Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Incurred but Not Reported Claims Estimation

Abstract. Incurred but not reported (IBNR) claims estimation reconstructs insurance obligations associated with insured events that occurred before a valuation boundary but have not entered the insurer's observed claim inventory. The problem joins an economic obligation, a partially observed event process, an administrative reporting mechanism, and a future payment distribution. Each component changes the estimand. This monograph establishes a precise distinction between pure IBNR, development on reported claims (IBNER), reported-but-not-settled liabilities, and total unpaid claims. It derives count- and amount-based estimators from reporting-delay distributions, contrasts them with aggregate chain-ladder and expected-loss approaches, and examines stochastic models, late-emerging claims, data defects, calendar effects, claim severity, and financial reporting. Two reproducible synthetic calculations show how reporting maturity, development factors, and changing claims administration alter reserve estimates. The analytical conclusion is that IBNR estimation is an inference from the observed reporting mechanism as well as from the insurance loss process: a stable historical triangle can conceal changes in reporting speed, severity selection, coverage, and claims handling. Robust estimation therefore requires an explicit valuation perimeter, reconciled data, distinct reporting and payment models where feasible, multiple coherent methods, time-respecting validation, and transparent uncertainty. The study synthesizes established actuarial research and professional requirements; it establishes no new actuarial theory or empirical finding.

Keywords: Incurred but not reported; IBNR; IBNER; loss reserving; claims reporting delay; reporting lag; chain ladder; Bornhuetter–Ferguson; claims development; stochastic reserving; marked point processes; actuarial uncertainty; IFRS 17; Solvency II; data quality; retrospective validation.

Introduction

At an accounting date, an insurer knows some claims and owes for other insured events about which it has yet to receive a notification. Reporting is an observation process that takes place after occurrence, often through several intermediaries. A reserve must incorporate the consequences of the events that have already happened, although their number, size, or payment trajectory remains uncertain. An accident year's apparent maturity is therefore an interaction among exposure, occurrence, reporting, investigation, litigation, case estimates, payments, and closure.

The central research question is: what can be inferred about obligations from events occurring before a defined date when the insurer sees only reports received by a second defined date? Three subsidiary questions follow. What data and assumptions identify the missing events? How should claims reporting be separated from the subsequent development of their cost? What amount and distribution are appropriate for an actuarial or accounting purpose? The question spans general insurance, liability coverage, reinsurance, health insurance, and other contracts, while their legal terms and professional standards remain specific.

The early foundational discussion of IBNR actuarial practice by Bornhuetter and Ferguson (1972) emphasized the problem of using claims experience before full maturity. Schnieper's (1991) separation of true IBNR and IBNER demonstrated the analytical importance of separating late reporting from development of known claims. Mack's (1993) chain-ladder standard error and England and Verrall's (2002) stochastic reserving review place reserve uncertainty at the center of estimation rather than treating a single point figure as the full result.

This dossier uses a structured comparison of primary actuarial publications, standards issued by relevant institutions, independently derived mathematical identities, and hypothetical numerical examples. Claims about individual articles rest on the examined publisher descriptions or accessible institutional text; proprietary insurer datasets and paywalled full texts are not represented as independently replicated. Assumptions are stated beside formulas. All examples are constructions for explaining methods, not assertions about actual insurers. The treatment develops original comparative conclusions within existing research traditions.

Claims Occurrence, Notification, and the Valuation Perimeter

Event, report, payment, and accounting times

Let TT denote the time a covered occurrence or contractually defined incurral takes place, RR the time the insurer recognizes a claim report, and D=R−TD=R-T the reporting delay. For a conventional occurrence-based observation D≥0D\ge 0. Let AA be the accounting date determining which events belong to the incurred population, and VV be the valuation data cutoff. The most common exercise has A=VA=V; nevertheless, an accounting cutoff, data transaction cutoff, and review-of-information date can differ materially. A policy claim reported by VV may be supported by an incurred event dated before AA, while an event after AA lies outside the incurred cohort.

The U.S. unpaid-claim estimation standard ASOP 43 explicitly distinguishes accounting, valuation, and review dates and requires clarity about intended measures, gross versus net amounts, claims covered, and expense scope. An actuary must first determine whether the calculation concerns the ultimate undiscounted loss, the future payment obligation, a statutory booked reserve, an accounting fulfillment-cash-flow component, or a designated risk measure. They answer related but different financial questions.

Define NAN_A as the number of covered claims incurred by AA, NA,VrepN_{A,V}^{\mathrm{rep}} as those in that cohort reported by VV, and NA,Vunrep=NA−NA,VrepN_{A,V}^{\mathrm{unrep}}=N_A-N_{A,V}^{\mathrm{rep}}. This is a latent population count, not necessarily a count of complaints in the claims system: events may be subject to deductibles, attachment points, exclusions, duplicate records, notification thresholds, or uncertain coverage. The incurral definition must match contractual wording, the insurance class, and the valuation purpose.

Event time is a legal and operational choice

For some occurrence policies, physical loss supplies a natural accident date. A disease, progressive injury, professional liability exposure, or contractual trigger can make incurral more complex. A claims-made contract may cover claims first made, or made and reported, within specified periods; reserve accounting follows the actual coverage trigger and any relevant reporting conditions. Applying an occurrence-lag estimator to all such policies without classifying attachment and notification terms can project events outside the insurer's obligation. This is a contract-boundary identification problem to be resolved before estimation.

For health insurance, claims often reach the payer after service through providers, billing systems, and benefit adjudication. The relevant ASOP 5 incurred-claims standard describes incurral dates, processing patterns, organizational administration, plan provisions, and benefits. ASOP 5 governs its stated U.S. professional scope; ASOP 43 has a different property/casualty scope. The distinction matters when transferring methods between medical and casualty payment systems.

Observation histories, rather than a snapshot alone

A claim record at VV combines several clocks: date of occurrence, first report, reopening, notice to insurer, case estimate dates, partial payments, final settlement, recovery receipt, and any change of data extraction or financial close. Reports received by a broker but entered into an insurer system later can have distinct timestamps. A change in reporting interface can alter the measured distribution of DD without changing insured losses. Historical snapshots are essential because an overwritten current field cannot reconstruct what was known at a past valuation.

The event-report distinction is developed formally in Norberg's (1993) prediction of outstanding liabilities, which separates settled, reported-unsettled, incurred-unreported, and covered-unincurred categories. This partition helps align observation windows with the insurance liability boundary.

