RESEARCH DOSSIER
Marine and Aviation Claims Reserve Uncertainty
Abstract. Marine and aviation unpaid-claim liabilities are particularly sensitive to the temporal separation between a casualty, recognition of a potentially insured occurrence, notification, coverage evaluation, liability determination, physical repair or remediation, negotiated settlement, actual payment, and reinsurance recovery. This monograph investigates how these distinct clocks shape paid and incurred claims triangles and why a development pattern estimated from ordinary transport losses can substantially underrepresent delayed, disputed, litigated, catastrophic, or cross-border claims. The analysis differentiates marine hull, cargo, protection and indemnity (P&I), aviation hull, international carriage liability, aviation products liability, and related reinsurance, while retaining explicit distinctions among policyholder, risk carrier, transportation operator, claim event, legal obligation, insured interest, deductible, and accounting unit. It combines primary statutory and regulatory materials with professional actuarial standards and peer-reviewed reserving literature. It evaluates chain ladder, Bornhuetter–Ferguson, stochastic generalized linear models, individual-claim scenario analysis, tail factors, and uncertainty decompositions. Two reproducible synthetic illustrations show how ordinary triangle extrapolation and contingent liability adjudication generate materially different views of outstanding losses. The principal finding is that the economic and legal settlement mechanism is part of the reserving data-generating process: reserve uncertainty can only be interpreted reliably after claim eligibility, reporting practice, case-estimate policy, claim aggregation, and gross-to-net transfer arrangements are made explicit. The study presents an auditable analytical synthesis and a research agenda; it proposes no new doctrine, named theory, or empirical estimate of the global market.
Keywords: Marine insurance; aviation insurance; unpaid claims; claims reserving; IBNR; IBNER; loss development triangles; claims settlement delay; casualty liability; protection and indemnity; hull claims; cargo claims; aviation liability; model uncertainty; stochastic reserving; reinsurance recoverables; IFRS 17; Solvency II; tail risk.
Introduction
A reserve for an incurred marine or aviation claim is an estimate of the insurer's remaining obligation at a measurement date. The payment can occur long after the accident and after multiple revisions to the estimated liability. Distinct legal parties may also bear different proportions of a single economic loss, depending on the subject insured, insured peril, contractual limits, defense arrangements, and reinsurance. These features make reserve uncertainty a problem of evidence and measurement as well as actuarial inference.
The scope includes liabilities for events already incurred at the valuation date, whether notified or unnotified, and the claim adjustment expenses properly attributable to the valuation. The professional starting point is ASOP No. 43 on unpaid claims, which distinguishes gross estimates, net estimates, and recoverables. The analysis remains international: the United Kingdom's statutory rules and the Montreal Convention's international air-carriage regime are applied only where their respective jurisdictional and contractual conditions are satisfied. Reporting and solvency prescriptions differ by jurisdiction and by the financial statement under consideration.
Three questions govern the investigation. First, what information do paid and incurred triangles actually measure when the observation system is disrupted by late notifications, coverage disagreements, salvage, or large losses? Second, how do laws, claims administration, and settlement incentives affect the transition from a casualty to a recognized and paid insurance obligation? Third, which probabilistic methods remain credible when portfolio history is short, heterogeneous, subject to changing exposure, and thin in the most consequential tail?
The research uses documentary triangulation rather than treating a casualty report, an insurer's financial statement, and an actuarial model as interchangeable evidence. The NTSB's accident docket system identifies investigative evidence; the IMO casualty investigation framework addresses marine safety investigation. Neither record independently establishes policy attachment, ultimate indemnity, claims-handling expenses, or recoverable reinsurance. Those require additional contractual and financial evidence.
The following analysis integrates the source-level concepts of legal coverage, operational loss adjustment, and reserve probability. Synthetic calculations are explicitly labeled and are not represented as observations of any insurer, airline, shipping company, P&I club, accident, or claimant. The journal literature is interpreted within the portions actually verified through journal abstract and bibliographic records; no unpublished portfolio or proprietary claim microdata is assumed.
Conceptual and Historical Foundations
The claim, the accident, and the obligation
A casualty is an occurrence involving a vessel, cargo, aircraft, passenger, or other legally relevant interest. A claim is a demand, notification, or recorded request for insurer response, which may contain multiple coverage heads and may ultimately be accepted, reduced, defended, denied, or withdrawn. An insured loss is an obligation within the operative policy and legal regime. A reserve is a dated estimate of the outstanding amounts arising from those obligations. The relationships are many-to-many: a single casualty may generate numerous claims under different policies, and one policy limit may aggregate several demands.
In marine business, the economic interest may be vessel hull and machinery, cargo, freight, general average, collision liability, or a shipowner's third-party exposure. Aviation business spans physical hull damage, passenger and baggage liability, cargo, airport and ground operations, product-related liability, and war or allied perils where specifically covered. The timing of casualty recognition differs from the timing of legal proof. The distinction is essential because a development triangle measures recorded insurance amounts conditional on an administrative classification. The scientific object cannot be inferred from the label “transport insurance” alone.
The applicable law and wording must be identified before a reserve scenario is mapped into insured liability. Section 55 of the UK Marine Insurance Act 1906 ties liability, subject to policy terms and other law, to a loss proximately caused by an insured peril and identifies excluded causes. The exclusion for a loss proximately caused by delay in section 55(2)(b), subject to contrary wording, illustrates why a delayed commercial delivery, vessel casualty, and insured physical damage are separate questions. A higher freight bill or longer voyage is an exposure indicator, not automatically an indemnifiable claim.
Marine cargo contracts are particularly wording-dependent. The IUA's Institute Cargo Clauses (A), 2009 identify a concrete market form, but an investigator must obtain the exact terms incorporated into the policy, endorsements, governing law, insured transit, value, exclusion language, deductible, and claims provisions before evaluating coverage. The presence of a model clause in an industry catalogue does not prove that any given shipment was protected by it.
International aviation liability illustrates an additional difficulty. The Montreal Convention 1999 framework addresses relevant international carriage of passengers, baggage, and cargo; it is not a universal code for aircraft hull or product liability. The convention's liability-limit revision effective December 28, 2024, demonstrates that a reserve must apply the correct legal regime and temporal limit rather than applying a historical limit mechanically to all unsettled claims. Relevant state participation, time of carriage, relationship between carrier and claimant, applicable court, and any other available cause of action remain case-specific questions.
