, cohorts and currencies. The financial layer maps actuarial estimates into accounting and regulatory measurement bases. Case incurred is generally a ledger construct affected by claims operations; actuarial projected ultimate is a conditional statistical estimate; a reporting liability follows the applicable measurement rules. These quantities should not be labeled interchangeably.

The IFRS 17 accounting framework measures groups of insurance contracts using fulfillment cash flows, including timing and risk adjustments, within its scope and effective jurisdiction. Its official key-term explanations distinguish the liability for incurred claims and the risk adjustment for non-financial risk. A raw expected-undiscounted individual-claim amount may be an input to fulfillment cash flows, but it is not the full financial reporting liability by itself. Accounting grouping, discounting, expenses and risk adjustment must be reconciled separately.

The claims process supplies the cash-flow timing distribution necessary for discounting, but discounting introduces interest-rate and economic assumptions external to individual claim status. A nominal expected payment at time t can be multiplied by an appropriate discount factor under the chosen measurement basis; doing so on already discounted case estimates would misstate the result. Reinsurance recoverables and claim adjustment expenses may also follow different timing and accounting treatments. Each transformation should be shown as a bridge from the common event ledger to the reporting measure rather than embedded silently into fitted settlement probabilities.

Censoring and Case Development

Delayed entry, administrative censoring and long tails

Individual-claim data contain several distinct observation problems. Reporting delay creates incomplete observation of the occurrence population: claims that have not reported by valuation time are invisible as claim records, although their exposures may be known. Administrative right censoring affects the observed subsequent development of reported claims: their future transactions and final settlement have not yet occurred. Delayed entry arises when the analyst observes a system only after some claims have already been developing, so past event histories may be partial. Left truncation is a more precise statistical concept when inclusion in the sample requires an event or survival condition before entry, and its applicability must be established from the actual extraction mechanism rather than inferred from an incomplete database.

For a claim reported at r and observed until T, the observed follow-up is T−r. Comparing late-reported young claims with early-reported mature claims without adjusting for follow-up can confound reporting dynamics with settlement speed. Censoring may be informative when claims disappear from an extract because of transfer to another administrator, portfolio sale or business rule rather than merely reaching a common valuation cutoff. The conditional independent-censoring assumption should be tested against operational descriptions and available status data. The correct exposure denominator for a payment hazard is the at-risk history, with appropriate restrictions for state, duration and operational visibility.

Long-tailed liability payments present an especially difficult extrapolation. A training set with fifteen years of observation may still contain unsettled cases with potentially material future litigation or medical costs. Fitting a highly flexible function to the observed ten-year payment frequencies does not determine its behavior at twenty years. Extreme claims may violate ordinary parametric severity assumptions. The model should document the extent of observed tail maturity and examine sensitivity to plausible changes in duration and ultimate cost, with specialized legal or claims expertise where appropriate. The point is substantive: the absence of late payments in a young dataset cannot itself establish that the modeled tail is negligible.

Open, closed and reopened trajectories

A claim's state should be interpreted as part of its development path. Administrative closure may follow final indemnity payment, denial, withdrawal or a zero-payment outcome; sometimes it is followed by supplemental payment, appeal, litigation or reopening. A single terminal closed flag discards these differences. A multistate representation can encode open, closed, reopened, litigated and settled states, but it should not proliferate states beyond what the data can reliably distinguish. Transition probabilities may depend on time since report, time since prior closure, amount already paid and coverage-specific conditions.

The collective claim-openness model by Lindholm and Zakrisson (2021) is particularly relevant to this issue: it explicitly retains closing and reopening in a process related to aggregate reserving. Its relevance is structural rather than a claim that its assumptions apply to every portfolio. For instance, a large property catastrophe portfolio may experience coordinated administrative closure and subsequent supplemental adjustment after litigation. A workers' compensation portfolio may include periodic payments whose administrative state has a different interpretation. The same numerical closure hazard can have different future-cash-flow implications in different claim systems.

