RESEARCH DOSSIER
Aviation Loss Severity and Sparse Data Estimation
Abstract. Aviation insurance combines an unusually selective accident process with potentially exceptional physical-damage, passenger, third-party, product, and consequential liability outcomes. Statistical scarcity arises because the observations most relevant to upper-layer insurance and capital decisions are rare, commercially confidential, historically heterogeneous, incompletely developed, and governed by changing contracts and legal regimes. This monograph develops a source-based and analytically reproducible treatment of aviation loss severity under these conditions. It distinguishes aircraft occurrences from insured events; occurrence losses from claims and settlements; hull values from tort and convention liabilities; individual claims from event accumulation; and the frequency of qualifying events from the conditional size of their insured consequences. Using current institutional accident statistics, aviation liability instruments, actuarial standards, and foundational literature on extreme values, credibility, and stochastic reserving, it examines the joint problems of exposure definition, selection, censoring, threshold choice, expert elicitation, hierarchical borrowing, uncertainty propagation, and dependence. A fully labeled synthetic example derives the effect of zero observed extreme events on a gamma-Poisson posterior and the expected recovery from a generalized Pareto excess layer. The central finding is that the effective information content of a sparse aviation portfolio depends on its observation process and contract mapping as much as on its count of losses. Scientifically defensible estimation therefore requires explicit estimands, dated legal and coverage transformations, prior-predictive and posterior-predictive diagnostics, and sensitivity ranges that preserve both statistical and structural uncertainty. All calculations are illustrative; no proprietary loss sample, real-world insurance price, or new actuarial doctrine is claimed.
Keywords: Aviation insurance; loss severity; sparse data; rare-event estimation; aircraft hull; airline liability; event aggregation; exposure measurement; heavy-tailed distributions; generalized Pareto distribution; Bayesian credibility; expert elicitation; claims development; reinsurance; model risk; Solvency II; Montreal Convention.
Introduction
The economic task addressed here is the estimation of a distribution of future covered aviation losses, with particular attention to layers for which direct portfolio experience is very small. The relevant random variable must first be defined. The loss suffered by an aircraft operator is not automatically the loss paid under its insurance contract; an accident report records a safety occurrence rather than its legal and contractual settlement; a report of fatalities expresses human consequences rather than a monetary claims distribution. Each may be informative, but each answers a different question.
Aviation provides a vivid empirical illustration of the difference between occurrence count and consequence. The International Civil Aviation Organization's 2026 aviation safety analysis reports 85 accidents in 2025 for its defined scheduled-commercial-air-transport population, compared with 95 in 2024, and four fatal accidents in 2025 compared with ten the previous year. At the same time, its reported fatalities rose from 296 to 387. The International Air Transport Association's 2025 safety report announcement identifies 51 accidents under IATA's own classification criteria, eight of them fatal, across approximately 38.7 million sectors. These are different statistical frames, with different admissibility and classification rules. Their numerical difference establishes a practical requirement: every aviation rate must be paired with an explicit population, event taxonomy, data vintage, and denominator.
The analysis applies across international airline fleets, business aviation, cargo, lessors, airports, aviation products, and the insurer or reinsurer carrying the resulting risk, while treating each as a potentially different statistical population. The principal discussion concerns civil aviation hull and liability. Specific sanctions, war risks, local statutory remedies, and policy endorsements are evaluated at the date and jurisdiction of the insured exposure. Evidence reviewed through November 17, 2026 is used for descriptive claims. The study uses analytical constructions rather than pretending that institutional safety data are a complete insurance claims panel.
Conceptual and Historical Foundations
The unit of observation is part of the estimand
Let an insurance portfolio contain exposure units indexed by aircraft, operator, route, policy period, or event. Let denote the number of qualifying covered claim events over a defined time interval , and let be the ultimate insured cost of event after the specified coverage transformation. The aggregate ultimate loss can be expressed as
This familiar collective-risk representation conceals material choices: the unit of exposure in the frequency model; the grouping rule for multiple claims arising from one occurrence; the ultimate-versus-paid valuation date; the inclusion of legal defense and loss-adjustment expenses; and the effect of policy limits, deductibles, currencies, proportional participation, reinsurance, and contract years. An aircraft ground-damage event may yield one hull claim, several liability claims, and a contingent loss for a lessor. Conversely, several legal claims can arise from a single collision. Counting claim files as independently occurring accidents would distort both frequency and the tail dependence of severity.