Why long tails occur
The claim tail has at least four dimensions. The reporting tail spans accident to notification. The evidentiary tail covers investigation and proof. The adjudicative tail spans allocation of fault, coverage, quantum, settlement discussions, and litigation. The payment tail runs from established or expected obligation to transfer of funds, potentially through multiple interim payments. Reinsurance adds a separate recovery tail that need not end with the insurer's payment to its insured.
Marine hull repair may develop in discrete jumps when a drydock inspection reveals hidden damage, specialist parts are unavailable, or an agreed replacement valuation changes. Cargo loss may require a survey, proof of damaged goods and salvage values, warehouse and storage decisions, cargo-owner documentation, and carrier or bailee recovery. P&I exposures may depend on third-party liabilities for pollution, cargo, crew, collision, or other risks within the club rules. An aviation claim can turn on airworthiness findings, passenger injury progression, international carriage law, products causation, or court proceedings. These are alternative mechanisms, not a single generic explanation called “slow settlement.”
For very serious marine casualties, the IMO's Casualty Investigation Code provides a mechanism for collecting and analyzing safety facts. Such evidence may affect litigation and actuarial assumptions, yet a safety report's purpose and legal effect differ from an insurer's claim adjudication. Similarly, EASA's Annual Safety Review 2026 describes 2025 aviation safety experience, allowing qualitative exposure context, while containing no standardized transaction-level insurance reserve record. The evidentiary bridge from event frequency to insurance ultimate requires policy scope and claim severity information.
Financial valuation and the boundaries of the reserve
An actuarial estimate of ultimate paid claims is an input into several reporting regimes, each with its own measurement prescription. IFRS 17 Insurance Contracts measures groups of insurance contracts using current estimates of fulfillment cash flows, timing, and a risk adjustment for non-financial risk, alongside other relevant contractual elements. The liability for incurred claims must be distinguished from the liability for remaining coverage; insurance contract presentation cannot be reconstructed solely from a gross undiscouted run-off triangle.
Under the EU's Solvency II Article 77, technical provisions generally comprise a best estimate and a risk margin. The best estimate is a probability-weighted present value of future cash flows, and reinsurance recoverables are calculated separately. The IAIS valuation principles similarly separate a current estimate from a margin over current estimate. These are related but distinct frameworks. An insurer cannot equate Mack prediction error, the IFRS 17 risk adjustment, and the Solvency II risk margin without identifying their conceptual, regulatory, and measurement differences.
ASOP No. 43 requires consideration of the purpose and applicable basis of the estimate. Claims handling expense, discounting, gross-to-net adjustment, salvage, subrogation, reinsurance collections, and defense outlays can change the final liability. A reserve should be represented by an explicitly described measurement tuple: valuation date, covered occurrence set, claim and accident assignment, currency, value basis, gross/net definition, expense basis, legal/contractual scope, and model. Without those fields, two equal numerical “reserves” need not measure the same thing.
Historical interpretation
Marine risk transfer historically predates contemporary statistical loss reserving; the Marine Insurance Act 1906 codified important aspects of insurance contract law in its jurisdiction. The Casualty Actuarial Society records its historical reserving principles as educational history rather than current professional standards. Modern stochastic reserving literature developed methods to move from observed development patterns to estimates of future claims and their precision. Mack's distribution-free standard error analysis and England and Verrall's stochastic reserving synthesis are landmarks in expressing uncertainty quantitatively. Their assumptions belong to statistical models, while the legal coverage and reporting processes determine what their input arrays contain. A historical connection between maritime insurance and actuarial estimation therefore requires both contractual interpretation and statistical observation rather than the assumption that long-standing marine practice yields stable development factors.
Reserving Triangles and the Information Architecture of Claims
Development coordinates and triangle definitions
A claims development triangle groups claims by an origin period and successive elapsed development periods. The origin can be accident year, underwriting year, policy year, reporting year, or another defined cohort. The choice is substantive. In accident-year analysis, a casualty enters the origin associated with when the insured event occurred. In underwriting-year analysis, the claim is linked to the inception of the relevant insurance contract; a contract covering casualties across months or years combines exposure and event timing differently. In reporting-year analysis, older accidents first reported today enter a contemporary origin cohort, altering the apparent distribution of future payments. A triangle is interpretable only with its origin, valuation lag, development interval, currency, valuation convention, and inclusion rules recorded.
Let C(i,j) denote the cumulative paid amount for accident cohort i observed at development age j. The incremental payment is X(i,j) = C(i,j) − C(i,j−1), with an explicit initial-period convention. The triangle contains cells for which observation dates do not exceed the valuation date; later cells are unobserved future cash flows. In an idealized homogeneous portfolio, development from j to j+1 is summarized by a factor f(j), estimated as the ratio of summed cumulative amounts for cohorts observed at both ages. An ultimate claim amount is then projected as U(i) = C(i,j) multiplied by the remaining factors. The unpaid amount is U(i) − C(i,j), before appropriate modification for other defined cash-flow categories. This basic chain-ladder reserving technique is intelligible precisely because its assumptions can be stated.
A paid triangle reflects money disbursed, adjusted for recovery and currency conventions if so specified. An incurred triangle typically combines paid amounts and contemporaneously booked case estimates, again only under a specified accounting definition. An outstanding-case triangle is a stock at successive valuation dates. A count triangle may track claims reported, open claims, closed claims, or reopenings. These arrays answer different questions. They may disagree systematically when a claims team raises an estimate at notification and pays only after a complex investigation. A low early paid share can coexist with a high incurred share; neither necessarily establishes that the other is wrong.
IBNR is often used broadly for liabilities attributable to incurred but not yet reported claims, sometimes incorporating future development beyond current case estimates in ordinary reserving usage. For analytical precision this monograph distinguishes true unreported claims (narrow IBNR) from insufficiency or excess in estimates of reported claims (IBNER). Mack's 1993 chain-ladder analysis addresses variability conditional on its model. The separation between IBNR and IBNER is essential when a maritime casualty is public knowledge but an injured claimant has not yet notified an insurer, or when reported claims remain subject to material quantum revision.