Identification of Pure IBNR and IBNER

A claims status decomposition

Pure IBNR denotes obligations for events incurred before the cutoff that have not been reported by the observation cutoff. Incurred but not enough reported (IBNER) denotes additional ultimate cost, relative to the currently recognized valuation, on claims already reported. Reported-but-not-settled (RBNS) concerns the future cost of known claims, potentially including payments after the valuation date and claims adjustment expenses. Some accounting and industry conventions use the label IBNR more broadly to include all development beyond case estimates. Consequently the label alone cannot identify an amount; the ledger and intended accounting definition govern.

Let UU be the ultimate undiscounted gross cost of claims incurred through AA; PVP_V the cumulative payments on these claims through VV; CVC_V the total outstanding case estimates for reported claims at VV; and EVrepE_{V}^{\mathrm{rep}} the conditional expected ultimate total for the already reported claims, including payments. Define:

PureIBNRA,V=E⁡[∑k:Tk≤A,  Rk>VSk∣FV],\mathrm{PureIBNR}_{A,V}=\operatorname{E}\left[\sum_{k:T_k\le A,\;R_k>V}S_k\mid\mathcal F_V\right],

where SkS_k is each claim's ultimate gross cost and FV\mathcal F_V is information available at valuation. Define expected development on reported claims as

IBNERA,V=EVrep−(PV+CV).\mathrm{IBNER}_{A,V}=E_V^{\mathrm{rep}}-(P_V+C_V).

When terms share the same expense and coverage perimeter, the expected total unpaid liability has the additive identity

L^A,Vunpaid=CV+IBNERA,V+PureIBNRA,V=U^A,V−PV.\widehat L^{\mathrm{unpaid}}_{A,V} =C_V+\mathrm{IBNER}_{A,V}+\mathrm{PureIBNR}_{A,V} =\widehat U_{A,V}-P_V.

The IBNER quantity may be negative if reported case estimates are collectively higher than expected final settlement amounts. Its sign is therefore informative about case adequacy. These definitions use a gross undiscounted ultimate-cost basis; paid recoveries, salvage, reinsurance, discounting, and different expense inclusions require explicit reconciliation.

Decomposition depends on the reference ledger

An incurred triangle usually records cumulative paid plus outstanding case estimates for reported claims. An aggregate projection from that triangle estimates total development, which commonly contains both the emergence of currently unreported claims and revision of known-claim case estimates. The projected excess over current booked incurred amounts cannot automatically be called pure IBNR. The decomposition requires reporting status, case data, or additional assumptions.

Schnieper (1991) examined the separation problem. Hesselager's (1994) state-transition treatment of reserving provides another route to separating IBNR and reported-but-not-settled cost within an individual-claim process. Both indicate why aggregate ultimate-minus-paid estimates are conceptually wider than the late-reporting population.

Example of a complete decomposition

Suppose a synthetic portfolio has 120 units already paid, 80 units in outstanding case estimates on reported claims, expected IBNER of 15 units, and expected ultimate cost of 25 units for events not yet reported. Then the ultimate expected incurred cost is $120+80+15+25=240$ units; unpaid expected claims are $240-120=120$ units. The 120120 unpaid splits into 80 case, 15 development on reported claims, and 25 pure IBNR. Calling all 40 additional units above case estimates pure IBNR would conflate separate mechanisms; calling the entire 120 units IBNR would instead reproduce a broad accounting convention, not the scientific component defined here.

Reporting-Lag Distributions and Observation Mechanisms

Right-truncated reporting observation

At valuation VV, a claim occurring at t≤Vt\le V enters the reported database only if D≤V−tD\le V-t. Define the conditional reporting cumulative distribution

FD(d∣t,z)=Pr⁡(D≤d∣T=t,Z=z),d≥0,F_D(d\mid t,z)=\Pr(D\le d\mid T=t,Z=z),\qquad d\ge0,

where zz represents risk, claim, channel, product or operational characteristics known for the target population. The reporting probability for an occurrence at tt is q(t,z;V)=FD(V−t∣t,z)q(t,z;V)=F_D(V-t\mid t,z). An event near the valuation date has less opportunity to be reported than an old event; old claims observed now are selected toward longer delays. The data are observed through a calendar-based reporting triangle with a triangular truncation frontier.

A key selection consequence follows: the empirical distribution of delays among claims already reported is not generally the delay distribution for all incurred claims. Recent long-delay claims are missing from that empirical sample. Fitting a delay distribution directly to recent reported records as though all delays had completed systematically exaggerates early reporting. The likelihood or estimating procedure must respect each occurrence period's available follow-up.

Under a discrete model, let pj=Pr⁡(Dp_j=\Pr(D falls in development period j)j) and Fa=∑j=0apjF_a=\sum_{j=0}^{a}p_j. If nijn_{ij} is the number of claims from occurrence period ii reported at age jj, only observed cells satisfying the valuation boundary belong to the likelihood. Given known total occurrence count NiN_i, counts are multinomial with cell probabilities pjp_j and a latent unreported remainder of probability $1-F_{a_i}.If. IfN_i$ is unknown, claim frequency or other mature cohorts are required to identify its scale. Inferring both occurrence rate and reporting probabilities from a single immature total is generally underidentified.

Competing mechanisms and conditional distributions

Reporting delays can depend on claim size, type, litigation complexity, deductible, geography, season, distribution channel, broker practices, and reporting incentives. The mechanism may also change after electronic intake, staff disruption, catastrophe congestion, or a change in benefit processing. A single portfolio-average FDF_D may give an accurate historical total yet misallocate cost across recent cohorts if severity and delay depend on one another.

The individual-claim micro-level reserving study by Antonio and Plat (2014) explicitly considers occurrence, report lag, payment times, payment size, and settlement in a unified longitudinal dataset. Its contribution is relevant as a methodological alternative to compressed runoff triangles. Empirical superiority reported for its particular case study is not evidence that one micro-level algorithm dominates for all portfolios.

Hazard and tail behavior

For a continuous delay DD, the density fD(d)f_D(d), survival function SD(d)=1−FD(d)S_D(d)=1-F_D(d) and hazard hD(d)=fD(d)/SD(d)h_D(d)=f_D(d)/S_D(d) describe the timing of reports among still-unreported events. A piecewise-constant hazard can represent different reporting rates in the first week, month, quarter, and subsequent years; a mixture can represent prompt routine claims and delayed legal claims. Parametric exponential or Weibull models are convenient yet impose tail shapes that require diagnostics. A semi-parametric hazard or discrete-time generalized linear model can incorporate covariates, with stabilization or partial pooling for sparse strata.

A tail probability of even one percent can be material when per-claim severity or aggregate exposure is high. This makes the final reporting tail an actuarial assumption with direct financial significance. A tail factor must be supported by sufficiently mature comparable cohorts, external but relevant experience, contract knowledge, or structured scenario evidence; an arbitrary extra development year does not establish that the liability has expired.