Case reserve development deserves its own time-series discipline. An assessor may update a case estimate after receiving documents, after paying an installment or after observing a medical prognosis. A sequence of case estimates is informative only if versions and update timestamps are preserved. Claims with frequent case changes may appear more severe because administrators focus on complex cases; this selection should be modeled or described. In historical tests, case estimates are predictors only from the snapshots actually available at a valuation. A final case estimate assigned retroactively to prior snapshots creates future-data contamination, often improving apparent historical results by an implausible amount.

Leakage, fairness and distribution shift

Data leakage can enter through explicit outcome fields, through derived features computed on a full development history, and through subtle extraction choices. A final claim status as of the database export date is not the status as of a historical valuation date. A total paid amount covering the entire claim lifetime cannot be used as paid-to-date. Case reserves restated after a system migration may contain information absent at the original cutoff. Randomly splitting claim-month rows between training and evaluation allows identical claims and future months of their history to appear on both sides and may inflate performance. Historical-origin and valuation-time splits, with controlled development windows, offer a better test of deployable performance.

Predictive attributes also raise governance questions. Geography, medical history or legal representation may be both predictive of payment amounts and correlated with social disadvantage. Estimating a legitimate obligation requires attention to policy terms and actual claim outcomes; statistical convenience is insufficient justification for using sensitive or unstable proxies. A historical settlement dataset may encode systematic claims-handling disparities. A model trained on those outcomes can reproduce disparities even when its loss is numerically low. Assessment should examine subgroup calibration, differential error and sensitivity to operational practices while respecting local privacy and anti-discrimination rules. Such review does not turn the reserve model into an underwriting score or claims-denial algorithm.

Distribution shift occurs when the joint relationship between predictors and payments changes. It can arise from inflation, case law, portfolio mix, deductibles, policy wording, claims fraud, automation or extraordinary events. A model with precise conditional predictions on its development sample may fail in a later regime. The selection of historical windows and discounting or inflation assumptions must follow the purpose of valuation. Calendar effects are necessary when experience materially changes; however, extrapolating an arbitrary recent calendar intercept as if it were a permanent causal trend can produce spurious future growth. Sensitivity analysis should consider coherent alternative regimes rather than only tiny perturbations around fitted coefficients.

Professional and financial reporting contexts

Individual-claim reserving is a modeling approach; its outputs serve institutions with distinct obligations. In the United States, ASOP No. 43 supplies professional guidance on developing unpaid property/casualty claim estimates within its stated scope. Among its considerations are the intended measure, multiple methods, external conditions, data, recoverables and reasonableness. The ASOP No. 23 standard addresses suitable data and known limitations. These standards are professional requirements for relevant covered services, not international statutes automatically binding all insurers.

The IFRS Foundation's IFRS 17 standard overview identifies a financial reporting framework effective from 2023 in applying jurisdictions. It uses risk-adjusted present-value concepts and defined insurance-contract groupings. A claim-level statistical mean is one source of prospective cash-flow information, and its implementation must be reconciled to the contract boundary, expense categories, financial assumptions, risk adjustment and reporting presentation relevant to that framework. The precise cash-flow scope and discount method require review of the actual accounting standard and local implementation. The public standard overview is not a substitute for a licensed or authoritative full-text accounting interpretation when a technical reporting decision turns on a detailed paragraph.

Professional reporting should preserve model version, data cutoff, accounting date, review date, intended gross/net basis, assumptions on future payments and long tails, material operational changes and any overrides. The difference between a modeled distribution and a management-selected booked amount must be recorded. An internal mathematical expectation cannot be presumed to satisfy the purpose of regulatory capital or a prudential stress scenario, and a percentile from a simulated reserve model should identify its confidence level, risk horizon, included risks and dependence assumptions.

Validation and Reproducibility

Historical valuation designs and baselines

A useful reserving validation exercise simulates the information available at a succession of past valuation dates. For each historical cutoff T_k, build the frozen feature tables, estimate the model using only earlier data permitted by the proposed live process, and forecast amounts payable during a fixed future horizon or through sufficiently mature runoff. Compare predictions to subsequently observed payments, keeping an explicit record of tail incompleteness. Retraining for each cutoff should follow realistic calendar constraints and model development rules. It is unsound to optimize a model using outcomes that would have been unknown at T_k and then report the fitted performance as a retrospective independent test.

