Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Predicting Claim Reporting and Settlement Delays

Abstract. The interval between an insured event and its first report, and the interval between reporting and resolution, determine when an insurer learns of a liability, how an open claim develops, and when a contractual obligation is discharged. These intervals are generated by distinct mechanisms: occurrence, reporting, investigation, payments, closure, and sometimes reopening. Their observation is shaped by valuation dates, recording systems, policy coverage, organizational processes, competing claim outcomes, and the availability of covariates at a forecast origin. This monograph develops a unified analytical account of reporting and settlement delay prediction while retaining separate estimands and observation processes. It critically examines delay distributions; incident-based and reported-claim likelihoods; survival, accelerated failure time, discrete-hazard, recurrent-event, and multistate methods; right truncation, delayed entry, and censoring; time-varying predictors; predictive evaluation under incomplete follow-up; and the relation between timing forecasts, incurred-but-not-reported claims, reported-but-not-settled claims, operational service, and consumer protection. An explicitly illustrative calculation shows how a time-to-closure distribution becomes a horizon-specific probability without implying empirical accuracy. The central conclusion is that trustworthy delay forecasting begins with an identifiable time origin, an auditable event history, a defined prediction horizon, and validation against the information available on the prediction date. A faster recorded settlement is meaningful only when the distinction between genuine resolution, administrative closure, and later reopening is preserved. The study synthesizes established statistical and actuarial methods, offers original comparative interpretation, and makes no claim to a new statistical theory or empirical portfolio discovery.

Keywords: Insurance claims; reporting delay; settlement delay; claim development; IBNR; RBNS; survival analysis; event history; right truncation; censoring; competing risks; multistate processes; reopening; probabilistic calibration; claims operations; predictive validation.

Introduction

A claim is simultaneously a possible insured liability, a developing information record, and a sequence of operational events. The period between the underlying loss and its notification controls the insurer's immediate visibility into the portfolio. After notification, investigation and adjustment shape payment and closure. These two durations influence loss development, claims servicing, liquidity, reserving, and the experience of claimants. Their statistical interpretation requires clear separation because a claim that has not yet been reported is absent from the ordinary claims register, whereas a reported claim may remain observed, open, and administratively active for many years.

Jewell's continuous-time analysis of IBNYR delays (1989) demonstrates why the dates on which losses occur and are observed matter to predictions of unreported events. Norberg's continuous-time liability decomposition (1993) distinguishes incurred but unreported claims from reported claims still unsettled within a marked point-process approach. Later actuarial studies recover further information from the individual development histories that aggregated run-off triangles compress. Antonio and Plat's micro-level stochastic reserving study (2014) connects occurrence, reporting, payments, and final settlement through event histories; Verrall and Wüthrich's reporting delay analysis (2016) examines two real lines of insurance and the importance of trend and seasonality.

The research question is therefore precise: given a defined population and the information available at a particular date, what is the predictive distribution of the time until an admissible claim event, and how do we establish that the distribution remains reliable in subsequent periods? Reporting and settlement models answer different conditional questions. A reporting model may forecast the number and timing of previously unobserved claims, while a settlement model predicts the future of a known but unfinished claim. A common framework of temporal indexing, probability, event histories, and validation connects these questions without conflating their populations.

The scope is international and interdisciplinary. The study concerns statistical prediction and the interpretation of claims operations rather than any particular insurer's operational rules. Regulatory examples are identified by jurisdiction. The assessment is analytical: no insurer-level restricted dataset, external benchmark replication, or new field experiment is represented as having been conducted. Claims about empirical comparative performance are attributed to the source study and are not generalized beyond its observed design.

Loss Date, Report Date, and Settlement Date

A chronological claim ontology

A valid delay analysis begins by identifying what happens, when it happens, and which record first makes the event visible. Let LiL_i denote the legally and operationally defined occurrence date of loss event ii, RiR_i the earliest qualifying claim report received by the relevant insurer or administrator, PikP_{ik} the timestamp of payment transaction kk, and CiC_i the timestamp of the first closure that satisfies the declared closure definition. When reopening is possible, let UijU_{ij} be a subsequent reopening timestamp and CijC_{ij} the successive reclosure dates. An effective dataset preserves the ordered sequence of recorded transactions and status transitions rather than a single overwritten field labeled “closed.”

The occurrence-to-report duration is DiR=Ri−LiD_i^{R}=R_i-L_i. The report-to-first-closure duration is DiC=Ci−RiD_i^{C}=C_i-R_i. The event-to-first-closure duration is DiT=Ci−Li=DiR+DiCD_i^{T}=C_i-L_i=D_i^{R}+D_i^{C} whenever all three times describe the same claim and satisfy chronological consistency. Payment delays must specify both the start and the endpoint: time to first payment, time to a final indemnity payment, and time to accounting settlement are distinct statistical variables. A claim may be closed without an indemnity payment, and its final payment may precede or follow a recorded administrative closure. These definitions are estimands, not interchangeable labels.

Recorded loss dates may be uncertain for long-tailed liability, occupational, latent-property, and cumulative injury claims. “Date of knowledge,” occurrence date, legal trigger date, report-received date, and first notification to a broker may differ. The reporting clock therefore requires an explicit convention and an audit field for uncertainty or later correction. A corrected timestamp is part of provenance; overwriting the old date without preserving a revision history can manufacture apparently negative reporting delays or alter historically reproducible results.