Estimating the Frequency of Unreported Claims

Incidence and reporting as a marked process

Suppose covered events arise with incidence intensity λ(t,z)\lambda(t,z) and reporting delays have conditional CDF FDF_D. Under the usual independent marking conditions of an inhomogeneous Poisson arrival process, expected reports received by VV among events incurred through AA equal

E⁡NA,Vrep=∫−∞A∫λ(t,z)FD(V−t∣t,z) dz dt.\operatorname{E}N^{\mathrm{rep}}_{A,V} =\int_{-\infty}^{A}\int \lambda(t,z)F_D(V-t\mid t,z)\,dz\,dt.

Expected unreported count is

E⁡NA,Vunrep=∫−∞A∫λ(t,z){1−FD(V−t∣t,z)} dz dt.\operatorname{E}N^{\mathrm{unrep}}_{A,V} =\int_{-\infty}^{A}\int \lambda(t,z)\{1-F_D(V-t\mid t,z)\}\,dz\,dt.

Both expressions require that zz has a defined measure, and that occurrence time aligns with coverage. They are expectation identities under the assumed occurrence/report process; they are not a claim that the insurer observes the events absent from its system. Marked Poisson arrivals yield independent Poisson counts for disjoint report/unreported classes conditional on the fixed intensity and mark law. Random common intensity, catastrophes, and varying exposures change unconditional dependence.

Norberg (1993) models outstanding-claim arrival classes through marked Poisson processes. Badescu, Lin, and Tang (2016) develop a marked Cox model for IBNR counts incorporating temporal dependence and fluctuating exposure. The latter is important where a Poisson model understates count variation because risk intensity changes over time.

A single-cohort completion estimator

For an occurrence cohort of comparable risks, suppose CaC_a claims have been reported by development age aa, and a well-identified reporting completion proportion is Fa>0F_a>0. The standard completion estimate is

N^ult=CaFa,N^IBNR=Ca(1Fa−1).\widehat N^{\mathrm{ult}}=\frac{C_a}{F_a}, \qquad \widehat N^{\mathrm{IBNR}}=C_a\left(\frac1{F_a}-1\right).

The estimator assumes that the observed count is representative under a stable applicable delay law. In a Poisson thinning model with fixed known FaF_a and E[Ca]=ΛFaE[C_a]=\Lambda F_a, the estimator of ultimate expected claim frequency is unbiased for Λ\Lambda. That expectation statement is conditional on the specified model, not a universal unbiasedness guarantee for any business calendar. If FaF_a is estimated from other cohorts, uncertainty in that estimate must be included.

For FaF_a to be transferable, comparable strata must have similar reporting processes, or the model must adjust for changes in reporting covariates. Multiplying counts by a severity average from the observed claims can compound selection error if slow-reporting claims have different severities.

Predictive precision for a simple thinning construction

Let total claim count be Poisson with mean Λ\Lambda, let reporting completion equal a known p∈(0,1)p\in(0,1), and let C∼Poisson(Λp)C\sim\mathrm{Poisson}(\Lambda p) and M∼Poisson(Λ(1−p))M\sim\mathrm{Poisson}(\Lambda(1-p)) be independent counts of reported and unreported claims. If Λ\Lambda were known, the optimal squared-error predictor of MM is Λ(1−p)\Lambda(1-p), with conditional process variance Λ(1−p)\Lambda(1-p). If Λ\Lambda is estimated solely as Λ^=C/p\widehat\Lambda=C/p, the plug-in predictor is M^=C(1−p)/p\widehat M=C(1-p)/p.

An unconditional mean squared prediction error, considering the random reported count and future unreported count, is

E⁡{(M−M^)2}=Λ(1−p)+Λ(1−p)2p=Λ1−pp.\operatorname{E}\{(M-\widehat M)^2\} =\Lambda(1-p)+\Lambda\frac{(1-p)^2}{p} =\Lambda\frac{1-p}{p}.

The first term is unreported claim-count process variation; the second arises from estimating incidence through CC with known pp. Additional uncertainty from estimating pp itself increases the realistic prediction uncertainty. This exact algebra is restricted to the stipulated Poisson construction. It also demonstrates why a variance reported only for the unobserved event process can materially understate the uncertainty of an estimated IBNR reserve.

Severity, Dependence, and Monetary IBNR

Counts and amounts are different estimands

For event kk, let SkS_k be its ultimate covered gross loss including only the expense categories specified in the valuation. The pure IBNR amount is the sum of SkS_k for events occurring by AA and remaining unreported by VV. In a marked Poisson model its expected value is

E[PureIBNRA,V]=∫−∞A∫λ(t,z)E[S 1{D>V−t}∣t,z] dz dt.E[\mathrm{PureIBNR}_{A,V}] =\int_{-\infty}^{A}\int \lambda(t,z) E[S\,\mathbf 1\{D>V-t\}\mid t,z]\,dz\,dt.

Only under conditional independence between severity SS and reporting delay DD does this simplify to

∫−∞A∫λ(t,z) μS(t,z)[1−FD(V−t∣t,z)] dz dt.\int_{-\infty}^{A}\int \lambda(t,z)\,\mu_S(t,z) [1-F_D(V-t\mid t,z)]\,dz\,dt.

The factorization is a genuine modeling assumption. Severe bodily injury claims may report promptly while their cost develops slowly; a latent injury can report very late and ultimately be large. Both patterns violate a naive universal independence assumption. Conditioning on sufficiently granular features can help, but information about a claim's actual severity is unavailable for unreported events. The actuary needs mature experience, external comparisons, or credible structural assumptions about the latent severity distribution.

Aggregate amount models and changing mix

Aggregate paid and incurred triangles encode monetary development without explicitly modeling report incidence. They may work well under stable experience precisely because reporting and settlement behavior jointly create a persistent pattern. A structural break in one component can destabilize the other. Mack (1991) connected IBNR estimation and cross-classification models, including the connection between multiplicative patterns and chain ladder.

When rare claims dominate cost, count completeness can approach one while amount completeness stays low. Conversely, rapid recognition of large claims can make cumulative incurred cost appear highly mature even as many small claims remain missing. A reserve report should therefore disclose whether its percentages refer to claim numbers, reported incurred dollars, or cumulative paid dollars.

Severity process and claim evolution

Reported case estimates are themselves revisions of anticipated severity, contingent on medical information, legal decisions, investigations, repair estimates, inflation, and settlement strategy. The event history comprises both a reporting process and, for reported claims, an evolving payment/valuation process. A stochastic model may treat SS as one ultimate mark; a more detailed model treats partial payment and case revision as events over time. These choices change the available information and the inferential target. Aggregate dollar estimates, while operationally simpler, cannot explain every source of their own emergence.

