RESEARCH DOSSIER
Large-Loss, Latent Claims, and Long-Tail Reserving
Abstract. Large-loss and long-tail insurance reserves arise where claim severity, the time of emergence, the evolution of legal responsibility, and the date of payment are jointly uncertain. A small number of extraordinary losses may dominate ultimate cost, while latent bodily injury, environmental damage, occupational disease, and other delayed liabilities may emerge years after the original exposure. This monograph develops a source-based account of large-loss thresholds, claim-level excess layers, delayed reporting, occurrence and claims-made coverage, exposure scenarios, claims inflation, actuarial development methods, and the statistical distinction between observed transactions and future liability. It compares paid and incurred development, Bornhuetter–Ferguson estimation, experience-adjusted and scenario methods, frequency–severity models, survival approaches, excess loss factors, and stochastic predictive distributions. It shows analytically why a fixed nominal reporting threshold changes meaning over time, why excess-layer projection cannot be inferred by applying a layer function to aggregate ground-up losses, and why the same apparent tail pattern can arise from inflation, reporting delay, claims handling, case strengthening, or coverage change. Four transparent hypothetical calculations examine a threshold adjustment, an excess-layer expected payment, a long-tail reserve estimate, and a finite scenario distribution, with numerical uncertainty and sensitivity analysis. The international discussion distinguishes actuarial unpaid-claim estimates under ASOP 43, the current ASOP 20 guidance for cash-flow analysis, IFRS 17 accounting measurement, and Solvency II technical provisions. An original analytical synthesis integrates contractual mapping, exposure reconstruction, statistical estimation, scenarios, and governance into an auditable sequence without proposing a new named theory or statistical doctrine. Its central conclusion is that credible long-tail reserving requires distinct estimates of what can be observed, what can legally be insured, when a claim may be reported or paid, and how uncertainty over each component affects the portfolio obligation. All worked figures are synthetic; no empirical insurer-specific reserve conclusion is asserted.
Keywords: large-loss reserving; latent claims; long-tail liabilities; claims development; casualty reserving; reported but not settled; incurred but not reported; incurred but not enough reported; asbestos; environmental claims; social inflation; judicial inflation; excess layers; tail factors; exposure reserves; severity distributions; generalized Pareto distribution; Bornhuetter–Ferguson; Mack method; claims inflation; IFRS 17; Solvency II; reserve uncertainty; sensitivity analysis.
Introduction
Reserving for unusually large and delayed insurance claims is an exercise in incomplete observation. A claims database records notices, payments, case estimates and legal milestones already known at a valuation date. The financial obligation also concerns losses that have occurred but are unreported, losses whose cost will develop beyond today's case estimate, claims subject to lengthy litigation, expenses needed to settle them, and in some contexts recoveries from reinsurers or other parties. These components can evolve over decades. A reserve at a particular date is therefore a conditional estimate based on an information set and a defined measurement purpose.
The governing distinctions must be explicit. An insured event is determined by the relevant coverage and applicable law. A claim report is the arrival of notice in a particular reporting system. A case estimate is an operational assessment at a particular moment. An unpaid-claim estimate is an actuarial estimate directed at a specified measure of future obligation. A booked reserve is an amount reported in financial statements under the applicable accounting requirements. The U.S. ASOP 43 unpaid-claims standard expressly distinguishes the actuarial estimation task from the amount booked and requires clarity about intended measure, valuation date, expense coverage, gross or net status, and relevant recoverables. These distinctions provide an appropriate baseline for the entire subject even where other jurisdictions have different binding rules.
Actuarial history explains the persistence of several complementary methods. Bornhuetter and Ferguson (1972) developed their IBNR reserving approach where emerging claims experience was too immature to justify simple extrapolation. Berquist and Sherman (1977) showed that reserve adequacy testing requires adjustments for operational and case reserve practices. Mack (1993) derived a chain-ladder prediction-error calculation under explicit assumptions. Their contribution remains relevant because rare severe losses and latent injury each weaken a different premise of straightforward extrapolation: representative average severity, stable claim emergence, comparable accounting practices, and a sufficiently credible sample.
A large insured loss can be reported promptly yet remain unpaid for years; a latent claim can be modest in size but unreported for decades; and a long-tail line can contain many small as well as a few enormous losses. The three descriptors refer to severity, emergence mechanism and development horizon respectively. Treating them as interchangeable leads to incorrect selection of thresholds, origin periods, development factors and scenario populations. The relevant scientific object is a joint process linking exposures, legal coverage, claim notices, defense and settlement, potential escalation, and payments.
The empirical and normative settings also differ. The Casualty Actuarial Society's latent-exposure reserving study describes exposure reserves and reporting scenarios for asbestos and environmental liabilities; it is evidence of established reserving approaches, not a transferable estimate of another insurer's loss. Lynch and Moore's (2022) social-inflation research studies U.S. industry patterns with a defined observational scope and does not establish that every liability class or jurisdiction follows the same trend. These limits are integral to scientific inference.
Methodologically, this dossier uses mathematical identities, probabilistic examples, source-linked literature analysis, professional guidance, and controlled synthetic illustrations. A hypothetical calculation is explicitly a demonstration of arithmetic under its stated assumptions. It is separated from empirical findings about the historical record. The research relates principally to Insurance Pricing and Reserving and has substantive connections to Casualty and Liability Insurance, Commercial Insurance, Claims Science, and Insurance Statistics and Predictive Modeling.
Large-Loss Reserving Thresholds
The unit of loss and the object of the estimate
The first question is what qualifies as one large loss. A single claimant's bodily injury may create a sequence of payments; one industrial site may generate many claims and policies; an occurrence may produce several injured persons; one insured can have multiple layers of coverage. A threshold applied to each claimant generates a different sample than a threshold applied to the total occurrence or the aggregate insured event. Data preparation must specify claimant, occurrence, policy, insured, coverage layer, and legal matter identifiers and must preserve the relationships among them. Without that hierarchy, duplicate recoveries and false independent observations become possible.
Let the ultimate ground-up cost of a defined occurrence be the nonnegative random variable . Let be paid claim cost through time , the case reserve for future payments on reported claims, and the eventually realized ultimate amount, measured on a consistent gross or net basis. An observed large claim can be defined by the current incurred amount exceeding a threshold , or by its ultimately realized amount . These are different events. At early valuations, changes in case reserve practice alter membership in the observed large-loss sample even when the underlying ultimate severity distribution does not change. A coherent analysis therefore stores the historical threshold test and separately models final-threshold crossing.