Data relationships and record-level reconciliation
A defensible input system links, at minimum, casualty ID, claim ID, coverage section, insured and claimant identity, policy year, accident date, notification date, claim opening date, event country, governing law, adjusting organization, original currency, payment transaction date, case reserve date, valuation timestamp, gross/ceded allocation, claim status, expense component, and relevant layer/limit. These attributes support reconstruction of multiple triangles rather than relying on a single irreversible summary table.
The ASOP No. 23 data-quality standard establishes expectations surrounding data selection, review, and appropriate disclosures. Its importance to transport reserving becomes concrete when a single collision is assigned both a hull claim number and a P&I matter number, or when an adjuster's reopened file creates a second administrative ID for the same underlying exposure. Deduplication must preserve legitimate distinct coverage heads. A casualty-level grouping prevents artificial independence assumptions; a claim-level grouping preserves separate legal outcomes.
Data must also distinguish source currency from reporting currency. If a claim is incurred in currency A, paid in currency B, and reported in currency C, mechanical conversion using current exchange rates can inject apparent development that is unrelated to claims experience. A transparent choice between transaction-date, accident-date, and valuation-date exchange rates, followed by sensitivity testing, improves interpretation. Net-of-reinsurance datasets introduce further complexity because reinsurance recoverable estimates may be revised on different dates from primary case reserves. Assembling an accurate net triangle requires a clear convention, not mere subtraction of unrelated valuation snapshots.
Reopenings and reductions in reported case reserves can produce negative incurred increments even when cash payments are nonnegative. Renshaw and Verrall's generalized linear model interpretation recognizes negative incremental observations as part of an enlarged statistical treatment. Model selection must still match the empirical sign structure; a straightforward nonnegative count or payment distribution cannot represent unrestricted negative increments without an additional specification.
Missing cells, right censoring, and observation selection
The upper-left observed triangle does not constitute a random sample of all ultimate claims. Older cohorts are more mature; recent cohorts are selected for early development only. Reported claims are additionally selected by the notification mechanism. A newly arising liability class, revised claims-reporting protocol, litigation funding change, new aircraft system, or maritime route change may yield different observation probabilities across origin periods. The missing lower-right cells are structurally censored by valuation time and cannot be treated as zero.
There are at least three selection problems. The first is ordinary development censoring: future settlements have not yet occurred. The second is notification selection: certain accidents are in the claims system only after evidence or legal action reveals a potential claim. The third is valuation selection: claims-handling teams revise case reserves in response to information such as litigation, expert reports, or changed counsel advice. A paid triangle may reduce the direct influence of case-estimate philosophy while making it harder to detect emerging severe claims that have not produced payment.
A fourth issue is portfolio composition. Suppose a historical marine segment consisted mainly of short-settling cargo damage and recent business is dominated by third-party pollution liability. Pooled triangle development factors would apply one mixture's settlement kinetics to another. Geographic differences in court procedure, liability limits, tort damages, salvage practice, and claims staffing can create the same distortion. An aggregation that improves arithmetic stability can simultaneously make the model less faithful to the risk. The right aggregation level is therefore a research decision supported by materiality, comparability, and available observations.
Case-estimate quality and actuarial adjustments
A claims professional may establish an initial case estimate before the insured's liability has been proven. An early case estimate might equal an estimated repair cost, a policy limit, a best-estimate settlement, a defense budget, or a deliberately cautious allowance. These quantities cannot be mixed across teams without reconciliation. Their development often reveals the insurer's reserving policy as much as the insured event's underlying severity.
The professional method should compare paid and incurred triangles, reported and closed counts, average payment per closure, open-claim case estimates, large-claim lists, expense development, and accident-to-notification/notification-to-settlement lags. Segmentation must be anchored to a feasible number of observations. Diagnostics should investigate whether a discontinuity corresponds to a system migration, new repair procurement practice, external inflation, exchange-rate translation, altered underwriting mix, portfolio transfer, new claims-handling instruction, or an actual change in risk. Where a discontinuity is systematic, adjustment to the triangle should be documented cell by cell with retained raw data.
ASOP No. 43's estimate basis and recoverables guidance provides the professional setting for these judgments. A reproducible reserve estimate records selected development factors, tail selection, exclusions of exceptional observations, parameter judgments, model changes, exposure normalization, and reconciliation from historical cumulative paid to the final current estimate. An actuary should explain an exclusion's effect on both the center of the estimate and the tail. Removing an exceptional aviation hull loss without recording the corresponding stochastic large-loss exposure may improve fitted historical stability while underestimating the next severe event.
Triangles versus individual-claim development
Aggregate triangles are strongest when a portfolio is sufficiently homogeneous and its development history contains informative repetition. Marine and aviation portfolios often violate these conditions in their financially dominant tail. A large vessel collision, multijurisdictional passenger claim, pollution response, or manufacturer-related aviation liability can dominate a year's losses. The reserve estimate should then combine aggregate development with named-claim or event-level analysis: legal scenario, liability allocation, cost-to-complete, specialist assessment, expected inflation, expected timing, defense costs, salvage, subrogation, and collection risk.
This is a methodological conclusion drawn from differences in observation units. An individual claim is a conditional distribution over settlement outcomes, not simply a cell in a development matrix. The existence of a catastrophe or litigation tail does not imply an automatic failure of chain ladder for smaller routine claims. It implies that the financial contribution of the thinly observed tail should be measured separately or explicitly accommodated, and that dependence among multiple claims arising from one occurrence must be retained through aggregation.
Settlement Delays and the Legal-Economic Causation of Development
The multiple clocks of claim settlement
A useful claims timeline records the accident, discovery, insured's notification, insurer acknowledgment, initial case estimate, survey or technical investigation, coverage position, liability allocation, demand valuation, mediation or legal decision, settlement agreement, initial and final indemnity payments, legal cost finalization, and recovery from counterparties and reinsurers. Some clocks run in parallel; some return to an earlier stage after evidence changes. In a large incident, distinct claimants and coverage heads follow distinct timelines while sharing accident causation.