Development Triangles and Aggregate Reserving

Basic definitions

For occurrence period ii and development age jj, let XijX_{ij} denote incremental paid or incurred amounts as carefully specified, and Cij=∑k≤jXikC_{ij}=\sum_{k\le j}X_{ik} cumulative amounts. At a valuation cutoff, only cells with i+j≤Ki+j\le K are observed under a regular index scheme. The historical period-by-age matrix is a triangle. A triangle for reported counts measures reporting maturity; an incurred-loss triangle mixes counts, case estimates, and payment recognition; a paid-loss triangle measures disbursement maturity.

The classical chain-ladder age-to-age development factor, estimated over cohorts with both ages observed, is

f^j=∑i∈IjCi,j+1∑i∈IjCij,\widehat f_j= \frac{\sum_{i\in I_j} C_{i,j+1}}{\sum_{i\in I_j} C_{ij}},

where IjI_j is the exact set of comparable cohorts. For the most recent observed age aia_i, the projected ultimate is

U^i=Ci,ai∏j=aiJ−1f^j f^tail,\widehat U_i=C_{i,a_i}\prod_{j=a_i}^{J-1}\widehat f_j\,\widehat f_{\mathrm{tail}},

and projected development is U^i−Ci,ai\widehat U_i-C_{i,a_i}. For a paid triangle that difference estimates unpaid amounts on the selected ultimate basis. For an incurred triangle it estimates additional development beyond current incurred valuation, often a mixture of pure IBNR and IBNER. The distinction must be retained in subsequent decomposition.

Statistical reading and limits

Chain ladder is a form of conditional development extrapolation. It relies on historical experience being informative about the development of remaining cohorts and on sufficient coherence of age-to-age factors. The data often include zero, negative, or extraordinary increments. Negative case reserve changes can lead to decreasing incurred cumulative values or unstable factors; forcing every factor above one hides legitimate release experience.

Mack's (1993) distribution-free chain-ladder approach derives reserve standard-error expressions under specified conditional moment and cross-origin independence assumptions. It does not establish distribution-free robustness against inflation regime changes, altered claim definitions, or systematic reporting delay. The phrase distribution-free refers to the parametric distributional assumptions in that mathematical method, not to the absence of assumptions altogether.

England and Verrall (2002) consider stochastic extensions and predictive distributions, including generalized linear models, bootstrap approaches and methods for tails. Wüthrich and Merz (2008) give a systematic treatment of stochastic reserving and prediction error. Both emphasize that projecting expected development and measuring variability answer different questions.

Accident, development, and calendar dimensions

The same triangle cell corresponds to calendar time c=i+jc=i+j. Accident-period effects represent changes in the business underwritten or events incurred; development-period effects represent typical reporting or settlement maturity; calendar effects represent inflation, court decisions, bulk claims booking, system migration, or claims handling changes affecting several ages simultaneously. A pure age-factor technique may confound calendar changes with development and generate misleading recent-cohort ultimates.

The mathematical index identity c=i+jc=i+j also creates exact linear dependence in additive accident/development/calendar parameterizations. Imposing identification constraints, such as reference periods or smoothness, is necessary to estimate identifiable contrasts. Numerical convergence alone does not imply that a separate calendar inflation and development effect has been empirically identified. Reserve sensitivity across reasonable decompositions is often more informative than the appearance of a uniquely estimated parameter.

Expected-Loss and Bornhuetter–Ferguson Methods

Expected loss as an external prior input

A pure experience-development projection can be unstable on an immature cohort whose earliest reported claims are unusually sparse or severe. Expected-loss approaches start from exposure and an a priori loss ratio, severity-frequency estimate, or other justified expected ultimate amount EiE_i. The a priori amount must be defined independently enough to supplement the immature observations and must reflect pricing adequacy, policy limits, changing exposure and trend.

For an age aa with estimated reported incurred proportion ra=1/CDFar_a=1/\mathrm{CDF}_a where CDFa\mathrm{CDF}_a is the ultimate development factor, a conventional Bornhuetter–Ferguson expression is

U^iBF=Ci,a+Ei(1−ra),Δ^iBF=Ei(1−ra).\widehat U_i^{\mathrm{BF}} =C_{i,a}+E_i(1-r_a), \qquad \widehat{\Delta}_i^{\mathrm{BF}}=E_i(1-r_a).

Here rar_a refers to the selected cumulative amount-development maturity, not necessarily the probability an individual claim has been reported. Using a count reporting-completion proportion in this amount equation without showing compatible severity maturity is a category error.

Bornhuetter and Ferguson's (1972) original IBNR discussion supplies the historical reference point. England and Verrall's (2002) Bayesian treatment of expected-loss inputs illustrates one route to formally expressing prior-information uncertainty. BF is not immune to error: an inflated expected loss ratio, an incorrect reporting percentage, or an unrepresentative portfolio will bias the result.

Complementary rather than interchangeable information

Chain ladder emphasizes observed development. Expected-loss methods emphasize independently supported future expectation. When a cohort is very immature, BF limits the mechanical scaling of a volatile first development cell, but it imports whatever bias sits in the expected ultimate. As the maturity fraction approaches one, its prospective addition diminishes. A selection between methods therefore depends on actual exposures, credibility, stability, purpose, and information quality. A transparent comparison may combine them for governance, provided the combination's rationale and dependence are disclosed rather than pretending two methods yield independent evidence.

Stochastic Claims Reserving and Predictive Uncertainty

From conditional point estimates to forecast distributions

A deterministic ultimate estimate is a functional of observed data and assumptions. A stochastic reserve model additionally assigns probability to future development. Its output may be an expected reserve, a predictive standard deviation, a distribution of one-year remeasurement, or a lifetime cash-flow distribution. These are not the same risk measure. The choice must follow the use: reserve adequacy, pricing, earnings attribution, solvency capital, or discounted financial reporting.

A common mean square error of prediction decomposition expresses the expectation of squared future deviation as process variability plus uncertainty arising from estimated parameters, with covariance terms treated where present. If θ\theta represents a random or estimated parameter vector, the law of total variance gives

Var⁡(Y∣FV)=E[Var⁡(Y∣θ,FV)∣FV]+Var⁡(E[Y∣θ,FV]∣FV).\operatorname{Var}(Y\mid\mathcal F_V)= E[\operatorname{Var}(Y\mid\theta,\mathcal F_V)\mid\mathcal F_V] +\operatorname{Var}(E[Y\mid\theta,\mathcal F_V]\mid\mathcal F_V).