A reserve is also contingent on the chosen valuation object. If the objective is expected unpaid nominal cash flow, the target is for a covered occurrence where denotes the information available at the review cutoff. If the intended output is a quantile, a contractual stress amount, a discounted probability-weighted estimate, or the financial-statement liability, the target measure changes. The scope and measure requirements in ASOP 43 make this an explicit documentation choice. A high percentile is not interchangeable with an expected value, and a conditional case forecast is not automatically a complete portfolio IBNR estimate.
Monetary thresholds, inflation and changing composition
A nominal threshold such as $1 million does not represent the same level of claim severity in every accident year. Rising medical expenses, labor costs, legal costs or average compensation can move an otherwise unchanged real severity distribution across a fixed cutoff. To compare origin years at a common cost level, define an index for the cost environment of year , choose a base year , and transform a historical severity to . Equivalently, a threshold in base-year currency corresponds to the historical nominal threshold . This is a mechanical equivalence, not a claim that any broad consumer price index captures class-specific claims inflation.
Threshold selection affects three quantities simultaneously: selected claim frequency, conditional claim severity, and the amount of loss excluded below the threshold. If a portfolio uses only records already above an incurred threshold, it may miss late-developing claims that will cross it after valuation. The truncated sample cannot in general identify a full ground-up severity distribution without assumptions or independent data on the unselected population. Likewise, an insurer's large-loss count can rise from increased exposure volume, higher real severity, changed recording practices or a lower effective real cutoff. Comparing raw counts without denominators invites false trend identification.
A defensible historical extract should retain original nominal observations, the exact original threshold and date of application, the inflation index and version used, and an as-if-common-price restatement. Reporting both the nominal-criterion result and a standardized-price result is informative when inflation is itself part of the object under investigation. The EIOPA insurance-sector inflation report documents the broad influence of claims costs and financial rates; it does not supply a universal large-loss trend index for every jurisdiction.
Excess layers, policy limits, aggregates and recoveries
A layer with per-occurrence attachment and limit pays, before other contractual adjustments, , where . The priority and cap are applied to individual covered occurrences, not to the sum of all occurrence costs. In general, . Even where ground-up aggregate ultimate cost is accurately forecast, layer-specific ultimate cost may be highly uncertain because the distribution of individual severities and their dependence is unobserved.
Reinsurance introduces further distinctions: treaty versus facultative arrangement, per-risk versus per-occurrence attachment, aggregation, reinstatements, policy and treaty limits, exclusions, reporting and settlement clauses, and recoverability. A large claim may be grossly significant yet largely ceded, or moderately sized but near a decisive threshold for net retention. An actuary should identify gross ultimate cost, contractual ceded entitlement, credit and collectibility assumptions, salvage and subrogation, and net ultimate cost as separate objects. Coverage litigation or disputed aggregation can affect the ceded estimate independently of the underlying claim severity.
Bloom and Patrik (2013) discuss excess loss development pitfalls in long-tailed business. Their research identifies the volatility of excess development triangles and explores more stable ground-up experience transformed through excess loss factors, with necessary adjustments for changes in retentions. This gives a methodological reason to preserve both ground-up and layered histories rather than selecting development factors solely on the basis of sparse excess payments.
Extreme severity and sample selection
A classical heavy-tail approximation represents losses above a sufficiently high threshold by a generalized Pareto distribution. For excess , one possible parameterization has survival function for within the support and ; the exponential limit obtains as . This is a candidate model requiring diagnostics, not an empirical law. Positive corresponds to a heavy right tail; the theoretical mean excess exists for and the variance for . Estimation is especially fragile where the sample contains only a handful of independent threshold exceedances, where settlements are censored, or where policy limits cap the observable amounts.
Method choice must respect the distinction between right censoring and payment limits. If final ultimate loss is still unknown but current payments are observed, its outcome is development-censored. If a policy limit prevents observing the ground-up amount above the limit, insured payment is contractually censored or truncated relative to the unlimited loss process. A model that uses capped payments as if they were uncapped severity observations biases upper-tail estimation. A robust analysis conditions the likelihood on the observation mechanism and includes sensitivity to cap changes and claims adjusting behavior.
For a threshold exceedance model, a useful conceptual decomposition is only under a compatible frequency/severity definition and adequate independence assumptions; frequency and severity dependence, cross-claim litigation shocks and changing limits call for a more general joint expectation. A reserve prediction requires the distribution of future unpaid cost and not merely a fitted marginal severity curve. The observation mechanism is part of that prediction.
Latent Injury and Liability Development
The multiple clocks of a latent claim
A latent claim has multiple dates that serve different analytical functions: exposure commencement and cessation, injury or property damage, manifestation or diagnosis, first notice to an insured, notice to an insurer, assertion of legal demand, coverage litigation, allocation decision, payment, closure and reopening. They may occur in different calendar years and differ across policies and courts. A single accident-year label often compresses several distinct processes. Long-tail analysis should preserve the time intervals explicitly and should identify the calendar convention and legal interpretation supporting an origin-year assignment.
For occupational illness, exposure may occur over many years while the diagnosis is recent. For environmental damage, physical contamination, discovery and remediation may be widely separated. For mass-tort claims, changes in scientific evidence, claimant screening, filing practice and settlements can generate distinct waves of reported counts. For each cohort of potentially covered exposures, an actuary must model both the number of claims that might become compensable and the time and severity with which those claims enter the insurer's system. These components need not share the same statistical trend.
Haidu (1996) investigated asbestos and environmental reinsurance liabilities using exposure reserves, report-year experience and scenarios because historical emergence alone might not adequately represent changing coverage, loss size or trigger interpretations. His setting shows why an exposure-based inventory is analytically distinct from development of currently reported claims. This lesson is methodological; historical estimates from the study cannot be transferred unchanged to contemporary portfolios.
Contract triggers, liability and portfolio reconstruction
Coverage depends on policy language, policy period, limits, deductibles, exclusions, notice obligations, reinsurance terms and applicable law. Occurrence policies and claims-made policies can attach differently to identical physical events, and legal systems differ in the treatment of continuous injury, aggregation, causation, defense costs and allocation across years. A general reserving model cannot independently determine which legal theory a court will adopt. It can, however, represent plausible coverage constructions as distinct conditional scenarios and document the policies and factual records supporting each.