For an insurer, one useful object is the ultimate ground-up economic loss , with a contractual payment function . The covered amount is , where describes limits, attachment points, exclusions, indemnity basis, liability allocation, and the applicable insured interest. The function may also depend on the event, jurisdiction, and claim date. For a simple indemnity with occurrence deductible and insured limit , the illustrative mapping is
Real aviation programs may contain multiple insuring agreements, different sublimits, aggregate limits, reinstatements, allocation provisions, or separate war-risk arrangements. The simple formula is a component, not a representation of every policy. The first methodological duty is to determine which payments are being estimated: gross ground-up event cost, gross insurer liability, net retained loss, ceded reinsurance recovery, or financial-statement reserve.
Four sources of apparent scarcity
Sparse experience has at least four substantively different causes. Physical events may be uncommon relative to a large number of flight sectors. The insured portfolio may be small in aircraft-years despite a substantial international aviation system. The insurer may observe only losses that attach to its layer or that have been reported and assigned to it. The data record may exclude outcomes governed by another insurer, excess layer, jurisdiction, or reporting period. Each cause implies a different remedy. Adding more flight observations addresses a rare-event frequency problem only when the additional flights belong to a relevant risk class. Acquiring claim-development history addresses valuation incompleteness. Recovering policy terms addresses contract observability. Borrowing external experience addresses sample size but creates transportability questions.
The history of aviation risk illustrates why loss variables change over time. The mix of hull values, aircraft size, seating capacity, traffic density, navigation systems, safety oversight, legal liability, and capital protection evolves. A severity amount from an older event requires re-expression in a common economic and contractual basis before entering a contemporary model. Adjusting a historical claim merely by a general consumer price index may miss aircraft replacement prices, engine repair capacity, legal settlements, aircraft utilization, and shifts in currency and insurance wording. Modelers should document an identifiable chain from original claim to normalized observation.
Safety records and claim records serve different public functions
The U.S. National Transportation Safety Board maintains an aviation accident data census, with downloadable files and coding guides for different historical periods. Under the U.S. accident definitions in 49 CFR 830.2, aircraft accident, fatal injury, incident, and substantial damage have specified meanings. These definitions provide a transparent event taxonomy for the jurisdiction; they do not transform a record into a verified payment under a specific insurance policy. Investigation and claim adjudication also have different evidentiary functions. Safety findings, causation allegations, litigation outcomes, and insurer settlements must be reconciled rather than substituted for one another.
The International Civil Aviation Organization publishes annual safety reports and offers a separate aviation data service listing accident and incident endpoints. Those resources identify occurrence-level information and denominators that may assist risk analysis, subject to access, definitions, completeness, and licensing. Insurer claim extracts must additionally contain the policy, coverage, reporting, and valuation fields necessary to compute ultimate insured losses. This distinction is foundational for severity estimation because a safety database is rich in event circumstances but generally incomplete in contract and paid-claim variables.
Insured persons, carrier obligations, and limits
Aviation severity extends beyond the aircraft hull. The Montreal Convention 1999 regulates specified international air-carriage claims for passenger injury or death, delay, baggage, and cargo, subject to treaty application and case-specific facts. According to the International Air Transport Association's Montreal Convention overview, the treaty's relevant limits undergo periodic review. The Canadian Transportation Agency's 2024 revision notice records the revised amounts effective December 28, 2024: 151,880 special drawing rights (SDR) for the Article 21 first-tier passenger injury/death threshold; 6,303 SDR for passenger delay; 1,519 SDR for baggage; and 26 SDR per kilogram for cargo. The Article 21 figure is a liability-tier threshold, not an absolute ceiling on damages for death or bodily injury. Currency conversion dates, treaty membership, exceptions, and proof requirements matter for the liability estimand.
In the European Union, Regulation (EC) No 785/2004 establishes aviation-specific minimum insurance requirements for specified air carriers and operators in relation to passengers, baggage, cargo, and third parties. The regulation's minimum insurance coverage is conceptually different from a private policy's negotiated limits and from the amount ultimately awarded to a claimant. Minimum insurance requirements, treaty damage rules, and policy payment provisions occupy three different stages of the loss mapping. The UK Civil Aviation Authority guidance illustrates how another competent authority expresses minimum insurance obligations and SDR-based categories for its jurisdiction. Its published baggage and cargo minimum figures should not be silently treated as the later Montreal Convention revised damage limits; legal instruments and applicable amendments must be checked separately.
Consequences for actuarial scope
The variable of interest for a commercial airline hull risk could be a repair amount per aircraft-damage claim, an agreed-value total loss, or the annual aggregate net retained hull burden per fleet. A liability book might require bodily-injury damages, defense costs, passenger counts, settlement lags, and jurisdictional outcomes. An airport liability portfolio could be exposed to an accumulation of surface-damage and injury claims associated with one ground movement. Products liability requires a different unit: potentially common design, manufacturing, or maintenance causes across aircraft and policy years. The statistical model must follow the insured obligation, and the insured obligation must be reconstructed from the actual legal and contractual instruments.