This formula is exact when θ\theta is modeled probabilistically under the stated conditional structure. In frequentist treatments one typically derives an analogous process-plus-parameter approximation under repeated sampling; interpreting those two objects as identical without stating a framework is misleading. Model-form error, changing operations, and legal surprises sit outside a narrowly fitted predictive distribution unless explicitly modeled.

Mack (1993) derives an uncertainty measure for chain ladder under moment assumptions, while England and Verrall (2002) examine predictive distributions and bootstrap methods. Wüthrich and Merz (2008) organize stochastic reserving models and their parameter and process uncertainty. Each supports a model-based assessment; none implies that a narrow estimate automatically covers a change in the underlying coverage or data process.

Generalized linear and Bayesian approaches

For incremental amounts XijX_{ij}, an overdispersed Poisson-type model may specify E[Xij]=μijE[X_{ij}]=\mu_{ij}, Var⁡(Xij)=ϕμij\operatorname{Var}(X_{ij})=\phi\mu_{ij}, and log⁡μij=αi+βj\log\mu_{ij}=\alpha_i+\beta_j for origin and development factors. The mean pattern can reproduce chain-ladder-like projections under suitable conditions. The dispersion parameter ϕ\phi and estimation uncertainty determine predictive variability. Such a model requires appropriate treatment of negative increments, structural zeros, data aggregation, and large claims; the simple Poisson variance function is neither a guarantee of fit nor a universal distribution for incurred dollar values.

A Bayesian model assigns a prior distribution to parameters or latent development effects and updates it with observed claims. This can incorporate prior knowledge of expected losses or tail behavior and yield a posterior predictive distribution. Prior influence should be reported for immature cohorts because scarcity of data can leave the result heavily dependent on assumptions. A posterior interval conditional on an incomplete structural model still excludes unmodeled regime shifts. For complicated claims, scenario ranges can be more intelligible than highly precise parametric percentiles.

One-year versus ultimate uncertainty

Lifetime runoff uncertainty concerns total remaining payments until claim settlement. One-year reserve-development risk concerns how the best estimate changes after another calendar year, including payments and the revised expected outstanding liability. An insurer may face a material one-year reserve deterioration even if the ultimate mean estimate at the initial date was unbiased over many comparable portfolios. Governance must identify which future information set and horizon underlie the reported distribution.

Micro-Level and Event-History Models

The observable claims ledger as a longitudinal process

Individual-claim models use event times rather than only triangle cells. A claim begins with an unobserved occurrence, may be reported after a delay, enters a case reserve and payment process, and moves among open, reopened, litigated, settled, and closed states. Different hazards or payment intensities can be conditioned on covariates, claim age, and calendar circumstances. The likelihood must include the truncation of the not-yet-reported population and the incomplete observation of open claims.

Antonio and Plat (2014) analyze individual-claim settlement and reporting histories. Hesselager (1994) derives Markov-process IBNR and RBNS expressions. A micro-level model can preserve information discarded by a triangle: rapid versus slow reporters, reopened claims, claim-specific payment patterns and within-year changes. Its additional parameters, incomplete covariates and data engineering can also make it less stable than a parsimonious aggregate method.

Overdispersion, dependent arrivals, and filtering

Suppose the insurer has a random time-varying incidence intensity Λ(t)\Lambda(t) conditioned on latent environmental or operational factors. A Cox process makes arrivals Poisson conditional on the intensity path but overdispersed when the intensity is integrated out. Reporting delays can be added as marks, and claims reported by the valuation date provide partial information about the latent intensity. Badescu, Lin, and Tang (2016) investigate a marked Cox IBNR-count construction; Badescu, Chen, Lin, and Tang (2019) subsequently develop its estimation and application. Avanzi, Wong, and Yang (2016) analyze overdispersion and reporting delays through a micro-level count model. These contributions make temporal dependency a testable modeling concern, not an automatic reason to increase reserve estimates.

An insurer can evaluate whether a random-intensity model improves sequential forecasts against a simpler completion model. Increased flexibility should earn its place through calibrated predictions, meaningful sensitivities, and accurate reconciliation to the accounting perimeter. Small, heterogeneous datasets may favor aggregation and explicit expert assumptions; rich, stable datasets may support finer-grained models.

Structural identifiability

Observed reports generally identify products of claim incidence and reporting probabilities more readily than the two components separately. For a cohort with only mm reported claims at maturity fraction pp, infinitely many pairs (Λ,p)(\Lambda,p) can produce the same expected count m=Λpm=\Lambda p. More reporting-age cells and mature historical cohorts help identify pp, while exposure and risk information help identify Λ\Lambda. Unobserved changes that affect both processes in opposite directions can remain indistinguishable in an aggregate triangle.

A model's complexity cannot overcome absent identifying information. This leads to an analytical rule of practice derived from the observations: each material unobservable claim component must be supported either by a stable observed transition, credible external frequency or severity information, contractually grounded constraints, or an explicitly documented sensitivity range. This is a methodological conclusion from existing inference theory, not the establishment of a new named doctrine.

Synthetic Calculations and Reproducibility

All values in this chapter are hypothetical. Numbers are retained with sufficient precision to reproduce the calculations and rounded only for presentation.

Case study: frequency completion and reporting delay

Assume an occurrence cohort has 160 reported claims at the valuation date. A comparable mature population gives an assumed reporting completion fraction p=0.80p=0.80 at that age. The count-completion estimate is $160/0.80=200$ ultimate claims and $200-160=40$ unreported claims. If assumed total mean occurrence count is Λ=200\Lambda=200 and the Poisson-thinning assumptions hold, the plug-in count-prediction mean squared error from the earlier derivation is $200(0.20)/0.80=50,correspondingtoarootmeansquaredpredictionerrorof, corresponding to a root mean squared prediction error of\sqrt{50}\approx7.071$ claims. This is a model-specific illustration, not a credible interval for actual business.

Change only the assumed maturity to p=0.65p=0.65, holding 160 observations fixed. The ultimate estimate becomes $160/0.65=246.153846,andtheexpectedunreportedcomponentbecomes, and the expected unreported component becomes86.153846.Theincreasefrom40isapproximately46.154claims.With. The increase from 40 is approximately 46.154 claims. Withp=0.75,theIBNRestimateis, the IBNR estimate is53.333333;with; withp=0.85,itis, it is28.235294.Thesesensitivitiesarelargebecausematurityappearsinthedenominator.Aseeminglysmallcalibrationerrorin. These sensitivities are large because maturity appears in the denominator. A seemingly small calibration error inp$ can dominate random Poisson fluctuation for a recent cohort.