The forecast origin and available information

Let AA be the valuation or operational forecast date, and let Hi(A)\mathcal H_i(A) be all information that could actually have been known for claim ii by AA. The predictive object for a reported, open claim is

Pr⁡(Ci−Ri≤a+h∣Ci−Ri>a,Hi(A)),a=A−Ri,\Pr(C_i-R_i\leq a+h\mid C_i-R_i>a,\mathcal H_i(A)),\qquad a=A-R_i,

where hh is the forecast horizon. For an unreported occurrence cohort, the object instead includes the probability that a loss exists but has not been registered, with future reporting conditional on loss characteristics available at the relevant informational level. A score computed later from the completed claim file belongs to Hi(A′)\mathcal H_i(A') for some later date A′>AA'>A, not to the old prediction at AA.

This distinction is foundational for valid model evaluation. The analysis in Crevecoeur, Robben, and Antonio (2022) illustrates how policy, claim, and evolving event-history covariates enter distinct layers of an individual development model. For a contemporaneous forecast, the record must be reconstructed at the precise prediction cutoff rather than queried from a final-state database. Predictors created from later reserve revisions, final legal findings, eventual claim closure, or settlements negotiated after AA are future information even if a retrospective extract attaches them to the original claim row.

Observable populations and event ledgers

An occurrence cohort consists of losses under a specified exposure definition during a specified calendar interval. A reporting cohort consists of claims first notified during an interval. A closure cohort consists of claims whose closure was registered during an interval. These cohorts answer different questions and have different selection mechanisms. A closure-cohort sample excludes the unfinished tail by construction. A reporting-cohort sample omits unreported occurrences. For example, a median duration computed among only claims closed this quarter estimates the duration distribution of this quarter's closures; it is generally not an unbiased median of the current open inventory's ultimate time to closure.

A research-grade event ledger includes stable claim and exposure identifiers; line of business; occurrence, report, status-change, payment, reopening, and extraction timestamps; transaction amount; recorded actor and version; and reason codes for transitions. It also records withdrawn notifications, duplicates, linked catastrophe events, split or consolidated claims, and updates to coverage or claim cause. These fields are indispensable for identifying when a numerical improvement is actually a change in event coding.

What the data cannot directly identify

The joint distribution of loss occurrence, reporting, and settlement cannot generally be recovered from a register that contains only reported claims without additional assumptions about claim incidence, exposures, and reporting selection. Two portfolios can show identical reported-case delays but have different unreported tails because their unobserved occurrence frequencies differ. Conversely, two portfolios can have similar ultimate report counts but different operational settlement speeds. Identifiability depends on the sampling frame and the likelihood; a flexible algorithm does not overcome the absence of an observable denominator.

This conclusion follows from the structure of Jewell's reporting models (1990) and the granular event-history approach of Antonio and Plat (2014). The first analyzes the information loss arising when continuous dates are reported only in coarse epochs; the second conditions future development on the information registered during a claim's lifetime. Their complementarity supports a practical division between occurrence-level, report-level, and operational models.

Reporting Delay Distributions

Survival functions and reporting hazard

For a claim generated under a specified occurrence process, define reporting delay TR≥0T_R\geq0, conditional distribution FR(t∣x)F_R(t\mid x), survival SR(t∣x)=1−FR(t∣x)S_R(t\mid x)=1-F_R(t\mid x), and hazard λR(t∣x)=fR(t∣x)/SR(t∣x)\lambda_R(t\mid x)=f_R(t\mid x)/S_R(t\mid x) where a conditional density exists. The survival here means “not yet reported”; it does not describe life insurance survival or contractual survival of a claim. For a discrete-day or discrete-month grid, qj(x)=Pr⁡(TR=j∣TR≥j,x)q_j(x)=\Pr(T_R=j\mid T_R\geq j,x) is a conditional report hazard, and SR(j∣x)=∏k≤j(1−qk(x))S_R(j\mid x)=\prod_{k\le j}(1-q_k(x)), with indexing chosen to match the time intervals.

The mean of a delay distribution, when finite, is ∫0∞SR(t) dt\int_0^\infty S_R(t)\,dt. Quantiles tp=inf⁡{t:FR(t)≥p}t_p=\inf\{t:F_R(t)\ge p\} support operational interpretation more directly than a mean when reporting lags are heavily skewed. A median reporting delay cannot characterize the long unreported tail required for a high reporting-completion proportion, which is why a full distribution and its uncertainty matter.

Exponential models supply a constant hazard but are restrictive where short notification lags are common and late reports form a persistent tail. Weibull and lognormal distributions offer different hazard and tail shapes; mixtures can represent qualitatively different reporting mechanisms. Flexible duration models allow calendar time, contract provisions, insurer communications, and loss type to shift the distribution. The correct model is selected by calibrated out-of-time forecasts, rather than by a family name alone.

Occurrence intensity and the observed reporting process

Let ν(s,x)\nu(s,x) denote the intensity of actual loss occurrences at calendar date ss among exposures with covariates xx. A simple thinning representation gives expected reported count up to valuation date AA:

E{Nreported(A)}=∫s≤A∫ν(s,x)FR(A−s∣s,x) dGs(x) ds.\mathbb E\{N_{\mathrm{reported}}(A)\}=\int_{s\le A}\int \nu(s,x)F_R(A-s\mid s,x)\,dG_s(x)\,ds.

For a finite accident interval, the expected incurred-but-not-yet-reported count at AA has the complementary factor SR(A−s∣s,x)S_R(A-s\mid s,x). These equations require a defined exposure population and an assumed relationship between the occurrence intensity and the conditional delay distribution. They do not say that unreported counts can be obtained by simply dividing observed cases by a universal percentage.

Jewell (1989) studies event generation and observation delay under uncertain Poisson rates and delay parameters, and Norberg (1993) embeds a related marked occurrence-process representation within the wider liability prediction problem. The representations show that reporting speed and occurrence intensity are separately relevant; an increase in reported cases could arise from greater underlying incidence, more rapid notification, or both.