Delay affects a reserve through two mechanisms. First, it delays the observation of the event or of its final economic amount, which affects parameter inference from development histories. Second, it changes the underlying future cash-flow distribution: repair and medical costs can inflate, legal fees accumulate, interest and damages may depend on applicable law, and salvage value can deteriorate or improve. Accordingly, a settlement lag cannot be modeled merely as a shift in the date of a fixed payment when delay may also change the amount.
The law provides an example. The United Kingdom's Insurance Act 2015 section 13A contains an implied obligation to pay sums due within a reasonable time, allowing for investigation and assessment and addressing reasonable grounds for dispute. This provision does not mean every slow claim is unlawfully delayed, and it does not apply automatically to every foreign policy. In appropriate cases, however, insurer behavior and legal remedies can influence expected outlay and timing. The reserve analysis should separate the contractual indemnity, claim adjustment expense, and any liability or cost associated with delayed performance according to the governing legal and accounting rules.
Marine hull and machinery
A marine hull file may initially be described as a machinery failure or collision, while the full scope of repairs is established at drydock. Alternative outcomes include partial repair, total loss, constructive total loss subject to policy requirements, betterment adjustments, salvage, and recovery from a third party. The policy's insured value, deductible, exclusions, sue-and-labor provisions, and valuation mechanism influence insured costs. Some losses are resolved quickly; others depend on contested causation or scope of reinstatement.
A common reserving error is to use gross shipyard repair cost as though it equals the insurer's ultimate claim. The relevant reserve is the expected insurer obligation, considering cover, value, limits, and recoveries. It can include expenses with different settlement clocks. When an adjuster initially records a gross repair quotation and later recognizes a subrogation potential, both incurred development and the recoverable schedule change even if the physical loss remains unchanged.
The UK marine proximate-cause provision helps explain why technical causation must precede coverage quantification in affected cases. The investigator's operational findings may be persuasive evidence, but they do not substitute for the insurance contract or legal interpretation. A machinery failure caused by ordinary wear, an external maritime peril, or a combination of causes can produce materially different insured outcomes.
Cargo, general average, and recovery paths
Cargo loss progression depends on the insured transit, transfer of risk and title, condition evidence, surveys, documentation, contractual delivery obligations, and potential recourse against carriers and other parties. Where general average or salvage arrangements arise, the insured's cash flow and subsequent reimbursement are governed by their precise legal and contractual basis. A claim transaction may thus appear as an early payment and a later net recovery; the net triangle can show negative development without any correction to the original physical loss amount.
The IUA's 2009 cargo clause record supplies a form-identification point, not a universal rule of adjustment. UNCTAD's Review of Maritime Transport 2025 documents rerouting and voyage delays in global shipping; those events may alter delivery risk and cost conditions but cannot be counted as insurance claims without confirming insured peril and loss. This separation avoids inferring an insurer's reserve from macroeconomic transport disruption.
A robust cargo reserve analyzes affected consignments and their values, condition and survey reliability, insured basis and deductible, likely salvage price, claims notification completeness, and the probability and timing of carrier recovery. Statistical models fitted to routine minor cargo losses may estimate a useful attritional component while larger transit disruptions require event-level correlated exposure assessment.
Protection and indemnity and third-party liability
P&I exposures can include claims by crew, passengers or others, cargo liabilities, collision, pollution, and specific expenses within club rules. Claims may involve multiple governments, courts, contractual counterparties, claimants, and insurers. Long lags are plausible where causation, environmental restoration, or quantum is disputed. One incident may contain many claimants and several coverage or reinsurance layers. A single initial reserve should not be treated as a stable ultimate merely because the casualty is known.
The International Group pooling agreements establish a market mechanism in which certain large risks are shared among member clubs according to defined terms and exclusions. The separately described 2026/27 reinsurance structure introduces a layer allocation and timing problem. The original club, the pool, a captive, and commercial reinsurers can have distinct share calculations, notifications, accounting dates, and collection histories. The group's published structure supports the existence of these mechanisms; an actual reserving decision must read the operative membership, pool, and treaty documents for the specific policy year and incident.
Consequently, gross ultimate paid loss, ceded ultimate, net retained loss, and expected collectible recoverables require distinct estimates. For a severe casualty with multiple claimants, a change in aggregate gross damages may cross an attachment point, causing nonproportional growth in ceded recovery. Pool or treaty exclusions may cause the opposite result: an apparent excess loss may remain largely retained. A portfolio's net loss development can therefore be less smooth than gross development, especially near the attachment and exhaustion points.
Aviation hull and liability
Aviation hull reserving frequently uses a well-defined physical asset, but material uncertainties still arise from agreed-value wording, damage investigation, repair versus total-loss decisions, components, parts availability, and the chosen legal/contractual mechanism. A technical grounding or regulatory action does not by itself imply that the hull insurer owes loss-of-use or business-interruption compensation. A covered insured event, the operative policy, any bespoke grounding extension, and exclusions must be established for each claimed head.
Aviation passenger and baggage exposures demand separate analysis from hull values. The Montreal Convention liability framework can determine relevant rights for qualifying international carriage, subject to the actual convention scope and limits. Other liability regimes may govern domestic transport, product defects, airports, ground handling, or claims against other defendants. A reserve must make legal scope an assumption or a verified fact; it cannot describe every aviation injury claim as governed by the same treaty.
Medical developments, injury prognosis, dependency claims, legal fees, choice of law, and settlement institutions may prolong liability claims. A manufacturer defect dispute can also divide ultimate economic loss among operators, manufacturers, component suppliers, maintenance organizations, and different insurers. The insurer's estimate should reflect the contractual insured, whose interest and policy respond, rather than summing every economic harm associated with an aviation event.
Litigation, information asymmetry, and reserving incentives
Dispute delays are endogenous: the time to settlement may depend on disputed liability, severity, litigation strategy, claimant information, jurisdiction, economic incentives, and the amount already reserved. A case estimate can influence settlement authority, and the settlement process supplies new information to the case estimate. This feedback makes simple independent-delay assumptions fragile. An observed increase in claim closure time can reflect a shift toward complex claims, a procedural backlog, a more adversarial legal environment, a deliberate case-management change, or worsening evidentiary uncertainty.
Scientific inference therefore requires comparison among like claims with comparable exposure and legal context. A claims office's median closure time is not a sufficient statistic for reserve sufficiency. It omits monetary weights, severity selection, open-claim censoring, and the last claims to close. The question is how the conditional distribution of unpaid dollars changes as a function of observable case features and elapsed time, including long-open cases and future reports.