The prospective reserving population should be linked to a structured exposure inventory: insured operations, sites or products, historical policy years, known injury mechanisms, coverage terms, jurisdictions, exposure intensity, potential claimants, and the evidence status of each item. Missing insurance policy records and uncertain corporate succession are sources of exposure uncertainty, not simply statistical noise. An assessment should explain which legal matters have been identified and which classes of unknown claims are represented by a modeled provision.
There is a fundamental distinction between an exposure that might cause loss, a legal claim that could be filed, and an obligation that is within the contracts under valuation. A model estimating the first cannot be used as an accounting liability for the third without a supported coverage and recognition bridge. Conversely, uncertainty about coverage does not eliminate the need to evaluate possible liabilities where applicable law and valuation rules require recognition or relevant disclosure. The insured-event mapping is a prerequisite to quantification.
Reporting, settlement and survival models
Let denote occurrence or a defined injury-origin date, the insurer's report date and a future payment date. Reporting delay is observed only for claims reported by valuation time , and the remaining unreported claims form a truncated part of the occurrence population. The reporting distribution should not be estimated as though the reported sample were a complete random sample of all eventual claims. If denotes the conditional reporting cumulative distribution for exposure covariates , then an illustrative known-event cohort of eventually reportable claims has expected unreported count . Estimating requires exposure, claim-frequency or prior information and is a separate identification problem.
Settlement time and payment timing are further conditional processes. A survival model may estimate a closure hazard , with representing the history of payments, litigation milestones and current status. A competing-risks specification can distinguish paid closure, withdrawal or defense verdict; reopenings require a recurrent or multi-state representation. The individual-claim reserving model of Larsen (2007) explicitly considers individual development, business mix and settlement, while Bladt and Pittarello (2025) demonstrate multi-state individual reserving with conditional Aalen–Johansen estimation. These studies establish available methods rather than proving that one is optimal for every latent portfolio.
An adequate model recognizes informative delay: severe cases may enter the system earlier because of visible injury, or later because clinical diagnosis and causation are difficult. Litigation can increase both severity and duration. A claim's eventual total size and its time to payment may therefore be dependent. The dependence becomes financially significant when discounting, policy limits, reinsurance attachment and inflation are applied at different dates.
The reserve identity and its components
For a fixed covered cohort at valuation , denote eventual aggregate ultimate covered cost by , paid-to-date by , and unpaid cost by . If the conditional target is an expected nominal undiscounted reserve, . The total can be partitioned, with mutually exclusive definitions, into future payments on reported claims and future payments on as-yet-unreported claims; potential development above current case estimates is incorporated within the reported component. Common operational labels include RBNS (reported but not settled), IBNR (incurred but not reported) and IBNER (incurred but not enough reported), but their implementation varies. All definitions and reconciliation rules should be stated in the source register.
This partition is an accounting of the prediction, not a license to add independently estimated components without removing overlaps. A case-reserve-based method may already reflect expected development, while an aggregate IBNR estimate may include both true unreported claims and development beyond case values. Double counting is common when exposure scenarios are added to aggregate tail factors that already contain historical latent development. A documented cohort-by-cohort bridge should identify the exposure set and payment streams included by each model.
The time value of money is a further dimension. For future random payments at times , an illustrative discounted conditional estimate is , where is the discount factor required by the valuation framework. Changes in payment lag alter present value even when nominal ultimate cost is unchanged. The current ASOP 20 cash-flow guidance applies to covered U.S. work products issued on or after June 1, 2026, and addresses analysis of property/casualty cash flows including discounting. Its scope should be interpreted directly rather than substituting superseded pre-2026 descriptions.
Judicial and Social Inflation
Economic and claims-specific price change
The word inflation covers several mechanisms. General price inflation affects the purchasing power of money. Claims cost inflation changes the price of medical care, building materials, wage replacement, rehabilitation, defense services and other claim-specific inputs. Judicial and social processes can change damages, filing incentives, settlement behavior or the probability of an adverse outcome. These effects may coexist and interact with economic inflation, and different line-of-business portfolios have different sensitivities. A rate applied uniformly to all future claim payments will usually conceal this composition.
One useful decomposition writes the expected ultimate cost per compensable claim in period as , with components such as medical care, income loss, defense expenses and noneconomic damages, weights reflecting claim composition and component-specific costs . A change in can arise from shifts in the weights as well as growth in underlying costs. The weights may depend on injury severity, jurisdiction, settlement channel and coverage limits. For a large-loss subset, changes in the selection threshold also change the weights through the conditioning event. A severity trend estimated after selecting claims above a fixed nominal amount is therefore partly a trend in the sample definition.
EIOPA's (2023) study of insurance-sector inflation examines macroeconomic and insurance consequences, including interaction of claim-cost pressure and interest-rate changes. It supports treating claim-cost assumptions and liability discounting separately. It does not determine an insurer-specific compensation-inflation rate. The relevant indices, currencies and exposure characteristics must be identified within the work's own valuation data.
Legal institutions and calendar-year effects
The phrase social inflation has competing industry uses; as an empirical label, it requires a chosen measure and identifiable comparison. Lynch and Moore (2022) examined social inflation in loss development using U.S. property/casualty industry triangle data, focusing on commercial automobile liability. Their study inferred patterns consistent with excessive claims inflation and related its estimates to stated periods and lines. Their later research update through 2021 revised the time window and emphasized the anomalous development conditions during the COVID-19 period. The published aggregate estimates are specific to these methods and populations; they are not universal adjustments to all casualty reserves.
From a reserving perspective, a change in court practice can affect the litigation success probability, damage amount conditional on success, the speed of resolution, the number of filings and the propensity of already closed claims to reopen. These channels have different payment footprints. A rise in average settlement size might reflect more severe cases entering the settled sample rather than greater awards for otherwise identical cases. A rise in reported incurred development might reflect stronger case reserving rather than actual settlement inflation. Observed data alone may not distinguish those explanations.
Calendar-year influences act across multiple occurrence-year cohorts. A legal change effective in calendar year can alter payments for accidents from many earlier years, breaking the stable origin-development pattern assumed by simple chain ladder projections. Conversely, the same accident year can pass through several legal regimes before closure. Identifying the legal event, affected jurisdictions, policy periods and claim types is necessary before applying a scenario. Where legal doctrine differs materially across jurisdictions, one worldwide 'social inflation' scalar is an inappropriate representation of the insured obligation.