A minimum benchmark set includes a paid chain ladder, an incurred-based or paid–incurred comparator where coherent case data exist, a simple frequency–severity reserve for IBNR and a transparent individual-level conditional average by claim age and status. Mack (1993) analyzed prediction uncertainty for chain ladder; the benchmark remains valuable because the new data-intensive methods have to justify the operational cost and complexity they introduce. Where legacy triangles are small or heterogeneous, the comparison should be reported by origin period and segment, not only as one overall percentage. One large claim can dominate a holdout year; results should show sensitivity to its presence and to alternative severity stratifications.

Evaluation dates must be representative of actual business use. A single year with unusually favorable closure patterns cannot support a general claim about performance. Rolling-origin assessment, where successive training cutoffs precede later held-out valuation periods, is preferable. Keep the latest chronological period as a final evaluation that was never consulted during feature or hyperparameter selection. For very long tails, distinct train, validation and test windows may still contain common development histories; claim IDs, origin dates and release timestamps are essential to construct appropriate partitions.

Forecast accuracy and distributional calibration

Point accuracy can be measured at multiple aggregation levels. Mean absolute error of a reserve forecast measures average absolute differences; squared error emphasizes large misses; weighted scores can reflect the actual portfolio significance of lines and cohorts. The denominator of percentage error must be defined carefully because a near-zero realized amount can make a relative error arbitrarily large. A model should report individual-level diagnostics and portfolio-level outcomes. Good aggregate performance can conceal compensating errors in reported versus unreported components, while accurate individual ranking may not imply a well-calibrated aggregate mean.

For a predictive distribution F_k of future payment Y_k, a proper scoring rule such as continuous ranked probability score evaluates the full distribution when its assumptions and computation are appropriate. Coverage of nominal prediction intervals should be examined by claim age, line, severity band and calendar period. A 90% interval is calibrated only if the rate of realized future outcomes inside such intervals is appropriate over comparable repeated predictions; no single observed realization proves calibration. Tail quality requires separate attention to large losses and dependence, since marginally calibrated claim quantiles can aggregate poorly when common shocks are omitted.

A practical scorecard includes total reserve bias by origin period, payment-horizon accuracy, IBNR count and amount calibration, RBNS cash-flow timing, closure and reopening probability calibration, observed versus expected recoveries, interval coverage and sensitivity to calendar assumptions. Record actual out-of-sample sample sizes and uncertainty around improvements. Do not declare a model more accurate solely from a lower training error or one unreplicated holdout year. Compare paired performance on the same valuation cutoffs, business cohorts, gross/net definitions and realized development horizon.

Sensitivities, auditability and governance

Stress testing should vary reporting hazard, claim frequency, settlement speed, future payment severity, tail inflation, case-estimate reliability, reopening propensity and recoverable collectibility. Changes should reflect plausible mechanisms rather than arbitrary percentage shocks. A shock to settlement delay can change cash-flow timing even when undiscounted ultimate remains constant; a severity inflation shock can change ultimate even if payment count does not. When assumptions are correlated, joint scenarios avoid underestimating aggregate uncertainty. The analyst should document which stresses represent observationally supported variation and which are constructed judgmental scenarios.

An auditable implementation preserves the source extracts or hashed snapshots, variable dictionary, inclusion and exclusion rules, development cutoff semantics, code revision, model configuration, parameter estimates, forecast distributions and aggregation bridge. The workflow should be reproducible on a frozen dataset. Privacy protection and access controls must be applied to claimant identifiers, medical data and legal records, while de-identified or synthetic artifacts can facilitate methodological replication. Model monitoring after deployment should track actual payments against forecasts and compare recalibration needs with earlier approval assumptions.

A reproducible synthetic exercise is useful as a unit test: regenerate claims under a known occurrence and development mechanism, freeze successive cutoffs, refit the selected model, project future payment paths, and compare conditional expectation and predictive interval coverage to the known generator. Gabrielli and Wüthrich's claims history simulation research is relevant to this methodological goal. Synthetic success establishes that an algorithm can perform under the chosen generator; real-data backtests are needed to evaluate robustness to data quality, operational transitions and tails outside the model's assumptions.