Model Uncertainty, Stochastic Reserving, and Tail Exposure
Three distinct sources of uncertainty
Process risk is the randomness of future claim notifications, judgments, settlements, cash flows, and recoveries conditional on a model and its parameters. Parameter risk is uncertainty in estimating the model's parameters from finite, heterogeneous observations. Model risk is the possibility that the selected probability family, cohort definition, trend adjustment, dependency structure, or settlement mechanism fails to describe the relevant process. The distinctions are operational: a narrow confidence interval computed under one model does not bound model misspecification, an unrecorded coverage dispute, or a sudden change in legal rules.
Mack (1993) developed a distribution-free chain-ladder standard error under explicit assumptions about claims development. It offers a measure of prediction uncertainty without imposing a full parametric claim distribution. However, the method's reliance on a stable relationship among successive development factors does not validate it automatically for a portfolio that changed the mix of hull, cargo, and third-party liability or faced a discontinuous new catastrophe. The standard error addresses a specified model's statistical variability; selecting the model is a separate inferential task.
England and Verrall (2002) examine stochastic reserving methods, including alternative stochastic models, development-curve smoothing, tail extrapolation, Bayesian treatments of prior ultimate expectations, bootstrap and simulation. Wüthrich and Merz's stochastic reserving reference extends the methodological vocabulary. These sources provide legitimate tools for assessing reserve distributions, but the predictive distribution depends on data and modeling assumptions. No model identified here is a universal estimate of transport-insurance reserves.
The chain-ladder decision
Chain ladder is a sensible benchmark when historical development is informative, claim processing is sufficiently stable, and origin cohorts are comparable. Selected factors should be audited against age-to-age plots, prior-year residuals, paid versus incurred differences, calendar-year inflation, exposure mix, changes in claims systems, and the effect of exceptionally large files. A tail factor extrapolates beyond the longest observed age; when no historical data exists past year ten, the year-twenty liability tail is an assumption supported by outside information and sensitivity rather than a direct observation.
A practical large-loss treatment may segment routine claims from exceptional events with documented thresholds. The segmentation needs a rationale: an absolute dollar threshold changes real meaning under inflation and currency movements, and percentile thresholds require enough claims to estimate a percentile. Both excluded and included claims should appear in a reconciliation showing the impact on the gross estimate, net estimate, and uncertainty. Removing a large P&I event and replacing it with its event-level estimate may improve interpretability; removing it without a replacement understates exposure.
A second risk lies in origin-period drift. Exposure-based normalization is valuable when vessel insured values, aircraft flight cycles, cargo throughput, or aggregate insured limits evolve. These denominators are not mutually interchangeable. Cargo declared value is related to severity, while departures and passenger journeys are closer to aviation event frequency and airline liability opportunity. Modeling a severity index per departure may be sensible for some purposes but cannot substitute for a claims record or liability allocation. A carefully selected exposure base should match the stochastic risk being modeled.
Bornhuetter–Ferguson and prior-expectation approaches
For immature origin periods, the Bornhuetter–Ferguson approach combines an external or independently estimated expected ultimate with an expected percentage unreported or unpaid. In a simplified cumulative paid setting, one expression is: estimated ultimate = paid to date + expected ultimate × (1 − expected paid proportion at the valuation age). The method can reduce the instability of multiplying sparse early payments by a volatile development factor. Its reliability then depends on the prospective expected ultimate, the development pattern, and their consistency with the current portfolio. The basic methods discussion places this approach within established actuarial reserving techniques.
For marine hull, expected ultimate might draw on credible exposure-adjusted hull values and expected loss ratios. For aviation products liability or unusual P&I exposures, a useful prior may instead require independent engineering, legal, or reinsurance-market analysis. A prior constructed from the same distorted triangle is not genuinely independent information. Where the uncertain quantity is legal attachment rather than expected frequency, event-level probability scenarios should be considered alongside loss-ratio expectations.
A Bayesian approach makes prior assumptions and updating explicit. That mathematical transparency can be valuable when only a few severe losses exist, but posterior distributions inherit the choice of prior and likelihood. A well-defined sensitivity analysis varies the probability of an adverse liability outcome, conditional quantum, payment lag, and reinsurance eligibility. An insurer that reports one posterior mean without the influential prior choices deprives the reader of the most material reserve uncertainty.
Generalized linear and claim-level models
A stochastic generalized linear model can represent incremental development with origin, development, calendar year, and covariate effects. Renshaw and Verrall (1998) describe the GLM foundation of chain ladder. A model can incorporate explicit inflation or observed reporting shifts, but identifiability is limited: age, origin year, and calendar year are linearly related in a regular triangle, so their separate effects require constraints or external information. Fitting an unconstrained combination and calling each coefficient causal would overinterpret the data.
For large claims, the data may support individual-claim models instead of aggregate cells. Covariates could include policy type, jurisdiction, cause category, insured asset, claimants, time since notification, prior payments, documented reserve revisions, and the status of coverage or litigation. An event-level dependency factor is needed when one casualty leads to correlated claims. Prediction should be validated out of time where enough history exists; validation must not leak future case reserve revisions into earlier prediction dates. A model that uses post-settlement legal outcomes as predictors at the initial valuation date may achieve artificially low retrospective error while having no prospective predictive meaning.
Aggregate event dependence and reinsurance
Ordinary accident-year triangles often implicitly pool claims as though cohort variation is sufficiently represented by observed history. A casualty with several correlated legal claims violates an assumption of independent claim severities. If common causation drives hull, cargo, liability, salvage and environmental components, summing independent variance estimates can understate aggregate tail risk. Event-level simulation can preserve shared incident severity, common legal findings, correlated inflation and contingent reinsurance triggering. Where a catastrophe model is used, ASOP No. 38 sets professional considerations relevant to assessing models; it does not establish that any particular event has catastrophe-insurance coverage.