Right truncation is central to reporting-delay estimation

Suppose only claims reported by AA enter a historical extract. For an occurrence at LiL_i, the observed delay satisfies TRi≤A−LiT_{Ri}\le A-L_i. The selection mechanism is right truncation in the reported-case sample, with the truncation bound varying by loss date. Under a conditional model, a case observed because it was reported has likelihood contribution

Li(θ)=fR(ti∣xi;θ)FR(A−Li∣xi;θ),0≤ti≤A−Li,L_i(\theta)=\frac{f_R(t_i\mid x_i;\theta)}{F_R(A-L_i\mid x_i;\theta)},\qquad 0\le t_i\le A-L_i,

when the conditioning scheme legitimately treats reported cases and their occurrence dates as the sample frame. Fitting fRf_R to the visible delays without the denominator overweights early reports, especially for recent accident periods. A joint occurrence-and-report model may be preferable when the exposure and incidence process are known; its likelihood then includes observed reporting counts as well as the missing-occurrence mechanism.

The underlying lesson also emerges from Verrall and Wüthrich's empirical comparison (2016), which explicitly treats claim-arrival and reporting-delay structure rather than treating recorded delays as a self-contained sample. Ordinary Kaplan–Meier censoring assumes the subjects at risk are known even when the event time is incomplete. Unreported claims lack a recorded claim-level row, so it is not defensible to label them automatically “right-censored observations” in a reported-only file.

Calendar, seasonality, policy and line-of-business heterogeneity

Delay is measured on two clocks: elapsed time since occurrence and absolute calendar time. Holidays, public-service outages, catastrophe surges, filing campaigns, litigation environments, and changes in digital intake can affect reports in calendar time. A summer disaster cohort may show short-term report acceleration followed by a delayed tail, while a liability cohort may be influenced by prolonged recognition and legal notification. An occurrence-year effect and an elapsed-delay effect should therefore be separately considered.

Verrall and Wüthrich's nonstationary reporting analysis (2016) relaxes assumptions to account for trends and seasonal patterns in real property and casualty datasets; this is evidence for testing calendar structure, not a license to assume the same parameters across jurisdictions and products. A defensible model may combine calendar-year indicators, month-of-year effects, cause-of-loss classes, policy reporting arrangements, and exposure-mix controls, with explicit smoothing or partial pooling where cells are sparse.

Coarse reporting periods and identifiable precision

A dataset may record only “loss month” and “report month.” Then a reported lag of two periods identifies an interval of possible continuous delays rather than an exact duration. A coherent discrete-time likelihood uses interval probabilities FR(b)−FR(a)F_R(b)-F_R(a). Treating every month's report as the first day of the month introduces artificial mass at round numbers and may distort the hazard near month-end. Jewell's discrete-time follow-up (1990) investigates the loss of predictive information from such temporal aggregation. The appropriate precision of a forecast cannot exceed that of the validated observation process without an additional defensible timing model.

Predictive examples and uncertainty

Consider a wholly hypothetical homogeneous occurrence cohort of 1,000 losses, already established from an independent exposure-generation source. Suppose the true conditional reporting probabilities by 30, 90, and 180 days are respectively 0.70, 0.90, and 0.97. The expected numbers reported by those dates are 700, 900, and 970, with expected unreported counts 300, 100, and 30. These are arithmetic implications of stipulated probabilities, not empirical estimates. If the denominator of 1,000 losses were unknown, seeing 700 early reports would not by itself identify the remaining 300; occurrence-rate and delay assumptions would be necessary.

Parameter uncertainty matters particularly in the tail. If plausible 180-day report completion ranges from 0.95 to 0.99, the same hypothetical 1,000-loss cohort implies between 10 and 50 expected unreported cases at that time, even before uncertainty in total incurred claim count is considered. An honest reporting forecast supplies a probability distribution for future reports or a calibrated interval, along with model and data limitations, rather than a single apparently exact count.

Survival and Duration Regression

Open claims and the right-censoring likelihood

For a reported claim with report date RiR_i, first-closure time TCi=Ci−RiT_{Ci}=C_i-R_i may be observed, or the claim may still be open at valuation date AA. Define observed age yi=min⁡(TCi,A−Ri)y_i=\min(T_{Ci},A-R_i) and event indicator δi=1{TCi≤A−Ri}\delta_i=1\{T_{Ci}\le A-R_i\}. Under independent administrative censoring conditional on modeled information, the familiar contribution is

Li(θ)=fC(yi∣xi;θ)δiSC(yi∣xi;θ)1−δi.L_i(\theta)=f_C(y_i\mid x_i;\theta)^{\delta_i}S_C(y_i\mid x_i;\theta)^{1-\delta_i}.

This distinction prevents the estimator from treating claims that have not closed yet as if they will never close. It also prevents the opposite error of dropping open claims entirely and estimating a systematically shortened distribution from the closed subset. The Kaplan–Meier product-limit estimator (Kaplan and Meier, 1958) is a foundational nonparametric tool when censoring assumptions hold and individuals at risk are identified.

An open claim with prior payments may still contribute to a time-to-first-closure analysis as an at-risk case; the number of earlier payments is then part of its history. A record with no payment and a record with several partial payments may have different future risks. The model's definition must state whether “settlement” means closure, first indemnity, last indemnity, agreement in principle, or another verifiable event.