To illustrate temporal drift rather than uncertainty in the input, suppose the actual cohort contains 100 incurred claims and that a process change reduces the current reporting completion from the historical 80% to 60%. The insurer sees 60 reports. Using 80% maturity estimates 75 ultimate and 15 unreported claims; the actual unreported count in this construction is 40. The 25-claim shortfall is produced entirely by the missed change in reporting speed. The estimator can fit past cohorts accurately and still fail systematically on the current cohort.

Case study: severity selection hidden by count maturity

Suppose 100 insured claims have occurred. Eighty are promptly reported and each has an ultimate severity of $1,000. Twenty are delayed and each ultimately costs $5,000. The observed group totals $80,000, the unreported group totals $100,000, and true ultimate total cost is $180,000. A count completion model correctly infers 20 missing claims from an 80% reporting fraction, but multiplying by observed average severity of $1,000 predicts only $20,000 of pure IBNR. This understates the synthetic pure IBNR amount by $80,000 because reporting delay is associated with severity.

This construction demonstrates why separate count estimation and amount estimation are necessary. It does not imply that slow-reported claims are generally more expensive: prompt catastrophic claims may generate the opposite selection. The analyst must investigate the empirical direction and stability of reporting/severity dependence by line, claim type, attachment layer, and contract.

Case study: a complete paid triangle

Consider the following cumulative paid-loss triangle, in arbitrary monetary units, at four development ages. Empty cells are not yet observed at the valuation date:

Origin cohort Age 0 Age 1 Age 2 Age 3
A 1,000 1,600 2,000 2,200
B 1,200 1,800 2,100 —
C 1,500 2,100 — —
D 1,800 — — —

Using volume-weighted age factors gives f0,1=(1600+1800+2100)/(1000+1200+1500)=5500/3700=1.486486486f_{0,1}=(1600+1800+2100)/(1000+1200+1500)=5500/3700=1.486486486, f1,2=(2000+2100)/(1600+1800)=4100/3400=1.205882353f_{1,2}=(2000+2100)/(1600+1800)=4100/3400=1.205882353, and f2,3=2200/2000=1.10f_{2,3}=2200/2000=1.10. Assume an additional tail factor of exactly one solely to keep this illustration self-contained. Projected ultimate amounts are A $2,200; B \2,310; C \2,785.588235; and D \3,549.205087. Total projected ultimate is \10,844.793322. Cumulative paid at valuation totals \8,200, so the projected unpaid amount is \2,644.793322$.

The estimate describes total unpaid amounts under paid-development assumptions. It does not directly identify pure IBNR. A reported-claims inventory, case estimates or an independently calibrated report-lag model is required to separate the unpaid amount into the components defined earlier. Treating this paid-development residual as only not-yet-reported loss would misclassify remaining payments on already reported claims.

For cohort D, the selected cumulative paid-development factor is fD=(5500/3700)(4100/3400)(1.10)=1.971780604f_D=(5500/3700)(4100/3400)(1.10)=1.971780604, and its associated paid maturity is approximately rD=0.507155r_D=0.507155. If a separately justified expected ultimate amount is ED=3,200E_D=3,200, a BF-style estimate becomes $1,800+3,200(1-1/f_D)\approx3,377.10. The chain-ladder result is about \3,549.21$. The difference arises from different weights placed on cohort D's early paid emergence and its expected-loss input; both require review against actual portfolio data.

Case study: accounting reconciliation

Returning to the earlier decomposition, take expected ultimate incurred 240240, paid 120120, current case 8080, IBNER 1515, and pure IBNR 2525 units. The gross undiscounted unpaid estimate is 120120. Assume separately estimated recoverable cash flows of 30 units before credit and timing adjustments, so a simple illustrative net unpaid amount is 90 units before any effects of discounting or noncollection. This example does not reproduce the separate best-estimate, risk-margin, contractual-service-margin, or financial statement procedures under any specific reporting regime. Its purpose is to expose arithmetic double counting before applying those regimes.

Delays, Inflation, Operational Discontinuity, and Tail Risk

Administrative and behavioral delay shifts

Reporting latency is partly determined by claims administration, which responds to operational capacity, broker networks, electronic notification, medical billing, and social behavior. A new online portal can accelerate first notifications while case development becomes slower due to more complete verification. A catastrophe can increase incurred claims and temporarily delay reporting because adjuster capacity is constrained. A backlog release can generate an unusual surge in later development ages. Each event can resemble a change in loss frequency if the analyst observes reports without their original occurrence dates.

ASOP 5 expressly discusses claims administration and processing changes in health incurred-claims analysis. For property/casualty claims, ASOP 43 requires the actuary to consider claim adjustment and external conditions relevant to the unpaid estimate. Translating these professional requirements into a data check means reviewing the reporting pipeline and operational events contemporaneous with a development anomaly.

Inflation and legal-cost severity

For a paid triangle, a change in nominal cost can move several calendar diagonals simultaneously. The actuary may distinguish exposure trend, frequency change, monetary inflation, medical inflation, repair costs, legal settlement environment, and altered claims mix. Explicit inflation adjustments should be consistent with the future timing of payments and the contract's limits, deductibles, index clauses and currency. A calendar effect is not always interpretable as inflation: late case strengthening, batch reconciliation, or a litigation event can create similar aggregate patterns.

For liability lines, delayed notification can coexist with multi-year settlement, reopening, judicial uncertainty and social inflation. The final reporting age and the final payment age are distinct. A well-fitted count-reporting tail tells only how many claims may appear, not how many years the insurer may continue paying. Tail assumptions should be stressed separately for count and monetary emergence.

Boundary and catastrophe stress

A frequency model calibrated to routine claims may fail after a systemic event, a policy change, a reporting law amendment, or a newly recognized class of claims. Stress analyses can vary occurrence intensity, reporting hazards, severity conditional on delay, reopening rates, case reserve adequacy and recovery assumptions together. The combinations should represent plausible coherent circumstances rather than adding isolated worst cases that could not jointly occur. A deterministic reserve margin added to compensate for unidentified model defects may obscure the actual mechanism; a documented sensitivity distribution or scenario table better informs governance.

Data Architecture and Reconciliation

The minimal reproducible dataset

A coherent IBNR data architecture records an immutable claim or event identifier; policy and coverage identifiers; exposure basis; occurrence/incurral time; insured notification time; insurer first-report time; recording time; status changes; paid transactions with dates and signs; case reserve transactions or snapshots; expense categories; recoveries and ceded allocations; currencies; claim counts and duplicate logic; and mapping to the accounting cohort. Data provenance should identify originating operational systems, extraction queries, reference calendars, cutoffs, and later adjustments. The same claim event may involve multiple coverages or exposures; a stable hierarchy of claim, feature, policy, and event prevents inflated counts.