Rare Events: Records, Exposure, and Legal Insurability
Denominator discipline and classification risk
For an airline operational event frequency , an approximate observed rate is , where is the count of events satisfying the selected rule and is the matching exposure. A published rate per million departures cannot be directly applied to aircraft-years, flight-hours, passenger-kilometers, cargo ton-kilometers, or sum-insured years. Conversion requires a defensible exposure relationship. Flight sectors measure takeoff-and-landing opportunities; flight hours place different weight on long-haul activity; passenger counts help determine certain liability accumulation magnitudes; insured hull value-years measure monetary assets under cover. Each denominator can be useful when connected to a corresponding causal and contractual mechanism.
The contrasting 2025 ICAO and IATA accident figures demonstrate the size of classification risk even before actuarial estimation begins. ICAO's July 2026 safety assessment states that its examined population comprises scheduled commercial aircraft above 5,700 kilograms maximum takeoff mass and reports approximately 38 million flights. IATA's March 2026 classification summary reports its accident count under a distinct industry classification. The respective figures of 85 and 51 are therefore not interchangeable estimates of one identically defined numerator. A model fit to one series and evaluated against the other must first harmonize scope, event definitions, data revisions, and exposure measurement.
The exposure frame must also match the insured book. A worldwide airline benchmark may be weakly relevant to a small fleet of specialized cargo aircraft, helicopters, training aircraft, or general-aviation operators. Distinctions among jet and turboprop operations, airfields, maintenance environments, route geography, dispatch and ground-handling practices, regulatory oversight, and aircraft types may affect both frequency and conditional severity. Data pooling increases sample size at the cost of mixing heterogeneous risks; stratification reduces apparent homogeneity but may leave extremely small cells. The model should state the exchangeability assumptions that make pooling credible.
Accident, claim, attachment, and settlement filters
A complete observation chain is: operated exposure → physical occurrence → detection and classification → legal or contractual responsibility → policy attachment → claim reporting → case valuation → litigation or settlement → ultimate paid outcome. Each transition has its own selection process. A recorded occurrence may have no insured damage. A covered ground-damage claim may never appear in a catastrophic-accident study. A severe event may produce several insurers' independent reported files or only a single consolidated settlement. An excess-layer dataset may contain only losses exceeding a reporting threshold, while losses below it are unavailable.
Let denote the ground-up size of a qualifying claim. If only claims above a disclosure threshold are visible, the observed density is
Ignoring the denominator overstates the prevalence of large losses in the underlying claim population. If losses below an attachment point are recorded as zeros, one must distinguish structural no-claim outcomes from censored or unobserved values. A monetary excess-of-loss program may create left-truncated experience. Upper limits can create right-censoring of the original ground-up severity: a payment of the full policy limit identifies a minimum underlying loss but does not reveal its upper tail. Claim development creates additional measurement noise, often correlated with severity. An estimator that treats all open losses as fully developed observations is exposed to reporting-lag and maturity bias.
Zero events are an observation, not a zero risk parameter
When independent qualifying claims are represented approximately by a Poisson process with rate per exposure unit, the chance of no observed claims over exposure is . A zero count is therefore consistent with a range of positive frequencies, particularly when is small. Exposure can be substantial in operations but small in the specific loss layer: for example, a portfolio may encompass many flight sectors yet provide no claims in a high excess liability layer over a decade. The rarity of layer attachment can dominate the precision of its severity estimate.
A pure frequency model also needs a convincing unit of independence. One severe weather occurrence or ground-service failure may affect several aircraft and create correlated payments; one systemic manufacturing defect may generate claims across years and locations. Poisson independence may approximate idiosyncratic events while failing to describe common-cause or aggregation risk. A point-process or scenario representation for cluster events, with explicit cross-policy aggregation, is often preferable to multiplying a single-aircraft probability by the fleet count.
Legal and policy changes introduce breaks
Treaty revisions, case law, policy limits, extensions, exclusions, deductibles, liability allocation, and reinsurance structure may change the part of the ground-up loss distribution that an insurer bears. The Montreal liability revisions effective December 2024 are a concrete example of a dated change relevant to some internationally governed claims. The European aviation insurance regulation is a separate source for minimum coverage rules. A historical payment dataset spanning several regimes must encode the relevant treaty applicability and policy response; otherwise apparent inflation or tail thickening may partly reflect changes in legal obligations and insurance coverage.
The responsible procedure is to preserve both the as-observed claim and its restatement under a declared evaluation basis. Restatement may require unavailable files, such as original policy wordings, litigation records, claim reserves, or exchange-rate dates. In their absence the result should be a scenario distribution across plausible mappings rather than an apparently precise corrected loss. The data's absence is part of the reported model uncertainty.