Hazard regression and interpretation

The Cox proportional-hazards specification (Cox, 1972) expresses an event hazard as λC(t∣x)=λ0(t)exp⁡(β⊤x)\lambda_C(t\mid x)=\lambda_0(t)\exp(\beta^\top x). The baseline hazard is unspecified, and the coefficients describe multiplicative changes in the instantaneous closure hazard conditional on a claim still being open at duration tt. A hazard ratio above one indicates a higher instantaneous event rate under its assumptions; it does not, by itself, assert a particular reduction in mean days to settlement or a causal benefit of an intervention.

Claim characteristics may violate proportional hazards. For example, a complex bodily injury claim may be especially unlikely to close during the first weeks while its hazard later increases after medical evidence is assembled; a simple property claim may close quickly or enter a smaller persistent tail. Schoenfeld-residual diagnostics, time interactions, and interval-specific effects can reveal misspecification. When nonproportionality is material, piecewise, stratified, or flexible hazard models may be more interpretable than one constant hazard ratio.

Accelerated failure time and conditional quantiles

An accelerated failure time (AFT) model parameterizes log duration, for example log⁡TC=α+β⊤x+σε\log T_C=\alpha+\beta^\top x+\sigma\varepsilon, where the distribution of ε\varepsilon determines a lognormal, Weibull-related, log-logistic, or other survival family. The exponentiated coefficient can be read as a time ratio under the chosen specification, which is often easier for a claims operations audience than a hazard ratio. However, a fitted AFT model still requires the correct censoring likelihood, temporal validation, and scrutiny of its implied late tail. A model predicting the median well can remain unreliable at the 95th percentile, which may drive consumer impact and resource allocation.

Quantile predictions are particularly useful where the operational question is how long it will take for a specified proportion of cases to resolve. The model should report whether a quantile refers to the full occurrence cohort, all reported claims, claims currently open after duration aa, or claims in a particular state. These conditional populations produce different numbers even under the same fitted distribution.

Discrete-time hazards and calendar resolution

If case statuses are captured weekly or monthly, a discrete-time hazard model estimates qit=Pr⁡(TCi=t∣TCi≥t,Hi(t))q_{it}=\Pr(T_{Ci}=t\mid T_{Ci}\ge t,\mathcal H_i(t)). A complementary log–log link approximates continuous-time proportional-hazard structure under intervals; a logit link models the discrete conditional probability directly. The conditional survival through interval hh is the product of the interval-specific non-event probabilities. Models must align interval inclusion, start dates, and status-change conventions; arbitrary conversion of month-level data into exact days produces false precision.

A discrete-time dataset has one at-risk interval per eligible claim-period. Covariates are attached as they were known at the start of that interval. This form is attractive for time-updated claim information, but interval rows from the same claim are dependent and need claim-level clustering or a coherent longitudinal likelihood for uncertainty quantification. Computational capacity permits complex models; sound predictive design still begins with a population and a defined observation window.

Time-dependent covariates and endogenous information

Claim size estimates, reserve changes, medical reports, legal representation, adjuster assignments, and earlier payment events may update the closure forecast as the claim develops. Such fields are time-varying covariates, but many are consequences of the same latent case complexity that affects settlement. A large reserve increase may signal delayed closure without causing it. A claims-team transfer may respond to already prolonged delay, so observational associations must not be interpreted as effects of staffing changes without a causal design.

The counting-process extension by Andersen and Gill (1982) provides a mathematical basis for intensity models with evolving at-risk histories and recurrent event structures. In claims operations, a feature's effective timestamp must be recorded, not merely the day on which it was eventually imported into a warehouse. Otherwise the model can inadvertently learn the outcome from its own future.

Recurrent payments and mark processes

A claim can make several indemnity or expense payments before closure. Let NiP(t)N_i^P(t) count eligible payment events up to elapsed time tt, and let marks MikM_{ik} record signed payment amounts or other payment attributes. A marked point-process or recurrent-event model distinguishes the intensity of a transaction from its size, and permits payment histories to inform the closure hazard. Antonio and Plat (2014) explicitly use a marked-process event-history framework for general insurance; it provides a natural counterpart to single-event duration regression.

An analyst must avoid using payment count since inception as a fixed baseline covariate when forecasting at first notification. That count is unavailable then and is positively associated with exposure to time at risk. A landmark or dynamic forecast updates payment history at each forecast origin and compares predicted outcomes among claims still at risk at that time.

Prediction at a landmark date

Suppose an estimated closure model yields conditional survival SC(t∣x)S_C(t\mid x) based on baseline-only predictors and the model's assumptions. For a claim open for aa days, its chance of first closure within an additional hh days is

1−SC(a+h∣x)SC(a∣x),SC(a∣x)>0.1-\frac{S_C(a+h\mid x)}{S_C(a\mid x)},\quad S_C(a\mid x)>0.

With a correctly specified exponential duration model whose median is 40 days, the constant hazard is λ=log⁡(2)/40\lambda=\log(2)/40 per day. For a remaining horizon of 60 days, 1−exp⁡(−60λ)=1−2−1.5≈0.64641-\exp(-60\lambda)=1-2^{-1.5}\approx0.6464, independent of elapsed age because the exponential law is memoryless. The number 0.6464 is a mathematical illustration, not a measured insurance rate. Under a non-exponential model, the forecast generally depends on how long the claim has already been open; this is precisely why distributions matter more than isolated average durations.

Reopened and Censored Claims

Closure as a state transition

A claim's life can be represented by states such as “reported/open,” “investigation,” “partial-payment/open,” “closed without payment,” “closed following payment,” and “reopened.” The necessary number of states depends on whether the purpose is claim duration, payment cash flow, or service-process analysis. For a basic multistate formulation, transition intensities λjk(t∣Ht)\lambda_{jk}(t\mid\mathcal H_t) govern movement from state jj to state kk. Probabilities of state occupancy after hh depend on the whole transition system, not merely the hazard of first closure.