Net retained loss can be represented symbolically as gross ultimate less expected collectible ceded amount less relevant salvage and subrogation, after consistent accounting for commissions, expenses, and other applicable contract terms. This expression is only a schematic reconciliation; recoverables may be assets rather than components of an insurance liability in the applicable reporting system. Nonproportional layers produce nonlinear net exposure. Reinsurance can reduce the mean of retained insured loss while leaving collection, contract interpretation, and counterparty risk. Reporting only a gross reserve standard error for net capital decisions or only a net estimate for gross financial reporting obscures relevant relationships.
The International Group's 2026/27 structure provides a contemporary example of a layered marine program. Its official summary does not determine coverage for a named claim. In event-specific analysis, the evaluator must identify the cedent's share, valid coverage, per-claim or per-event aggregation, retention, layer attachment, limit, reinstatement, notification, collectibility, and applicable claims-control arrangements. A reporting system that stores only a final net cash total loses the data needed to assess these possible changes.
Capital, margins, and financial-statement interpretation
A statistical range of gross unpaid claims is not identical to a reported provision or solvency charge. IFRS 17's current fulfillment-cash-flow measurement incorporates a risk adjustment for non-financial risk. EU IFRS 17 disclosure requirements include disclosure of the confidence level used or an equivalent level corresponding to another risk-adjustment method. Solvency II best estimate and risk margin are a different regulatory construct. The IAIS ICP 14 valuation framework provides principles for supervisory valuation, but local legal requirements govern implementation.
Actuarial risk communication must state whether amounts are nominal or discounted, gross or net, best estimates or margins, current incurred claims or future coverage obligations, and whether dependence and parameter uncertainty are included. The 2024 revision of ASOP No. 36 on reserve opinions concerns reserve-opinion communication under its defined scope; it is not a substitute for independent assessment of portfolio evidence.
Reserve adequacy also matters for enterprise governance. The current ASOP No. 58, effective May 2025, replaced the former ASOPs 46 and 47 in the US professional standard framework. An analysis citing the repealed standards as current without qualification would be historically inaccurate. Board-level reserving should show how changes in estimated liabilities, uncertainty ranges, reinsurance collection, and correlated shocks enter capital, risk appetite, and management actions under the enterprise's applicable regulation.
Evidence Appraisal, Stress Testing, and Reproducible Illustrations
A source hierarchy for this subject
Primary legal text supports propositions about the legal test when the law applies. Operative policy wording supports propositions about contractual cover and exclusions when incorporated in an actual contract. An official accident docket supports recorded investigative facts but rarely an actual insurer's final claim payment. A claims ledger establishes booked transactions subject to its completeness and controls. An actuarial journal article supports a method under described assumptions, not an empirical assertion about an unobserved portfolio. A supervisory or reporting standard supports measurement requirements within its regime. These are complementary forms of evidence with different observation powers.
The Marine Insurance Act 1906 provision and the UK late-payment rules have legal force within their applicable scope. The IUA cargo clauses catalogue entry documents a standard wording's identity but does not verify incorporation. The NTSB docket search and IMO marine casualty investigation page document investigative systems rather than claims-finance outcomes. ASOP No. 23 and ASOP No. 43 define professional approaches within their own jurisdictional framework. The inferential error to avoid is using one source category as though it directly establishes another category's claim.
Synthetic paid triangle: a transparent benchmark
Consider the following artificial cumulative paid triangle, expressed in millions of dollars. Accident-year labels are illustrative, and the latest valuation date is assumed to fall after the fourth development age of the oldest origin. Every cell is invented solely to make an established actuarial calculation reproducible; no cell is an insurer observation.
| Accident origin | Age 1 | Age 2 | Age 3 | Age 4 |
|---|---|---|---|---|
| 2022 | 10 | 16 | 18 | 19 |
| 2023 | 12 | 19 | 21 | — |
| 2024 | 15 | 23 | — | — |
| 2025 | 18 | — | — | — |
Using volume-weighted chain-ladder development factors over pairs observed at both ages, the first factor is (16 + 19 + 23) / (10 + 12 + 15) = 58/37 = 1.5675676. The second factor is (18 + 21) / (16 + 19) = 39/35 = 1.1142857. The third factor is 19/18 = 1.0555556. Assuming no further development beyond age 4, estimated ultimates are 19.0000 for 2022, 22.1667 for 2023, 27.0524 for 2024, and 33.1876 for 2025. The sum is 101.4067. Latest observed paid amounts total 19 + 21 + 23 + 18 = 81, producing a projected unpaid amount of 20.4067 million. Every figure follows directly from the stated cells and factor rule.
This example demonstrates algorithmic arithmetic, not reserve reliability. Only one cohort informs the oldest factor, no additional tail is assumed, and the most immature cohort's ultimate is driven by a large cumulative multiplier. To make the calculation suitable for real use, the analyst would test exposure changes, data coherence, model fit, factor variability, calendar inflation, large losses, expenses, discounting, and unreported exposure. Mack's standard error calculation cannot be meaningfully inferred from this small illustration without adopting and testing assumptions; the single age-3-to-age-4 transition also leaves inadequate empirical degrees of freedom for that factor's variability.
For the final origin the cumulative multiplier to age four is approximately 1.8438. A change of 10% in its future multiplier changes that origin's estimated ultimate by approximately 3.3188 million. This is a pure numerical sensitivity: it shows why immature cohorts should not be judged solely from the central chain-ladder point estimate. Conversely, supplementing an unstable early-year factor with a credible independent expected-ultimate view is compatible with the established Bornhuetter–Ferguson family of techniques.
Synthetic disputed liability and reinsurance scenario
Consider a single notified transport claim with two mutually exclusive gross settlement scenarios: a covered and economically limited outcome of $2 million, and a severe adverse outcome of $20 million. Assume for illustration a 40% probability of the severe outcome at valuation and 60% of the smaller outcome. Expected gross loss is (0.60 × 2) + (0.40 × 20) = $9.2 million. If new legal evidence changes the probability of severe loss to 25%, expected gross loss becomes $6.5 million. If it rises to 60%, expected gross loss becomes $12.8 million. The spread between those two updated means is $6.3 million even though no payment has been made and the originally reported casualty is unchanged.