Professional analysis should distinguish three claims: an observable association between calendar time and development; an empirical claim that the association reflects legal behavior rather than other factors; and a forward-looking forecast that the effect persists. The first may be demonstrable with a triangle, the second often requires granular legal and cost data, and the third requires supported scenarios. A source supporting one step cannot be silently used as proof of the others.
Calibration under data and identification constraints
A practical claims-inflation analysis reconciles occurrence-year, report-year and calendar-year data in a common valuation extract. It can track closed and open claim counts, the frequency of severe settlements, paid severities at a constant price level, changes in case estimates, disposal speed, defense expense, legal venue, injury mix and policy limits. Comparing an aggregate development pattern with case-level change tables helps locate a pattern but does not by itself establish causality. A trend attributed to inflation should be challenged by changes in the claims handling organization, threshold definitions, exposure mix and portfolio transfers.
A common statistical representation is for positive incremental amounts by origin period and development period , with calendar index . The three effect families are not uniquely identified without constraints because a linear trend can be redistributed among them. A fitted calendar coefficient should not be reported as a purely observed causal inflation effect. The chosen identifiability constraints, external index information and time-period segmentation must appear in the technical appendix or recorded assumptions.
Where the largest claims are sparse, case-level covariates can improve conditional predictions if available without leakage. Relevant covariates may include injury category, jurisdiction, age of the claim, lawsuit stage, coverage layer, defense organization, and inflation-sensitive payment type. Data on settlements should include the population of unsettled and dismissed claims to avoid selective inference from paid cases only. When a model's estimates depend heavily on one extraordinary observation, report influence diagnostics and selected alternatives rather than smoothing away the observation as inconvenient.
Inflation, interest rates and contract boundaries
Nominal ultimate loss, payment timing and present-value liability are different measurements. Suppose a base payment of is anticipated after years, grows at an assumed annual claims-cost rate , and is discounted at a flat illustrative rate . Its projected nominal amount is , and present value is . This expression makes the offsetting directions clear, but it is not a substitute for an actual discount curve, variable payment dates, stochastic inflation, or the requirements of a particular reporting standard.
Coverage and reinsurance limits may be fixed in nominal currency while the underlying loss grows. In such a case, gross claims inflation and net retained inflation are nonlinear: a growing loss can move through an attachment point, exhaust a layer and then shift further increases to a different carrier. A model that first inflates aggregate loss and then applies a single portfolio cap ignores the contract-level distribution of losses. Inflation must be applied to the relevant ground-up claim payments and expenses before policy and treaty treatment, subject to contract provisions and the actual order of operations.
An apparent financial benefit from rising discount rates does not establish reduced underlying insurance cost. Claims-cost inflation may increase the nominal estimate while present-value rates rise; the net effect depends on duration, cash-flow shape and rate term structure. IFRS 17's insurance-contract measurement requirements and Solvency II Article 77 technical provisions define distinct measurement regimes. An actuarial reserve review should state which regime it serves before combining claims inflation, uncertainty adjustments, margin and discounting.
Sparse Claim Experience
Development methods and their assumptions
A claims triangle is an informative but compressed view of recorded experience. Let be cumulative paid or reported loss for origin period at development age . The classical chain ladder estimates age-to-age factors, for example over rows observed at both ages, and derives an ultimate . This technique is sensitive to stable development patterns and adequate credibility. A single multimillion-dollar settlement can dominate a factor, especially in a smaller line or high excess layer. Estimated factors drawn from nonhomogeneous cohorts can have precise-looking arithmetic and poor economic relevance.
Mack (1993) supplied a distribution-free chain-ladder error analysis under assumptions about conditional means, conditional variances and independence across origin years. A formula for model-based prediction error should be reported together with these assumptions. It cannot capture unmodeled litigation regimes, missing exposure populations or future coverage reinterpretation simply by being stochastic. Historical chain-ladder uncertainty is one component of the total risk in an evolving liability system.
Bornhuetter and Ferguson (1972) introduced an expected-loss reserving basis that is useful when immature data make direct extrapolation unstable. Let be a prior expected ultimate amount and the expected fraction reported by the valuation date on a consistent basis. Then one common reported-incurred form is , with . It combines observed reported cost with an a priori expectation of unreported development. The prior must be justified from exposures, prices, expected loss ratios or external information; it is not a free adjustment chosen to reconcile a desired final reserve.
Mack (2006) examined Bornhuetter–Ferguson parameter selection, including distinctions between additive and multiplicative development assumptions. Ollodart (1999) analyzed paid BF dependence on case estimates, emphasizing that combinations of paid and incurred approaches require attention to shared underlying information. Averaging several methods that derive from essentially the same triangles does not provide independent confirmation.
Berquist and Sherman's (1977) reserve-adequacy methodology is particularly instructive when settlement rates or case reserving practices change. Stronger case reserves accelerate reported incurred emergence without accelerating actual payments. Faster settlements can increase paid development without increasing ultimate losses. A method should be normalized, or its assumptions revised, when such operating changes materially alter the observable pattern.
Individual-claim and exposure-based estimates
For known large claims, an individual-claim estimate can use the current case file, legal assessment, expected payment schedule, defense costs, coverage limits and plausible settlement alternatives. The individual prediction remains probabilistic: even a carefully documented case reserve may be undercut by a court ruling, later treatment costs or appeal. Claims-level models can be reconciled against aggregate expected development; disagreement is information about selection, case practices or unreported losses rather than proof of one model's error.
The individual reserving literature of Larsen (2007) illustrates how reporting, business mix and settlement characteristics can be introduced at a granular level. Bladt and Pittarello (2025) evaluate a conditional multistate reserving estimator and report performance on simulated and real data. These are methodological resources and do not provide a general empirical performance ranking for latent claims, where the unreported population and coverage history can dominate observed claim-level information.
For unreported latent exposures, consider an inventory of exposure groups . An illustrative expected unpaid cost may be written , where is a potentially exposed population or appropriate intensity measure, an eventual covered-claim probability or frequency factor, and an expected future cost on the correct coverage and payment basis. If a single event can involve several policies, the model should allocate the event to policy and treaty layers before summing by insurer. The factorization is an explanatory device; correlations, uncertain exposure inventories and legal eligibility can require a joint model.