A reproducible analytical synthesis

The literature shows that the scientific unit of a reserving model can be a transaction, a reported claim, an occurrence event or an entire portfolio. The unit determines which risks the model can directly represent. Event-time models describe reporting, payment and settlement paths; conditional aggregate models describe total payments by development cell; machine-learning models exploit nonlinear information available in the claim history. They are related techniques with different information needs. There is no general implication that retaining more attributes produces a smaller future prediction error. The additional information must be measurable at the actual prediction date, conditionally relevant, stable under operational change and supported by enough observations for reliable estimation.

The source-based comparison yields four methodological conclusions. First, identification of the unreported population is a separate problem from projecting the reported population. Reported-only learning can estimate RBNS but cannot identify full IBNR without an occurrence and reporting mechanism or suitable collective data. Second, case estimates are useful predictors only with historically accurate definitions and snapshots, since their changes reflect both claim information and managerial decisions. Third, portfolio mean reconciliation follows arithmetic identities, while distributional reconciliation additionally requires attention to common shocks and dependence. Fourth, a well-described aggregate benchmark is a necessary comparator for granular methods, even when the final decision uses a granular model. Each conclusion follows from the definitions and the reviewed scholarly record; none is asserted as a new proprietary scientific theory.

A practical decision procedure begins with the valuation target and the legal and accounting scope. It then establishes data provenance, reconstructs historical cutoffs, estimates the reporting and development components, assembles a coherent forecast, compares with aggregate baselines and reconciles gross, net, expense and financial measures. Material gaps trigger documented limitations or method modification. The final communication states the purpose, date, conditional measure, uncertainty, assumptions and realized results of validation. The process makes individual-claim modeling intelligible to an actuary, claims department, auditor or regulator without forcing all users to adopt the same computational architecture.

The mathematical identity of conditional portfolio expectation does not require a particular neural-network architecture, and no algorithmic result should be promoted beyond the distribution and duration of the available observations. Antonio and Plat (2014) reported gains for one European liability-claims application; Pigeon, Antonio and Denuit (2014) reported good performance in a paid–incurred general-liability case study; Gabrielli (2021) supplied a neural reported-claims proof of concept. These are scientifically important conditional findings. The empirical question for a new insurer remains whether similar benefits persist after its own exposure mix, claims systems, historical cutoffs, calibration choices and reporting scope have been applied.

Conclusion

Individual claim reserving brings the observable development of insurance losses into direct contact with actuarial projection. Its information base consists of occurrence and report dates, transactional payment histories, dynamic claim states, interim valuations, exposure characteristics and financial relationships. Its stochastic targets are future cash flows from already occurred events, separated into reported and unreported components and defined at an explicit valuation boundary. The expected portfolio amount is a conditional sum; the associated uncertainty requires coherent treatment of process risk, parameter uncertainty, long tails, reporting selection, reopenings and correlated outcomes.

The reviewed literature supplies several complementary routes to estimation. Marked-point-process and survival models describe event histories; paid–incurred models combine cash and case information; historical simulation retains heterogeneous claim markers; statistical learning represents complex conditional relationships; collective architectures link unreported and reported processes. Their validity rests on consistent data, identified observation rules and historical tests. The gain from granularity is strongest when recorded claim-level signals survive real prospective validation, and weakest when increased dimensionality amplifies missingness, operational artifacts and unobserved tails.

The resulting scientific synthesis is a disciplined relationship between claim-level prediction and aggregate obligation. Reconcile the event ledger to triangles, the estimated future cash flows to expected ultimate losses, and the actuarial target to the precise financial reporting or regulatory purpose. Disclose uncertainty, data limitations, source access and the scope of empirical claims. Use genuine historical valuation cutoffs and credible aggregate comparators to assess whether new methods improve on established approaches. These steps establish a complete analytical basis for micro-level loss development while keeping substantive novelty claims proportional to evidence.

References