Now assume an illustrative excess-of-loss treaty that attaches above $10 million and pays at most $10 million on an eligible gross claim, with no aggregate complication. The small scenario produces no ceded payment; the severe scenario would contractually generate $10 million. Under the original 40% probability, expected ceded recoverable before collection risk is $4 million and the expected retained loss is $5.2 million. If, in addition, only 90% of the contractual severe-scenario recovery is expected to be collected, expected collectible recovery is $3.6 million and the schematic retained economic expectation is $5.6 million. The collection factor is also entirely hypothetical. It is not a prediction of a named reinsurer's performance.
The scenario reveals a structural asymmetry: the variation in severe-liability probability affects gross loss, expected ceded recovery, and retained loss at different rates. Policy defenses, legal venue, coverage limits and treaty terms could change which outcomes are eligible. The real P&I pooling contract structure is more complex than this abstract layer and cannot be replaced by the example. The purpose is to show why liability probability, recovery eligibility, and recoverable collection must each be preserved as independently reviewable assumptions.
Settlement time, nominal amount, and discounting
If a fixed nominal payment of $10 million is expected three years after valuation and the applicable annual discount rate is assumed to be 4%, the present value is 10/(1.04)^3 = approximately $8.8900 million. At eight years it becomes 10/(1.04)^8 = approximately $7.3069 million. Later payment reduces present value when the nominal amount and discount rate are held constant. Real settlement delay can also increase nominal cost through repair inflation, claimant damages, legal expense, interest, or currency changes; the present-value impact of delay can therefore be positive or negative depending on those correlated changes.
For example, if the amount starts at $10 million in today's prices and grows at 3% annually until settlement, its three-year nominal value is 10 × (1.03)^3, while its discounted value under the hypothetical 4% curve is 10 × (1.03/1.04)^3. A separate stochastic inflation, interest-rate and duration analysis would be appropriate where these factors matter. The financial statement must use its required discounting and risk-adjustment framework, rather than adopting these round-number assumptions as accounting policy.
Stress-test matrix
| Driver | Base observation | Alternative scenario | Main reserve implication |
|---|---|---|---|
| Reporting delay | Stable notice lag | Material late-reported liability | Missing IBNR and recent-origin bias |
| Case reserving | Consistent initial estimates | Revised liability assessment | Incurred development distorted without cash movement |
| Repair economics | Stable real costs | Parts shortage or drydock inflation | Larger nominal hull and cargo payments |
| Legal regime | Known cover and venue | Contested jurisdiction and liability | Scenario probability and settlement duration change |
| Catastrophe clustering | Mostly independent routine claims | One incident with many claimants | Correlated tail and aggregate-limit effects |
| Reinsurance | Clear valid layer attachment | Exclusion, allocation or recovery dispute | Gross-to-net divergence and recovery lag |
| Currency | Stable translation basis | Volatile settlement currencies | Calendar pattern and reported reserve translation shift |
| Claims data | Continuous comparable records | Migration, missing notifications, reopened IDs | Apparent development may be data artifact |
The table is a diagnostic checklist, not a quantified scenario model. The quantitative effect should be assessed case by case under recorded assumptions. In particular, an accident report or UNCTAD's transport disruption analysis cannot by itself produce an ultimate claim distribution. Public risk reports describe exposures and events; financial claims data connect the events to insured values, payments, and coverage.
Limits of available empirical evidence
The cited public regulatory and scholarly sources verify the existence of legal rules, contractual market mechanisms, reporting requirements, and established estimation methods. They do not provide a common international database of claim-level paid and incurred histories spanning marine hull, cargo, P&I, aviation hull, and aviation liability. The official EASA 2026 safety review and IMO casualty investigation materials provide accident and safety evidence, not a validated multinational claims reserve sample. Neither reserve-development percentages nor severity tail quantiles for the worldwide market can responsibly be reported from those sources alone.
Some academic sources were checked through publisher abstracts and metadata, and proprietary policies and treaty schedules were not obtained. Their full quantitative results are accordingly not asserted. The UK legislative provisions are used to illustrate a legally bounded mechanism rather than global law. Treaty applicability and liability caps remain specific to the carriage and relevant state participation. Actual accounting applications require the reporting entity's basis, reporting date, contracts and financial circumstances. These limitations constrain the level of empirical generalization while leaving the conceptual and numerical distinctions testable.
Scientific Synthesis and Research Questions
What the evidence establishes
The sources jointly establish four connected propositions. First, legal coverage and the structure of the insured obligation precede statistical reserving: the same casualty can generate multiple contractual claims or no covered indemnity under a particular policy. Second, the triangle is an observation of booked developments conditioned by claim-reporting practices and settlement decisions; it is not a direct history of ultimate physical damage or economic disruption. Third, classical stochastic methods offer useful estimates and uncertainty measures under their assumptions, yet those assumptions must be checked against sparse and changing marine/aviation portfolios. Fourth, a gross estimate, net retention, reinsurance recoverable, risk adjustment and solvency margin answer different questions and require explicit reconciliation.
The legal and contractual sources support the first proposition through the Marine Insurance Act 1906, Montreal Convention overview, and P&I pooling documentation. The second is supported by the definition of data and unpaid claims in ASOP No. 23 and ASOP No. 43, together with explicit data-generating-process reasoning. The third is informed by the Mack 1993 study, England–Verrall 2002 study, and Wüthrich–Merz treatise. The fourth follows from the separate measurement requirements in IFRS 17, Solvency II Article 77, and the IAIS valuation framework.
Original analytical interpretation
The central interpretive contribution is a transparent causal description of why reserve uncertainty concentrates in the interface between insured event, claims administration, and legal liability. A claims triangle compresses the event-to-payment pathway into development cells. Compression is efficient when the pathway is repeated in a relatively stable population; it becomes fragile when a rare casualty rearranges coverage, evidence, claimant identity, or recovery obligations. This conclusion follows from the elementary information lost during aggregation and from the documented differences in legal and actuarial measurement. It does not require naming or asserting a new theoretical framework.
A second analytical conclusion is that settlement delay should be modeled simultaneously as an observation process and an economic process. An unreported claim is absent from the ledger but may already be an obligation; a reported dispute can be recorded yet highly uncertain in quantum; a settled claim can remain unpaid; a paid claim can still have a disputed ceded recovery. Each stage changes both the available information and the forward-looking cash flows. Explanations based only on a single average closure lag therefore discard information material to the reserve.