Haidu's (1996) exposure-reserving discussion is evidence for supplementing triangles with external exposure and future-reporting scenarios where historical losses are a poor representation. Scenario inputs should have independently documented sources such as policy records, claims studies, medical or environmental evidence, relevant legal developments and insured business histories. An unsupported expected claim count multiplied by an unsupported severity assumption remains speculative even when the multiplication is correct.
Worked example: stable thresholds and excess layers
The following amounts are entirely hypothetical, denominated in millions of currency units, and use deliberately simple distributions to make the arithmetic auditable. Suppose the claim-cost index is in historical year and in base year . A base-year large-loss criterion of corresponds to a historical cutoff (rounded). A historical claim with nominal ultimate restates to $0.800(150/100)=1.200$ in base-year units and therefore qualifies on a constant-price basis. A fixed nominal historical threshold of would have excluded it. This demonstrates a classification effect; it does not identify the appropriate actual index for any insurance class.
Now assume a per-occurrence excess cover attaching at with a limit, and three independent illustrative ultimate ground-up claims of , , and . The respective layer costs are , , and , totaling . The aggregate ground-up total is , but applying the layer function once to would produce only . The difference arises solely from incorrect aggregation before applying the contractual layer. A genuine contract with aggregate conditions could produce another result, which must be calculated under its own language.
To show a related expectation error, take one hypothetical occurrence with ultimate at probability , at probability , and at probability . Its expected ground-up cost is . Expected layer cost is $0.60(0)+0.30(0.500)+0.10(2.000)=0.350L(E[X])=0.2500.100$ below the actual expected layer cost. Nonlinearity explains the discrepancy: . Retentions, reinsurance recoveries and limits should therefore be applied at the scenario or simulated claim level before calculating an expectation.
These calculations establish three distinct effects: threshold classification changes with monetary scale; sums of layer costs depend on the grouping of occurrences; and expected layer cost depends on the severity distribution rather than the mean alone. Each is relevant even when claim counts are otherwise perfectly known. No assumption about future claims frequency was needed for these identities.
Worked example: long-tail ultimate and prediction
Consider one synthetic origin-year cohort. At valuation, cumulative paid cost is , reported incurred cost is , current case outstanding is therefore , and the prior expected ultimate is . Suppose the selected proportion of ultimate reported on the same incurred basis is . The BF ultimate is $5.000+8.000(1-0.55)=8.600, and the expected unpaid amount is \8.600-2.400=6.2002.6003.600$. If the entire difference is labeled IBNR operationally, the documentation must make clear that it can include both unreported claims and development beyond recorded case estimates.
A chain-ladder estimate based on the same gives $5.000/0.55=9.091$ and unpaid $9.091-2.400=6.691$ after rounding. The difference from the BF reserve results from the estimators' different treatment of immature emergence, not from calculation error. Neither is inherently correct without evidence about the prior expected ultimate and the development proportion.
A focused sensitivity replaces the assumed reporting proportion by and , holding all other synthetic observations constant. The resulting BF unpaid reserves are $5.000+8.000(0.55)-2.400=7.000$ and $5.000+8.000(0.35)-2.400=5.400$ respectively. The corresponding chain-ladder unpaid estimates are approximately and . The widening spread at low illustrates why long-tail maturity assumptions matter more for a multiplicative estimate when few claims have emerged. These are assumption sensitivities, not confidence intervals.
A further development-tail adjustment must be defined in relation to the proportion already used. Adding a separate tail factor to a BF reserve that was calibrated using ultimate maturity can double count future development. Similarly, adding a standalone latent-exposure reserve to a triangle estimate is defensible only for exposure populations demonstrably absent from that triangle's modeled ultimate. The data dictionary should identify origin periods, report-year effects, the terminal development age and the assumed unobserved tail beyond the available triangle.
Scenario and Case-Based Estimates
Tail risk, latent exposure and reinsurance
An actuarial scenario is a specified internally coherent description of possible future claims and payment behavior, together with assumptions about the probability or weight of that description. Large-loss scenarios can vary ultimate severity, claimant frequency, bodily injury progression, time to report, procedural law, judgments, defense costs, and insurance coverage. A scenario with a plausible legal outcome but impossible payment dates is not coherent; neither is a scenario that changes a coverage limit without identifying the contract allowing that change. The model must respect both the claim process and the financial instruments that transfer or retain its cost.
A latent-loss scenario starts with an exposure population. For example, it may distinguish previously identified sites from plausible but unidentified sites, current claimants from possible future claimants, and reported matters from exposure inventories that have no current notices. The probability of discovering an additional exposure is distinct from the probability of a compensable claim conditional on discovery. An inventory derived from litigation records can be selected toward disputed, high-severity cases. Claims frequency, defense expense and coverage eligibility should therefore be calibrated or bounded separately wherever the supporting information permits.
Claims-made versus occurrence policy characteristics also affect the development horizon. Notice, retroactive dates and extended reporting periods may matter for some claims-made covers. Long-tail occurrence contracts may remain relevant to claims reported many years after exposure. The precise contract language and governing law determine whether a given future payment is part of the insurer's obligation. In a synthetic analysis, it is acceptable to state an assumed coverage rule explicitly; it is not acceptable to present that rule as universally applicable across insurance jurisdictions.
A reinsurer's projection requires the same loss to pass through policy conditions and treaty conditions in the correct order. Large claims can reach the attachment of a facultative layer, erode an aggregate treaty retention, trigger reinstatement provisions or create disputes over event definitions. Haidu (1996) documented the special difficulty of reinsurance estimates for latent losses, including dependence on information from cedents. A ceded reserve should distinguish the gross economic claim, contractually recoverable amount, uncertainty over the interpretation of coverage and the risk that an established recovery is not collected.
For a portfolio of claims and a deterministic treaty transformation under fixed contracts, the net cost is , which may not equal a sum of separately modeled net claims if treaty aggregation is portfolio-wide. The treaty transform itself may be random under disputed coverage or counterparty outcomes. This is a reason to compute ceded and net results from common underlying scenarios rather than selecting independent gross and net reserve estimates that cannot reconcile at the scenario level.