The collective reserving model with claim openness of Lindholm and Zakrisson (2022) explicitly admits claim closings and re-openings and demonstrates why open-state dynamics can matter in aggregate development. The hierarchical reserving study of Crevecoeur, Robben, and Antonio (2022) integrates development events as layers. These studies motivate preserving a ledger of statuses rather than treating “closed” as a permanently absorbing category by administrative convention.

First closure, durable closure, and ultimate finality

“Time to first closure” is estimable from the first closing entry. “Time to final closure” is more demanding: after a claim closes, it may reopen, and proving that a closure is final requires a finite monitoring horizon, a substantive legal or administrative rule, or explicit assumptions about future reopening. A model that assigns zero future reopening probability after the extraction date may overstate the permanence of recorded resolution. For a verifiable objective, define “no reopening within 12 months after closure,” or another horizon justified by portfolio development and follow-up. This is a horizon-specific endpoint, not proof of permanent finality.

A difference in predicted first closure times can coexist with no difference in durable resolution when one workflow closes files early but reopens many later. This leads to a substantive interpretation: settlement timeliness should be examined jointly with reopening frequency and postclosure payments. Otherwise, the apparent speed of a claims department may simply be a reporting artifact.

Censoring, delayed entry, truncation, and competing events

Four mechanisms require separate handling. Administrative right censoring arises when an identified reported claim is still open at the valuation cutoff. Delayed entry arises when a duration cohort is first observed after its clock has begun, provided eligibility and survival to entry are explicitly represented. Right truncation arises in a reported-only occurrence cohort because later reporters are absent from the observed sample. Competing risks arise when distinct first terminal outcomes, such as paid settlement, denial, or withdrawal, preclude one another as the first endpoint. These mechanisms have different denominators and likelihoods.

The Fine–Gray competing-risk model (Fine and Gray, 1999) addresses subdistribution hazards and cumulative incidence for a specified event type. In insurance, the probability of a paid closure by day 90 is a cumulative incidence, whereas the hazard of paid closure among currently unresolved claims is a cause-specific quantity. Treating a denied claim as ordinary noninformative censoring when predicting paid closure can distort event probabilities. For all-cause first closure, the competing terminal types can instead be aggregated deliberately into a common event, provided the research question does not require their separation.

Informative follow-up and missing status changes

Administrative censoring at a common cutoff can often be treated as independent conditional on covariates in a stable system, but this must be evaluated. Insurers may archive claims after a transfer, migration, insolvency event, or outsourcing change. A missing status update is not automatically evidence that a file remained open. Cases with incomplete or fragmented data may be more complex than fully documented cases, producing informative data loss. Sensitivity analyses should vary assumptions about unavailable closure dates and investigate changes in migration windows, business units, and external administrators.

The model should distinguish three forms of absence: an unknown event time, a known event recorded with uncertain timing, and an event never ingested into the source platform. Each threatens a different part of the estimand. Missingness indicators can be predictively useful, but their relationship to case complexity, geography, reporting channel, and record quality should be disclosed.

Multiple claims from one occurrence

A catastrophe can produce many claims sharing an occurrence date and similar environmental conditions. A liability event can generate multiple claimants and linked legal proceedings. Shared causal factors produce correlated reporting and settlement durations. Claim-level bootstrap resampling without cluster structure may substantially understate predictive uncertainty. Where the intended deployment is to a new catastrophe or a new portfolio period, validation should respect event and calendar clusters. Randomly splitting records from the same catastrophe across training and test sets does not measure performance on a genuinely new event.

A practical approach records catastrophe or incident identifiers, common administrative channels, and grouped exposures. Frailty, multilevel effects, and cluster-robust inference are possible options when their assumptions fit. The relevant decision is whether the forecast should generalize to new claims within known events or to entirely new events, which determines the meaningful holdout unit.

Interpretive limits of the statistical record

A closed file may represent full payment, partial compromise, a denied claim, a withdrawn notification, litigation settlement, an administrative write-off, or another institutionally defined state. Reducing all of these to an identical “successfully settled” label can make a fast model look accurate while erasing the qualitative difference between outcomes. A published empirical study on an administrative category should be read under that category's definitions. The statistical reconstruction of status trajectories is therefore a precondition for any economic or fairness claim about settlement performance.

Forecast Validation

The fundamental prediction task

Predictive validation assesses the forecast actually available at a chosen date, for the specified cohort and outcome. Development data must be reconstructed as of each historical forecast origin. Training should precede the evaluation period; hyperparameter selection and feature engineering must not use future claim information or the test period. Repeated calendar-origin evaluation is particularly important when claim-reporting practice, legal processes, prices, staffing, and claim mixes evolve.

The hierarchical claims reserving analysis (Crevecoeur, Robben, and Antonio, 2022) includes moving-window evaluations and simulation scenarios, emphasizing the difference between a model's ability to reconstruct past development and to forecast future development. Insurance applications require both a prediction origin and an explicit target horizon. A model can perform well for 30-day closure prediction but fail for 365-day tail probabilities.

Temporal, cohort, and entity leakage

A random row split of a long claim-event table allows the same claim's early history to occur in training and its later history in testing. A customer, common incident, law firm, or adjuster may also span folds and create an unrealistically easy evaluation. A secure design first defines a historical cutoff, generates permitted predictors from a snapshot at that cutoff, and tests on a genuinely subsequent origin or held-out cluster. Historical target and input data must not be silently restated with future corrections unless a parallel “as-known-then” snapshot is available.