A third conclusion concerns diversification. Correlation is not located solely in asset classes or geographic catastrophe perils. It may originate in one legal fact affecting several coverage heads, a common component failure affecting multiple insureds, a common maritime casualty producing many claimant categories, or a shared reinsurance counterparty. Some of this dependence may be detectable through event identifiers even when triangle cells look smooth. The lesson for reserve analysis is to preserve shared-cause groupings before estimating aggregate uncertainty.
A fourth conclusion concerns communication. The number describing unpaid claims should always travel with its unit and measurement statement. An internal accounting estimate, an external solvency best estimate, an IFRS 17 incurred-claims liability, and an underwriting loss-ratio forecast can differ appropriately. Decision-makers need a bridge from the transaction ledger to each reported value and a sensitivity view showing which assumptions drive the differences.
Research agenda grounded in these limits
A credible international empirical program would collect pseudonymized claim records with accident and report dates, claim head, policy period, coverage status, origin currency, claim-level paid and incurred evaluations, expense data, key legal events, settlement dates, reinsurance attachment and recovery dates, and common-event identifiers. Privacy, contractual confidentiality, and antitrust constraints would need to be addressed before cross-company pooling. A harmonized data dictionary would distinguish marine hull, cargo, P&I, aviation hull, passenger/carriage liability, and other liability heads instead of combining them under one “marine and aviation” line.
The primary empirical tests would compare accident-year and underwriting-year development, estimate reporting and settlement lag distributions by coverage head, examine the effect of case-reserving changes, measure how error varies with sparse origin cells, and backtest gross and net estimates separately. Models should be compared through genuinely out-of-time observations and transparent scoring rules, with particular attention to large-event holdouts. Where sample sizes permit, comparisons would test whether event-level legal covariates improve reserve forecasts over triangle-only baselines, and whether dependence materially alters high quantiles of aggregate unpaid claims.
Further analysis should track how legal uncertainty translates into discounted cash-flow uncertainty and how real treaties allocate that uncertainty among risk bearers. It should examine the timing of reinsurance recoverables, contractual exclusions, and legal-expense treatment under the exact terms applicable to each occurrence. Specific jurisdictions and policy-year regimes must be stated in every study. A shared claims database with sound event IDs would enable these questions to be answered empirically; their present absence is an evidence gap rather than proof of a market-wide result.
Conclusion
Marine and aviation claims reserve uncertainty arises from the interaction of exposure, insured events, claim notification, coverage law, technical evidence, liability assessment, settlement behavior, cash-flow timing, and risk transfer. Accurate estimation begins by specifying what is being reserved and whose obligation it is. Development triangles remain important methods for summarizing historical payments and case estimates, but they are conditional records of administration as well as loss. The most material uncertainty often lies in long-tail liability, disputed causation, exceptional physical damage, concentration within one incident, and the timing and eligibility of reinsurance recovery.
The review supports a disciplined combination of aggregate development methods, independent expected-ultimate judgments, stochastic uncertainty analysis, and individual-event legal/engineering assessments. Its worked examples show how a small set of paid observations can yield a reproducible yet fragile chain-ladder result, and how disputed liability can materially alter gross and net reserves without changing recorded cash movements. The appropriate treatment of margins, discounts, recoverables, and disclosures depends on the applicable reporting framework, including IFRS 17, Solvency II, or other local regimes.
The final research conclusion is practical and verifiable: each significant reserve estimate needs a traceable connection from casualty and contract to claim file, valuation, statistical model, and financial statement. That connection permits uncertainty to be located, challenged, and updated when evidence changes. Broader claims about the global distribution of transport-insurance reserves require insurer-level empirical data not supplied by the public primary sources examined in this monograph. The text contributes a comparative analytical synthesis and a reproducible educational demonstration; it makes no claim of a new original actuarial theory, independently estimated international reserve distribution, or external peer review.
References
Actuarial Standards Board (2011). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates.
Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality.
Actuarial Standards Board (2024). Actuarial Standard of Practice No. 36: Statements of Actuarial Opinion Regarding Property/Casualty Loss, Loss Adjustment Expense, or Other Reserves.
Actuarial Standards Board (2021). Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas).
Actuarial Standards Board (2024). Actuarial Standard of Practice No. 58: Enterprise Risk Management.
Mack, Thomas (1993). Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates. ASTIN Bulletin 23(2), 213–225.
England, Peter D., and Richard J. Verrall (2002). Stochastic Claims Reserving in General Insurance. British Actuarial Journal 8(3), 443–518.
Renshaw, Arthur E., and Richard J. Verrall (1998). A Stochastic Model Underlying the Chain-Ladder Technique. British Actuarial Journal 4(4).
Wüthrich, Mario V., and Michael Merz (2008). Stochastic Claims Reserving Methods in Insurance. Wiley Finance Series.
IFRS Foundation (2026). IFRS 17 Insurance Contracts: Standard overview.
European Union / EIOPA (2026). Solvency II Directive 2009/138/EC, Article 77: Calculation of Technical Provisions.
International Association of Insurance Supervisors (2026). Insurance Core Principles and ComFrame, ICP 14 Valuation: Online Tool.
United Kingdom (1906). Marine Insurance Act 1906, Section 55: Included and Excluded Losses.
United Kingdom (2015). Insurance Act 2015, Part 4A, Section 13A: Late Payment of Claims.
International Air Transport Association (2026). Montreal Convention 1999: Framework and Periodic Liability-Limit Revisions.
International Maritime Organization (2026). Casualty Investigation Code and Marine Casualty Investigation.
National Transportation Safety Board (2026). Accident Docket Search.
International Group of P&I Clubs (2026). Group Agreements and Pooling Agreement.
International Group of P&I Clubs (2026). 2026/27 Pool and GXL Reinsurance Contract Structure.
International Underwriting Association (2009). CL382 Institute Cargo Clauses (A), 1 January 2009.
European Union Aviation Safety Agency (2026). Annual Safety Review 2026.
UN Trade and Development (2025). Review of Maritime Transport 2025: Staying the Course in Turbulent Waters.
European Union (2021). Commission Regulation (EU) 2021/2036 on Adoption of Amendments to IFRS 17: Disclosure.
Casualty Actuarial Society (2020). Statements of Principles: Historical and Educational Archive.