Worked example: a finite scenario portfolio
Consider a fully hypothetical set of three exhaustive future outcomes for the entire remaining gross covered liability of a mature industrial casualty portfolio, expressed in millions. Assign scenario A probability and unpaid gross cost , scenario B probability and cost , and scenario C probability and cost . The expected nominal unpaid cost is $0.60(4)+0.30(9)+0.10(22)=7.300. The second raw moment is \0.60(16)+0.30(81)+0.10(484)=82.300, so variance is \82.300-7.300^2=29.010$ million squared and standard deviation is approximately million. The distribution has a defined right tail even though its mean is well below the severe scenario.
With quantiles defined as the smallest value at which cumulative probability reaches the stated level, the 90th percentile is because the cumulative probability at B equals ; the 95th percentile is . The 95th-percentile value is more than three times the expected value. Nothing about these numbers establishes the appropriate carried reserve, confidence level or capital amount; those depend on the chosen measurement purpose and regulatory framework. Their purpose is to demonstrate that a central estimate and a high quantile answer different questions.
Now suppose the same fully specified hypothetical contract structure generates ceded entitlement of in A, in B and in C. The scenario-level net amounts are , and . Expected ceded entitlement is $0.60(0)+0.30(2)+0.10(12)=1.800, and expected net is \7.300-1.800=5.5001.620$ and expected cash retained is . The distinction is between a contractual entitlement and its realization, and the illustration does not establish how any jurisdiction requires a particular credit adjustment to be recorded.
Scenario probabilities are epistemic inputs unless anchored in credible frequency evidence. A risk committee might select the 10% severe scenario to investigate vulnerability, but calling the scenario weight '10%' does not make it a data-estimated occurrence rate. A complete record should identify whether weights were statistically fitted, elicited from specialists, inferred from analogous exposures, or assigned for illustration. If probabilities are unsupported but ranges can be defended, it is more accurate to present conditional outcomes and bounded weights than a misleading single probability-weighted mean.
Model disagreement, parameter uncertainty and diversification
A long-tail predictive distribution should reflect several sources of uncertainty: random claim and payment outcomes given fixed parameters; uncertainty in rates, severity scales, tail indices and reporting patterns; changes in claims handling or portfolio mix; and uncertainty in legal, medical or regulatory regimes. The second and third sources may be material even where routine model output produces narrow intervals. A bootstrap conditional on the historical triangle cannot create outcomes representing a missing class of latent claims unless those outcomes enter the model's assumed population.
Let index a set of supported model specifications. By the law of total variance, if the probabilities over are themselves justified. The first term averages uncertainty conditional on a model, and the second measures variation in model-specific means. This identity does not solve model selection or assign objective probabilities to legal scenarios. Its utility is diagnostic: treating a selected method as if all other plausible explanations had been ruled out can materially understate uncertainty.
Cross-claim dependence is particularly important for latent injury and mass litigation. Multiple claims may share a manufacturer, medical discovery, legal venue, defense strategy or claims administrator. A court ruling or scientific study can simultaneously change many estimated settlements. If , then . Independent-claim simulation implicitly sets the covariance terms to zero; that assumption is often a substantive judgment rather than a fact. The absence of a detailed dependence model should prompt an explicit common-shock sensitivity rather than an automatic zero-covariance assumption.
Reinsurance can concentrate uncertainty instead of removing it. Net tail risk may be sensitive to attachment and exhaustion points, recoverable collectibility, event aggregation, or one large disputed ceded claim. A portfolio with apparently stable expected gross loss may have an unstable net position when the distribution lies close to retention thresholds. Comparing gross and net distributions scenario by scenario reveals such nonlinear effects.
Independent evidence and competing explanations
Competing models should be selected to challenge distinct assumptions. A paid chain ladder tests the persistence of payment development. An incurred method tests case-reserve adequacy and reporting evolution. Bornhuetter–Ferguson introduces an independent prior expected ultimate, to the extent the prior is independently established. An exposure model tests whether claims not represented in the triangle could emerge. An individual-claim model tests conditional predictions on observed case characteristics. A legal scenario tests the conditional effect of a contract or tort outcome. Their agreement matters most when they draw on substantially independent information rather than multiple transformations of the same historical triangle.
Method disagreement should be traced to data and assumptions before a selected value is declared. For example, a much higher paid estimate than incurred estimate may reflect delayed case strengthening, acceleration in settlement, or one exceptional payment. An exposure model above triangle estimates may indicate genuine unreported cohorts or overlapping already-reported obligations. A higher excess-layer estimate than ground-up scaling may follow from retention erosion and a heavier severity tail. Each explanation has different observable implications and suggests a specific check.
The governing question is whether an alternative changes the conditional economic obligation and whether the record contains evidence capable of distinguishing it. Sensitivity analysis can produce bounded amounts without assigning false precision. A report may conclude that a specified legal scenario dominates reserve variability while the historical severity parameters remain reasonably estimated. That is a scientifically meaningful result even when the precise legal probability remains unresolved.
Reserve Sensitivity and QA
Audit-ready data and reproducible valuation snapshots
A source-locked long-tail estimate needs a valuation snapshot that another reviewer could reconstruct. The essential fields include claim and occurrence identifiers, policy and treaty identifiers, exposure years, report dates, claim and payment dates, status histories, historical case reserves, payment currencies, development periods, limits and retentions, salvage and subrogation, allocated and unallocated expenses, recoveries, legal venues, major claim milestones, and the provenance of each data extract. A data dictionary specifies the exact meanings of incurred, outstanding and IBNR amounts. Corrections and restatements require traceable versions, and historical snapshots should not be silently replaced by subsequent information.
ASOP 23's data-quality guidance addresses selecting, reviewing and relying on data and communicating known significant data limitations. It distinguishes an actuarial data review from a formal audit. A reserving analysis should record duplicate-event checks, inconsistent paid-to-incurred relationships, missing origin dates, reopened claims, exchange-rate conversions, recoveries posted before corresponding gross payments, and changes in coding practices. Exceptions should be resolved or bounded based on their effect on the reserve target rather than being treated as a single generic completeness score.
The review date is distinct from the valuation date. A legal decision learned after the accounting date but before the review cutoff may change the information available to an actuary, subject to the measurement framework's treatment of subsequent events. The estimate should therefore declare the accounting date, transaction valuation date and information review cutoff. ASOP 43's required valuation disclosures provide a clear example of this discipline. Without the time boundary, historical backtests are vulnerable to using facts that were unavailable when the estimate purportedly was made.