Examples of leakage include a settlement-negotiation code issued after the outcome, the ultimate paid amount stored in a baseline reserve column, an amended report date retroactively applied to old records, and the eventual final claim category used as if known on notification. Such leakage can generate impressively sharp predictions that are operationally unusable. Feature provenance and historical replay are therefore central components of statistical quality assurance.

Calibration for survival and competing events

For an hh-day closure forecast, calibration asks whether claims assigned probability pp actually close within hh at approximately that frequency, accounting for censoring and cohort selection. Calibration-in-the-large, calibration curves, horizon-specific observed-to-expected ratios, and recalibration slopes expose different failures. A model can correctly rank fast and slow claims while producing systematically optimistic closure dates. The prediction-performance framework of Steyerberg and colleagues (2010) emphasizes calibration and discrimination as complementary dimensions.

When different closure types compete, calibration must address cumulative incidence of the specific outcome; censoring competing closures as if they were lost to follow-up does not supply that quantity. Any reported calibration metric must state its target population, horizon, weighting method, and the amount of effective follow-up.

Proper scores and censoring corrections

For an uncensored binary event indicator Yi(h)=1{Ti≤h}Y_i(h)=1\{T_i\le h\} and forecast pi(h)p_i(h), the Brier score is n−1∑i(Yi(h)−pi(h))2n^{-1}\sum_i(Y_i(h)-p_i(h))^2. It rewards the accuracy of probability forecasts and is proper for this binary target. When some event statuses at horizon hh are unknown due to censoring, naive substitution of “not observed” with “did not occur” biases the score. Inverse-probability-of-censoring weighted approaches provide a principled alternative under conditions on the censoring distribution.

The Gerds–Schumacher analysis (2006) studies consistency of expected Brier score estimators under right censoring, including conditional independence of event and censoring times given covariates. This theoretical requirement should be stated when an insurer reports an “adjusted Brier score”: the adjustment is only as credible as the corresponding censoring model and overlap of observable histories.

A suite of evaluation summaries should include a proper scoring measure, horizon-specific calibration, time-dependent discrimination where meaningful, coverage of prediction intervals, and stability by accident/report period. A single c-index can obscure severely miscalibrated settlement probabilities. A model that is slightly less discriminative but demonstrably calibrated may be preferable for workload and claimant communications.

Baselines and comparative design

A complex method should be measured against clear alternatives: a recent-cohort empirical survival curve, a stratified Kaplan–Meier estimate for mature reported claims, a simple parametric survival regression, and a seasonally adjusted reporting hazard. Claims reserving exercises may also compare timing-linked aggregates with chain-ladder-style projections, while stating that an aggregate reserve metric and an individual closure forecast are different prediction targets. Antonio and Plat (2014) report favorable results for their own out-of-sample micro-level case study; the finding is conditional on that portfolio and design and is not a universal ranking of micro-level methods.

Model selection should prevent an unfair information advantage. If a sophisticated model uses detailed legal and payment histories, evaluate it at a date when that information exists; if an aggregate method uses only run-off counts, compare each method both for its own intended decision and under a clearly described shared information set. Different targets must not be declared winners and losers on the basis of incomparable loss functions.

Heterogeneity, fairness, and drift

Forecast accuracy can differ by coverage category, jurisdiction, reporting channel, claim amount, legal representation, catastrophe period, and vulnerable-customer circumstances. A materially accurate global estimate may conceal poor performance in small but consequential groups. Differences in group error rates require interpretation because the claim mix, event mechanisms, available data, and statutory or contractual processing rules may differ. The point is to locate miscalibration and harmful decisions, not to assert causality from subgroup correlations alone.

Data shifts include a new intake portal, a revised claims definition, altered court timelines, staffing discontinuities, large weather events, or changed inflation conditions. Monitoring should track input distribution, label completeness, reporting lags, predicted-versus-observed event rates, and tail instability on successive historical cutoffs. The NIST AI Risk Management Framework (NIST, 2023) offers general risk-governance principles for AI-enabled decision systems; it is guidance for governance, not an insurance-specific statute or a statistical guarantee of validity.

Sensitivity and stress testing

Reporting models need tail-sensitivity exercises for sparse long lags, calendar corrections, missing exposure, and misdated events. Settlement models need tests for closure redefinitions, undocumented reopening, terminal-outcome recoding, informative exit, delayed entry, and claims transferred between administrators. Forecasts should be recomputed with alternative valuation snapshots to quantify the effect of historical database revision. At portfolio level, uncertainty should incorporate within-event correlation, parameter variation, and plausible changes in future claim and process characteristics.

Where a model's late-tail output is unstable, the appropriate scientific statement is a broad probability interval accompanied by the source of uncertainty. Apparent numerical precision without sufficient tail information has no evidential advantage. A reproducible validation record should include code revision, data extraction schema and time, feature cutoff policy, model specification, resampling unit, sample counts, missingness checks, calibration method, and all evaluated horizons.

Reserving and Operational Decisions

Connecting delay probabilities with liability amounts

Reporting forecasts contribute to estimates of claims incurred but not yet reported. Settlement and payment-event forecasts contribute to expectations for reported but not settled claims. However, the cash-flow consequence of a future settlement also depends on severity, coverage, salvage and subrogation, expenses, economic conditions, and the timing of interim payments. Predicting closure dates alone does not identify future paid amounts. Under a properly specified joint event-and-mark model, a discounted expected future payment measure might be expressed schematically as

E ⁣[∑k: Pik>Av(Pik−A)Mik | Hi(A)],\mathbb E\!\left[\sum_{k:\,P_{ik}>A}v(P_{ik}-A)M_{ik}\,\middle|\,\mathcal H_i(A)\right],

where vv is a chosen discount factor and MikM_{ik} a signed eligible payment mark. This is a generic expectation, not a standalone accounting valuation. Reporting, occurrence, expense, and dependence models may require further information before the expression becomes an insurance liability estimate.