The ASOP 23 data-quality standard describes data selection, review and use, including relevant limitations. Solvency II's Article 82 data requirements and delegated Article 19 data-quality conditions address appropriateness, completeness, accuracy and consistent timing for technical-provision information. These sources operate in different professional and legal domains; their common analytical implication is that reserve models inherit errors introduced by unreliable observation.

Reconciliation invariants

Before modeling, the analyst should reconcile: reported claim counts by occurrence and report period against the claims inventory; paid transactions against the ledger; outstanding case estimates against financial close; gross to net development against recoveries; original and adjusted triangles; and the opening-to-closing reserve movement. In a consistent dataset, at the same valuation and gross perimeter, ultimate equals cumulative paid plus projected future paid. Reported-incurred equals paid plus current case estimates, subject to the exact ledger definition. Deviations require documented adjustments, not hidden balancing factors.

When snapshots are reconstructed, the evaluation must use values actually known by the earlier valuation date. A current final settlement value attached to a historic claim record creates look-ahead leakage if used as though it were available before settlement. Late-entered historical events can be used to score the earlier IBNR forecasts after their outcomes mature, but must be excluded from information used to make those forecasts.

Controls over large and reopened claims

Large losses often distort development factors and aggregate covariance. A principled treatment may segment large claims, make case-specific estimates, analyze attritional losses separately, and maintain a reconciliation to total insured losses. Reopened files can make closure status nonabsorbing. Calendar batch transactions and missing negative recoveries can similarly change apparent triangles. Material changes in data transformation rules should be recorded as model inputs and as independent sources of uncertainty.

Model Validation, Backtesting, and Reserve Governance

Time-respecting retrospective exercises

A retrospective test reconstructs what the analyst would actually have known at each past valuation VkV_k, computes estimates with only available information, and compares them with later-emerging report counts and ultimate payments when sufficiently mature. Randomly splitting current claim records into training and test samples generally leaks future development and may fail to measure the real forecasting task. Valid chronological testing includes changing maturity, reporting seasonality and operational conditions.

For each cutoff, compare predicted IBNR counts with subsequently reported incurred-event counts; compare payment and incurred-development forecasts with observed changes; review signed bias by age and cohort; and inspect whether predictive ranges have reasonable empirical coverage. A prediction interval that covers expected runoff on old cohorts but consistently misses recently changed reporting processes needs revision. Sparse high-severity cohorts often need qualitative diagnostics and wider scenario bounds because apparent interval coverage is statistically noisy.

ASOP 56 addresses model appropriateness and validation, including inputs, assumptions, model change controls and output testing in its U.S. applicability domain. ASOP 43 emphasizes method selection, assumptions and disclosures. These standards justify recording both statistical validation and operational evidence in the reserve review.

Diagnostics by the underlying mechanism

For a reporting-lag model, assess observed-to-expected reports by lag age, occurrence month and reporting month; conditional severity by lag group; late-report counts versus exposure; and residual correlations by calendar period. For an aggregate paid or incurred triangle, inspect age-to-age factors, residuals, calendar diagonals, large-claim effects, negative increments, and data definition changes. For a joint stochastic model, examine predictive residuals and robustness to dependence assumptions. A method can satisfy aggregate ultimate reconciliation while predicting the wrong reporting cohort or settlement timing; diagnostic choice must match the intended use.

A reserve committee can use explicit comparisons among chain ladder, BF, count-severity models, and claim-file review. Agreement is more persuasive when methods use genuinely different information. Multiple models constructed from the same triangle are correlated, so their agreement alone does not yield independent confirmation. Disagreement provides evidence about assumptions that should be investigated and documented.

Decision record and quality hierarchy

A defensible valuation record states the intended claim population, claims and expense definition, dates, materiality threshold, selected methods, source-data provenance, major assumptions, alternatives, uncertainty, sensitivity, movement since the last valuation, and responsible review. Method selection should follow causal and contractual features of the exposure, sufficiency of observations, and the financial purpose. A sophisticated model is not preferable merely because it is sophisticated; an appropriate transparent model with accurate data can outperform a complicated model whose latent assumptions are untestable.

No statistical metric can determine the booked liability without considering the relevant accounting or regulatory standard. No accounting disclosure can cure a biased scientific estimate. The relationship is sequential: identify the obligation, estimate its future consequences, translate those estimates to the required measurement basis, and disclose the judgments that remain.

Accounting, Solvency, and International Applicability

IFRS 17 and the liability for incurred claims

IFRS 17 Insurance Contracts is the internationally issued accounting standard for eligible reporting entities applying that framework; it became effective for annual periods beginning on or after January 1, 2023, subject to applicable adoption. The standard's liability for incurred claims concerns fulfillment of obligations for events already incurred. A statistical IBNR projection is one input to the contractual measurement of future cash flows, whose reporting amount can additionally reflect discounting and risk adjustment for nonfinancial risk under the applicable measurement rules. The sum of a gross undiscounted case reserve and a pure IBNR estimate is therefore not, by itself, an IFRS 17 financial statement line item.

The International Actuarial Association's ISAP 4 on IFRS 17 provides a model actuarial practice standard. Its status as a model means domestic professional standards and applicable law determine whether particular provisions bind an actuary. The insurer also needs to distinguish liabilities for incurred claims from any remaining-coverage liabilities and to maintain consistent claims and acquisition-cash-flow boundaries.

Solvency II and recoverables

In the European Union framework, Solvency II Article 77 requires technical provisions generally to comprise a best estimate and a risk margin, with a probability-weighted present value of relevant future cash flows and realistic assumptions. This is wider than a simple undiscounted IBNR amount. Article 82 requires processes ensuring technical-provision data quality, and Article 81 requires separate valuation of reinsurance recoverables, including timing and counterparty default considerations. The distinction prevents silently netting assumed reinsurance claims against the direct underlying obligation.

EIOPA's revised valuation guidelines further address consistent application of the European framework. Legal requirements must be interpreted for the relevant jurisdiction, reporting date and implementing rules. Their existence does not mean every insurer worldwide uses Solvency II valuation bases.

U.S. actuarial practice and health claims

Under relevant U.S. actuarial professional scopes, ASOP 43 addresses property/casualty unpaid claims, ASOP 5 addresses incurred health and disability claims, and ASOP 23 addresses data. Professional standards guide actuarial work and associated communications; they are not themselves an international reporting ledger or a substitute for jurisdictional financial-reporting rules. For workers' compensation benefits, property/casualty and health classifications may interact in ways explicitly treated by applicable scope rules.