A reproducibility manifest for every scenario and method should capture the underlying input snapshot hash, source inventory, program or formula version, inflation indices, policy mapping, legal scenario record, assumptions selected, simulation seed when applicable, output hash, reviewers and decision date. This is an operational implementation of auditable scholarship, not a proprietary theory of reserving. Sensitive claim-level data should be protected or anonymized in accordance with applicable privacy and access requirements; publishing raw personal injury files is unnecessary to reproduce the arithmetic of a synthetic example.
Backtesting, calendar effects and stress analysis
Historical validation recreates a series of genuine past valuation dates using only information actually available at each date. It compares the estimate with subsequent observed development, keeping the outcome horizon explicit; a ten-year validation cannot prove the correctness of a reserve for claims that mature over thirty years. Rolling-origin backtests can reveal persistent bias, subgroup deterioration and tail instability, but the scarcity of fully matured recent latent cohorts limits how far those results can be extrapolated.
A useful backtest table reports error by line, original exposure or origin cohort, development maturity, coverage layer, jurisdiction and severity band. For point forecasts, it may use absolute or scaled prediction error, but ratio errors on portfolios with tiny denominators can be unstable. For distributional forecasts, coverage of prediction intervals and proper scoring methods can be considered. Holdout observations should be grouped to avoid a large occurrence appearing partly in training and partly in test data, which would leak related information into a forecast purportedly made before that occurrence was reported.
Stress analysis complements out-of-time validation. Plausible changes include slower reporting, prolonged litigation, higher claim severity, unexpectedly high defense expense, changes in settlement rate, reopening of closed claims, a longer development tail, reinterpretation of policy allocation, common-shock mass litigation and a ceded recoverability reduction. Each shock should be mapped to a distinct model parameter or scenario and tracked from ground-up cost through policy terms to net cash flows. Overlapping scenarios should not be added together as though they were independent incremental liabilities.
A short, transparent inflation stress illustrates duration dependence. For an entirely hypothetical $5.000 million base cost due ten years later, a 3% annual claims inflation assumption gives nominal payment $5(1.03)^{10}\approx6.720$ million; 6% gives $5(1.06)^{10}\approx8.954$ million. At an independently assumed constant 4% annual discount rate, corresponding present values are $5[(1.03)/(1.04)]^{10}\approx4.540$ million and $5[(1.06)/(1.04)]^{10}\approx6.049$ million. The nominal difference is approximately million; discounting changes its present value but does not make it disappear. Actual financial reports use the discount and cash-flow treatment prescribed by the relevant measurement standard.
Regulatory and accounting interpretation
A forecast of unpaid claims serves several purposes, but the applicable measurement rules differ. Under ASOP 43, the actuary identifies the intended measure and whether the estimate is gross or net of recoverables, with documented expenses and sensitivities. For U.S. actuarial services with covered prospective cost estimates, ASOP 53 on prospective costs has a different scope from unpaid incurred-claim estimation. The revised ASOP 20 cash-flow standard took effect for its applicable issued work products on June 1, 2026. All professional standards must be applied in accordance with their stated jurisdictional and professional applicability, not as one global rulebook.
For entities reporting under IFRS 17, the standard's current measurement framework identifies fulfillment cash flows and the measurement of insurance groups. The IFRS Foundation's key-term explanations distinguish the liability for incurred claims, fulfillment cash flows, risk adjustment for nonfinancial risk and contractual service margin. A nominal actuarial undiscounted central estimate is an input to some accounting analysis, not automatically the entire reported IFRS 17 liability. Relevant contract boundaries, recognition criteria, risk adjustments, discounting and aggregation rules require separate treatment.
For Solvency II, Article 77 of the Directive specifies technical provisions comprising a best estimate and a risk margin, with the best estimate representing probability-weighted present value of future cash flows under the applicable prescribed rates. The valuation framework also distinguishes gross provisions from amounts recoverable under reinsurance and special-purpose vehicles. That is a different target from a simple nominal case reserve, from a statistical high percentile and from an IFRS 17 accounting presentation. Differences should be reconciled by the measurement basis instead of being attributed automatically to actuarial inconsistency.
Cross-border portfolios must state currency translation assumptions, claims settlement jurisdiction, coverage law, reinsurance jurisdiction where relevant and the financial reporting regime of the entity being analyzed. Accounting recognition does not establish legal coverage, and legal coverage does not alone determine which actuarial probability measure belongs in a financial statement. An international research monograph can give comparable mathematical structures while maintaining explicit local contractual and regulatory boundaries.
A complete analytical synthesis and decision record
The investigated sources support an integrated analytical workflow. First, establish an exposure and contract universe: define occurrences, potential future claimants, policy periods, limits, reinsurance terms and jurisdictions. Second, establish the observation process: distinguish exposure, injury, report, case estimate, litigation and payment clocks. Third, construct a matched claim and cash-flow history with stable identifiers and data-quality disclosures. Fourth, choose and compare development, frequency–severity, individual-claim and scenario methods according to their actual assumptions and data populations. Fifth, apply contractual transformations at the correct event or portfolio level and reconcile gross, ceded and net amounts. Sixth, separate nominal central forecasts from discounted, risk-adjusted or percentile-based measures required for a given reporting purpose. Seventh, conduct sensitivity, historical backtesting, source-to-claim checks and independent arithmetic review.
This is a synthesis of existing professional and scholarly practices, not a newly named reserving model. Its distinctive analytical point is the requirement to maintain an explicit mapping across four different objects: the exposure population, the observable claim population, the contractually covered claim population, and the financial measurement population. The mapping explains why a model can be internally well calibrated on reported claims while materially underestimating a latent exposure portfolio, and why a legally plausible aggregate loss estimate can still be wrong for an excess layer or reporting regime. Each mapping step can be documented and challenged with different evidence.
Several recurrent errors emerge from the comparison. A constant nominal large-loss threshold creates a changing real selection mechanism. A reported-claims database is incomplete by construction for latent unreported cohorts. Claims inflation can shift both severity and time to settlement. Aggregate development factors may conceal operational case strengthening. An insurer's direct net recovery depends on actual treaty terms and collectibility rather than gross loss alone. Statistical uncertainty conditional on one fitted distribution does not encompass omitted legal and exposure scenarios. Present-value measurements require treatment of cash-flow timing, while a pure unpaid-claims estimate can have an undiscounted target. These are separate errors with separate checks.