Norberg's four-way claims status decomposition (1993) highlights the conceptual distinction among settled, RBNS, IBNR, and not-yet-incurred exposure categories. Huang and colleagues' individual reporting and settlement model (2015) provides a tractable example with independent reporting and settlement delays under specific assumptions. An insurer should test any such independence assumption; complex claim types, late notification and evidence availability can link the two delays.

When correlation between reporting and settlement matters

The same event features may drive both reporting and closure: difficult loss attribution can postpone notification and require later investigation. A simplified model that treats DRD^R and DCD^C as independent may generate optimistic predictions of total development time. A conditional factorization p(DR,DC∣X)=p(DR∣X)p(DC∣DR,X)p(D^R,D^C\mid X)=p(D^R\mid X)p(D^C\mid D^R,X) shows how reporting lag can enter settlement prediction after notification, while remaining unavailable for a claim not yet reported. For claims known to exist at occurrence level, joint simulation can propagate uncertainty through both stages.

Dependence can also arise through selection: only some claims are visible by a particular date, so a model fitted to them may not represent the late-reporter population. The meaningful solution is to define the target population first, then use the appropriate occurrence, reporting, and subsequent development components with stated assumptions. A high-performing classifier trained on mature closed claims may fail as a forecast of today's newly reported queue even if its training error is small.

Workload, staffing, and expected open inventory

Let new report arrivals enter an operational queue at rate λR(s)\lambda_R(s). In a simplified stable, independent service system, expected active inventory at time AA can be related to the history of arrivals weighted by the probability of still being open. More generally, an approximate convolution takes the form

E{O(A)}≈∫s≤AλR(s)SC(A−s∣s) ds,\mathbb E\{O(A)\}\approx\int_{s\le A}\lambda_R(s)S_C(A-s\mid s)\,ds,

with corrections for case heterogeneity, service-system capacity, withdrawals, transfers, and reopening. This identity explains why two processes with equal report volume can produce different open inventories if their settlement-time distributions differ. It does not imply that case-handling capacity or arrival rates are unaffected by workload: queue feedback can invalidate the exogeneity embedded in a simple forecast.

Operational use requires separately estimating the number of new reports, age structure of the current queue, future payments, and probability of reopening. Workload should also distinguish claim complexity and required staff hours; a predicted 90-day duration does not necessarily represent 90 days of continuous handling. Capacity decisions demand service-effort information in addition to the time-to-event distribution.

Timeliness and consumer outcomes

The United Kingdom Financial Conduct Authority's ICOBS 8.1.1 claims rule requires insurers within its scope to handle claims promptly and fairly, provide reasonable claims guidance and progress information, and settle promptly once terms are agreed. This is a jurisdiction-specific conduct rule rather than a universal statutory day limit. In that setting, a delay predictor can identify cases likely to need more communication, specialist review, or additional resources; its governance must also consider the risk of using the model to defer difficult cases or privilege easily closed files.

An empirical study of external administrative rules provides a useful illustration of causal complexity. Heaton's mandatory reporting and settlement-delay study (2020) uses a difference-in-differences design around a U.S. Medicare reporting change and reports an increased settlement duration in the studied bodily-injury setting. The finding concerns those reporting rules and data; it does not establish that every new insurance data requirement necessarily delays claims. It demonstrates that legal reporting obligations may alter observed settlement timing independently of claim severity.

Finance, accounting, and the meaning of a reserve

The International Accounting Standards Board's IFRS 17 Insurance Contracts requires measurement of applicable groups using current estimates of fulfillment cash flows and a risk adjustment within its accounting framework, among other components. Statistical reporting and payment timing models can inform cash-flow projections used by insurers applying that standard, but the standard does not prescribe a particular survival model or convert a closure-hazard score directly into an IFRS 17 liability. Actuarial, accounting, and risk-management systems have related but distinct outputs.

There is also a temporal difference between financial closure and human claims experience: a settled claim may have expenses or recovery transactions outstanding, and a final cash-flow projection may extend beyond administrative closure. Reserving packages therefore need payment events and reconciliation to claims and general ledgers. A claims timeliness KPI cannot substitute for financial valuation, just as a reserve triangle alone cannot describe the quality of customer communication.

Research reproducibility and a principled implementation sequence

A reproducible project should begin with an event dictionary and temporal ledger; construct occurrence and reported-claim sampling frames; document right truncation, right censoring, and delayed entry; estimate interpretable baseline distributions; inspect seasonal and portfolio differences; introduce duration regression and, where needed, multistate or payment-event models; and evaluate at sequential historical forecast origins. Models and intervals should be compared within the same estimand and information set. Each material simplification must state the operational question it still answers.

A minimal audit trail for a reported open claim includes the predicted probability of first closure within a stated number of days, the event and source records available at that moment, model version, calibration date, data quality flags, and later outcome. Where outputs influence decisions affecting claimants, the decision record should preserve human accountability and avenues for correction. This operational use is compatible with the governance concerns articulated in NIST's AI RMF 1.0, while insurance-specific duties must be assessed under the applicable jurisdiction's rules and contract law.

Original analytical synthesis

The literatures on reporting delays, micro-level reserving, duration regression, and conduct regulation can be read together through three questions: which event initiates the clock, which states are visible at the valuation date, and which decision will use the forecast? Under this interpretation, temporal accuracy is a property of the entire statistical observation architecture, not merely of a fitted hazard curve. Two independently well-fitting component models may produce a poor total-development forecast if the interface between reporting and settlement is misspecified, the forecasts use incompatible cohorts, or they are validated at different information dates.