A health claims payment-lag triangle may be highly operationally useful, especially if claim service dates, provider submissions, adjudication and cash disbursements are reliably distinguished. A liability claim may require different development ages and expert claim-file adjustment. International comparisons should normalize these distinct mechanisms before treating their reserve percentages as comparable.

Integrated Interpretation and Limitations

Original analytical synthesis

The preceding evidence supports six connected conclusions about estimation practice.

First, the latent object is defined by covered incurral and observation dates. Moving either boundary changes the target population and may change the financial liability more than changing an estimator. An analysis without contractual and temporal perimeter controls cannot obtain a reliable answer merely through a more accurate numerical method.

Second, reporting is an observation mechanism. A report count records two simultaneous processes—occurrence frequency and reporting timeliness. The identity E[reports]≈incidence×reporting completionE[\mathrm{reports}]\approx \mathrm{incidence}\times\mathrm{reporting\ completion} explains why a speed change can imitate a claim-frequency change. Reliable IBNR work models or controls for both processes and recognizes limits to identification when only immature reports are available.

Third, the composition of missing events determines missing cost. A correctly estimated count can coexist with a materially incorrect dollar reserve when delay and severity are dependent. Gross IBNR amounts should be considered alongside count maturity, severity conditional on reporting patterns, large losses and contract attachment rules.

Fourth, total development and pure IBNR must be separated conceptually even when the available aggregate method cannot isolate them empirically. Paid ultimate-minus-paid describes all future payments; incurred ultimate-minus-current-incurred describes additional recognition; pure IBNR describes the ultimate cost of unreported prior events. Consistent ledger identities reveal mislabeling and double counting.

Fifth, stability assumptions deserve the same scrutiny as statistical fitting. Reporting transformations, settlement inflation, claims-made conditions, reinsurance and new claim categories can be missing from a fitted historical triangle. A well-calibrated model under old operations cannot serve as evidence that the old operating regime continues.

Sixth, reserve quality is an interdisciplinary chain from contract to dataset to model to financial measurement. Actuarial evidence, claims administration, accounting, finance and regulation supply different necessary information. Governance is strongest when the chain can be reconstructed from original dates, records and documented transformations.

These are original integrative inferences drawn from identified research, mathematical consequences, and synthetic examples; they are neither new named theory nor demonstrated empirical discovery.

Limits of the evidence

The researched sources include primary professional and regulatory pages, publisher bibliographic records, accessible article abstracts, and books identified through publisher descriptions. Several full scholarly texts are behind access controls and were not represented as having been read in full. No confidential insurer records were used, and no independent experiment on real-world accuracy was carried out. Formula-based examples establish logical consequences of stated assumptions, not measured reserve adequacy. Model-specific distributions may fail under correlated claims, policy boundary ambiguity, heavy-tailed severity, data migration, or nonstationary reporting. Legal applicability varies by jurisdiction and actual reporting framework.

The practical response is to preserve evidence provenance and model sensitivity rather than claim universal accuracy. This work's international comparisons are descriptive scientific synthesis, not individualized insurance, broking, financial-reporting, or legal advice.

Conclusion

Incurred but Not Reported Claims Estimation studies the recovery of obligations created by insured events before a valuation cutoff but absent from the insurer's reported claim population. Its central mathematical feature is partial observation: future reports depend jointly on historical occurrence and the delay with which those occurrences become visible. Claim costs introduce further uncertainty through severity, settlement, case estimate revision, inflation and recoveries.

The monograph derives the principal count and amount relations, distinguishes pure IBNR from IBNER and total unpaid loss, demonstrates aggregate development and expected-loss calculations, and evaluates stochastic and micro-level alternatives. The synthetic cases establish that reporting-maturity error and delay-severity dependence can dominate otherwise coherent point estimates. The resulting scientific interpretation places contractual coverage definitions, reporting histories, independent data reconciliation, model validation, and transparent uncertainty at the center of actuarial practice.

The actuarial calculation acquires institutional meaning when translated carefully to the intended measurement basis. IFRS 17, Solvency II, and professional ASOPs provide distinct but complementary contexts for assumptions, financial recognition, and disclosure. Appropriate methods follow the risk, contract, available evidence, reporting process, and jurisdiction. The analytical result is a complete causal accounting of what is estimated, from incurred event through observation and eventual payment.

References

  • Actuarial Standards Board. (2011, last revised). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Official standard.

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

  • Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 5: Incurred Health and Disability Claims. Official standard.

  • Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Official standard.

  • Antonio, K., & Plat, R. (2014). Micro-level stochastic loss reserving for general insurance. Scandinavian Actuarial Journal, 2014(7), 649–669. Publisher DOI.

  • Avanzi, B., Wong, B., & Yang, X. (2016). A micro-level claim count model with overdispersion and reporting delays. Insurance: Mathematics and Economics, 71, 1–14. Publisher article.

  • Badescu, A. L., Lin, X. S., & Tang, D. (2016). A marked Cox model for the number of IBNR claims: Theory. Insurance: Mathematics and Economics, 69, 29–37. Publisher article.

  • Badescu, A. L., Chen, T., Lin, X. S., & Tang, D. (2019). A marked Cox model for the number of IBNR claims: Estimation and application. ASTIN Bulletin, 49(3), 709–739. Publisher article.

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

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

  • European Insurance and Occupational Pensions Authority. (2022). Revised Guidelines on Valuation of Technical Provisions. Official publication.

  • European Parliament and Council. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 77: Calculation of technical provisions. EIOPA rulebook.

  • European Parliament and Council. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 81: Recoverables from reinsurance contracts and special purpose vehicles. EIOPA rulebook.

  • European Parliament and Council. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 82: Data quality and application of approximations. EIOPA rulebook.

  • European Commission. (2015, as amended). Commission Delegated Regulation (EU) 2015/35, Article 19: Data used in the calculation of technical provisions. EIOPA rulebook.

  • Hesselager, O. (1994). A Markov model for loss reserving. ASTIN Bulletin, 24(2), 183–193. Publisher article.

  • IFRS Foundation. (2017, subsequently amended). IFRS 17 Insurance Contracts. Effective from annual periods beginning on or after January 1, 2023. Official standard overview.

  • International Actuarial Association. (2019). International Standard of Actuarial Practice 4: IFRS 17 Insurance Contracts. Model standard. Official ISAP 4 page.

  • Mack, T. (1991). A simple parametric model for rating automobile insurance or estimating IBNR claims reserves. ASTIN Bulletin, 21(1), 93–109. Publisher article.

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

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

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

  • Wüthrich, M. V., & Merz, M. (2008). Stochastic Claims Reserving Methods in Insurance. Wiley Finance. Publisher book.