The appropriate conclusion of a reserving study may be a conditional range rather than a single scenario-weighted figure where probabilities cannot be established. An estimate's scientific quality is determined by the match between its scope, evidence, assumptions and intended measure, and by whether material alternatives were examined. For rare, severe and latent insurance losses, documentation of uncertainty is part of the estimate's content. The source-based nature of this study supports analytical interpretation and transparent worked mathematics; it does not provide original empirical access to confidential claim files, newly validated reserve predictions for a live portfolio, or proof of causal relationships beyond those studied in the cited literature.
Conclusion
Large-loss, latent-claim and long-tail reserving requires a coherent representation of how insured exposures become observable claims and, eventually, contractually payable cash flows. Severity thresholds, observation dates, legal triggers, claim handling and settlement patterns alter what the historical data mean. The resulting reserve is conditional on a precisely identified exposure population, claims information cutoff, contractual scope, measurement basis and set of assumptions.
The mathematics establishes several durable conclusions. A common-price threshold can change which claims belong to the large-loss population. Excess-layer expectations are nonlinear and must be computed on the appropriate event distribution. Bornhuetter–Ferguson and chain ladder estimates can diverge sharply in immature cohorts because they encode different assumptions about unobserved development. Latent exposure scenarios can supplement aggregate methods only after overlapping claim populations have been reconciled. The finite scenario example demonstrates that expected unpaid cost, high quantiles, ceded entitlement and net retained cost represent separate financial objects. Inflation and discounting interact through the future cash-flow horizon, with the direction and magnitude depending on the actual valuation standard.
The specialized literature supports combining data-driven development with individual case evidence, exposure inventories and legal scenarios where conventional triangles lack credibility. Professional standards require precise communication of the intended measure, data limitations, methods, assumptions and material sensitivities. International accounting and solvency regimes assign different meanings to the final measured liability. These differences are substantive and can be explained without obscuring the underlying cash-flow analysis.
The resulting research finding is a requirement for traceable consistency: the same insured event must be followed from exposure and coverage through claim emergence, expected payment, recovery and financial valuation, with each transformation recorded. Reserving confidence comes from identifying where the evidence is strong, which uncertainties remain material and how plausible alternatives affect the conclusion. This monograph contributes an original integrated analysis and fully specified synthetic calculations using established theory and verified source descriptions; it introduces no new theoretical system and makes no claim of external peer review.
References
Actuarial Standards Board. (2011 revision). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/propertycasualty-unpaid-claim-estimates/
Actuarial Standards Board. (2016 revision; effective April 30, 2017). Actuarial Standard of Practice No. 23: Data Quality. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017 revision; effective August 1, 2018). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Actuarial Standards Board. (2025 revision; effective June 1, 2026). Actuarial Standard of Practice No. 20: Analysis of Property/Casualty Cash Flows, Including Discounting. Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/asop-no-20-analysis-of-property-casualty-cash-flows-including-discounting/
Berquist, J. R., & Sherman, R. E. (1977). Loss reserve adequacy testing: A comprehensive systematic approach. Proceedings of the Casualty Actuarial Society, 64, 123–184. Casualty Actuarial Society. https://www.casact.org/abstract/loss-reserve-adequacy-testing-comprehensive-systematic-approach
Bladt, M., & Pittarello, G. (2025). Individual claims reserving using the Aalen–Johansen estimator. ASTIN Bulletin, 55(1), 29–49. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/individual-claims-reserving-using-the-aalenjohansen-estimator/38EE623E8A2056A238CFDE0D385A8792
Bloom, D., & Patrik, G. S. (2013). A methodology for avoiding the pitfalls of excess loss development. Casualty Actuarial Society E-Forum, Fall, 1–73. Casualty Actuarial Society. https://www.casact.org/abstract/methodology-avoiding-pitfalls-excess-loss-development
Bornhuetter, R. L., & Ferguson, R. E. (1972). The actuary and IBNR. Proceedings of the Casualty Actuarial Society, 59, 181–195. Casualty Actuarial Society. https://www.casact.org/abstract/actuary-and-ibnr
European Insurance and Occupational Pensions Authority. (2023, October 5). Report on the Impact of Inflation on the Insurance Sector. EIOPA. https://www.eiopa.europa.eu/publications/report-impact-inflation-insurance-sector_en
European Insurance and Occupational Pensions Authority. (n.d.). Solvency II Single Rulebook: Article 77, Calculation of Technical Provisions. EIOPA. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2160_en
Haidu, J. W. (1996). Estimating a reinsurer's liabilities for asbestos and environmental losses. Casualty Loss Reserve Seminar Transcripts. Casualty Actuarial Society. https://www.casact.org/abstract/estimating-reinsurers-liabilities-asbestos-and-environmental-losses
IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts. IFRS Foundation. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts—Key Terms. IFRS Foundation. https://www.ifrs.org/supporting-implementation/supporting-materials-by-ifrs-standards/ifrs-17/key-terms/
Larsen, C. R. (2007). An individual claims reserving model. ASTIN Bulletin, 37(1), 113–132. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/an-individual-claims-reserving-model/A218719775893E2DC2EC7AA4772143E7
Lynch, J., & Moore, D. (2022). Social Inflation and Loss Development. Casualty Actuarial Society and Insurance Information Institute. https://www.casact.org/abstract/social-inflation-and-loss-development
Lynch, J., & Moore, D. (2023). Social Inflation and Loss Development—An Update. Casualty Actuarial Society and Insurance Information Institute. https://www.iii.org/white-paper/social-inflation-and-loss-development-an-update-030623
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://www.casact.org/abstract/distribution-free-calculation-standard-error-chain-ladder-reserve-estimates
Mack, T. (2006). Parameter estimation for Bornhuetter/Ferguson. Casualty Actuarial Society E-Forum, Fall, 141–157. Casualty Actuarial Society. https://www.casact.org/abstract/parameter-estimation-bornhuetterferguson
Ollodart, B. E. (1999). A note on the paid Bornhuetter-Ferguson loss reserving method: Recognizing dependency on case reserves. Casualty Actuarial Society Forum, Winter, 405–416. https://www.casact.org/abstract/note-paid-bornhuetter-ferguson-loss-reserving-method-recognizing-dependency-case-reserves