A second implication concerns performance incentives. Improving the predicted time to first closure can be valuable, but a valid operational interpretation also measures reopening, paid-settlement quality, and incomplete follow-up. Otherwise the “fastest” workflow may maximize administrative closure rather than durable resolution. The evidence base provides existing tools for these distinctions: reopening-aware claim models, dynamic event-history reserving, and censoring-aware forecast assessment. Their combined use is a methodological recommendation arising from the present synthesis, not a newly named doctrine.

A third implication is epistemic. The inability to observe an unreported claim is a question of sampling and model identification; the inability to observe a final settlement for an open claim is a question of incomplete duration follow-up. These are distinct reasons for uncertainty. Reporting systems that preserve occurrence exposures and claims event histories improve the questions an analyst can ask. Statistical sophistication is valuable after these data conditions have been made explicit.

Conclusion

Predicting claim reporting and settlement delays is a problem of temporal definition, observation, stochastic modeling, and decision-specific validation. An event-to-report forecast concerns latent losses entering an insurer's knowledge; a report-to-closure forecast concerns the future of a known case; payment timing, reopening, and competing closure outcomes extend the development process beyond a single duration. The two clocks interact through claim characteristics, calendar conditions, and the selection of cases available for analysis.

The literature establishes powerful tools for these questions. Jewell (1989, 1990) shows why event and reporting-date information, along with temporal aggregation, matters for hidden claims. Norberg (1993) places distinct claim states within an aggregate liability analysis. Kaplan and Meier (1958), Cox (1972), Andersen and Gill (1982), and Fine and Gray (1999) provide statistical structures for censored times, regression, evolving event histories, and competing outcomes. Contemporary insurance research by Antonio and Plat (2014), Verrall and Wüthrich (2016), Lindholm and Zakrisson (2022), and Crevecoeur, Robben, and Antonio (2022) translates many of those ideas into claims data and demonstrates the usefulness and limitations of more granular development models.

The analytical findings are consequently concrete. Reporting-delay data collected only from reported claims are right-truncated under occurrence-based selection. Open claims contribute censored duration information and cannot be excluded from an unbiased description of the current queue without strong assumptions. A closure time is meaningful only under an auditable event definition; reopening changes the interpretation of finality. Forecasts must be conditioned on information available at their exact origin and assessed on future periods with censoring-aware scoring and calibration. Reserving and operational decisions require the distribution of future payments and case histories in addition to a predicted date of closure.

A scientifically defensible claims timing system therefore reports conditional probabilities for defined horizons, their uncertainty, the populations and data from which they arise, and the conditions under which their accuracy has been tested. Reliable forecasts serve financial understanding and claims management when their interpretation is joined to transparent event definitions, correct sampling mechanisms, and responsible use. These conclusions are an original comparative synthesis of established research and do not assert the invention of a new theory, the completion of an external peer review, or the empirical validation of a new insurer-specific model.

References

  • Andersen, P. K., & Gill, R. D. (1982). Cox's regression model for counting processes: A large sample study. The Annals of Statistics, 10(4), 1100–1120. DOI

  • Antonio, K., & Plat, R. (2014). Micro-level stochastic loss reserving for general insurance. Scandinavian Actuarial Journal, 2014(7), 649–699. DOI

  • Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202. DOI

  • Crevecoeur, J., Robben, J., & Antonio, K. (2022). A hierarchical reserving model for reported non-life insurance claims. Insurance: Mathematics and Economics, 104, 158–184. DOI

  • Financial Conduct Authority. (2025, current rule page consulted 2026). Insurance Conduct of Business Sourcebook: ICOBS 8—Claims handling. United Kingdom. Official source

  • Fine, J. P., & Gray, R. J. (1999). A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association, 94(446), 496–509. DOI

  • Gerds, T. A., & Schumacher, M. (2006). Consistent estimation of the expected Brier score in general survival models with right-censored event times. Biometrical Journal, 48(6), 1029–1040. DOI

  • Heaton, P. (2020). The effect of mandatory insurer reporting on settlement delay. American Law and Economics Review, 22(2), 397–433. DOI

  • Huang, J., Qiu, C., Wu, X., & Zhou, X. (2015). An individual loss reserving model with independent reporting and settlement. Insurance: Mathematics and Economics, 64, 232–245. DOI

  • International Accounting Standards Board. (2017, as amended). IFRS 17 Insurance Contracts. IFRS Foundation. Official source

  • Jewell, W. S. (1989). Predicting IBNYR events and delays: I. Continuous time. ASTIN Bulletin, 19(1), 25–55. DOI

  • Jewell, W. S. (1990). Predicting IBNYR events and delays: II. Discrete time. ASTIN Bulletin, 20(1), 93–111. DOI

  • Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. DOI

  • Lindholm, M., & Zakrisson, H. (2022). A collective reserving model with claim openness. ASTIN Bulletin, 52(1), 117–143. DOI

  • National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). U.S. Department of Commerce. Official source

  • Norberg, R. (1993). Prediction of outstanding liabilities in non-life insurance. ASTIN Bulletin, 23(1), 95–115. Official source

  • Steyerberg, E. W., Vickers, A. J., Cook, N. R., Gerds, T., Gonen, M., Obuchowski, N., Pencina, M. J., & Kattan, M. W. (2010). Assessing the performance of prediction models: A framework for traditional and novel measures. Epidemiology, 21(1), 128–138. DOI

  • Verrall, R. J., & Wüthrich, M. V. (2016). Understanding reporting delay in general insurance. Risks, 4(3), Article 25. DOI