Severity Tails: Statistical and Contractual Mechanisms
The severity distribution and the quantity of interest
A severity model should begin by separating the severity of a physical occurrence from the distribution of claim amounts conditional on a qualifying insured claim. If denotes the total economic consequence and the contractual payment, the transformation may produce a mass at zero, an accumulation at the policy limit, and an altered upper-tail shape. A very heavy-tailed distribution of ground-up liability does not imply that each insurer's capped payment has the same tail, although aggregate event losses and multiple limits can restore considerable tail exposure at the portfolio level. A model estimating for pricing, for capital, or for a treaty layer is answering a different question each time.
The loss process also has multiple clocks. The occurrence date determines exposure and, frequently, the applicable policy and treaty year. The report date determines when the claim enters a recorded triangle. The accounting evaluation date determines the valuation of incurred losses. Settlement and payment dates influence discounting, litigation exposure, and inflation. For rare large aviation liabilities, a single claim can develop over multiple valuation years. A data extract must identify each clock or its uncertainty before the modeler treats its reported severity as an ultimate amount.
A practical severity decomposition distinguishes hull repair, hull total loss, death and bodily injury, legal defense, airport or ground third-party property damage, product liability, investigation and adjustment expenses, and policy-specific ancillary covers. The amount and identity of the policyholder must be fixed. A lease agreement may assign an owner, operator, and financier different interests in one aircraft; their nominal insured amounts do not necessarily sum to the economic value of the machine. For liability, the number of claimants, the jurisdictional remedies, and the aggregation rule may matter more to insured severity than the physical damage to the aircraft.
Thin tails, subexponential tails, and extreme-value approximations
Standard actuarial candidates include lognormal, gamma, Weibull, Pareto, Burr, mixtures, and distributions with separate body and tail components. A lognormal body may describe many routine repairs while failing to reproduce rare liability settlements at the reinsurance attachment point. A Pareto-type tail can allocate more probability to exceptional payments, but with few observations the shape can be overwhelmingly controlled by modeling assumptions. Distributions should be compared on out-of-sample and decision-specific quantities, including high-threshold exceedance probabilities and layer expected losses, rather than by an in-sample global goodness-of-fit number alone.
Extreme-value theory supplies a conditional approximation for excesses above a sufficiently high threshold. Balkema and de Haan's residual-life limit analysis and the subsequent peaks-over-threshold literature provide the mathematical context for the generalized Pareto distribution (GPD). McNeil's loss-severity tail estimation analyzes why such methods are relevant to insurance loss tails. For , with threshold , scale and shape , the GPD survivor function is
for positive support satisfying $1+\xi y/\beta>0\xi=0$ is . Positive represents an unbounded polynomial-type tail. Negative imposes an upper endpoint. In an idealized uncapped GPD, the mean excess exists for and equals ; the variance exists only for . Actual insured payments may be capped by contracts, so mathematical divergence of an assumed ground-up distribution must be kept separate from the finite maximum payment under a single capped policy.
A robust peaks-over-threshold fit requires sufficient exceedances, a threshold high enough for an appropriate tail approximation, and stability checks over a defensible threshold range. The central aviation difficulty is that these conditions can conflict. Raising the threshold removes the ordinary body of the distribution but leaves very few independent extreme claims. Lowering the threshold produces more observations but can introduce severe model bias. Threshold stability plots, profile likelihoods, confidence or posterior intervals for , and alternative severity families are necessary diagnostics. An estimated from three exceptional claims should be described as an assumption-sensitive estimate, even if the optimizer returns many decimal places.
A reproducible synthetic tail calculation
Consider a wholly hypothetical ground-up loss severity distribution for qualifying aviation claims, expressed in millions of U.S. dollars at one common evaluation date. Suppose the probability that a qualifying claim exceeds million has been specified separately, and suppose that conditional excesses above follow a GPD with million and . These values are not fitted to any actual aviation portfolio. The assumed conditional probability that a claim greater than million also exceeds million is
The expected excess above million, conditional on exceeding that threshold, is $10/(1-0.25)=13.333$ million, so the conditional expected ground-up amount equals million. These computations describe the assumed model; they provide no empirical aviation severity statistic.
Now consider an illustrative insurance or reinsurance payment equal to the part of one such loss between million and million. Writing this -million limit excess of million as , its expectation conditional on is the area under the conditional survivor curve:
Substituting the stated GPD parameters yields
The result is conditional on a claim already exceeding million. A full expected layer cost per qualifying claim additionally multiplies by ; an annual technical loss cost further requires the frequency of qualifying claims and the exact aggregation and contract rules. Capital costs, expenses, insurer profitability provisions, and reinstatement premiums also remain outside this conditional integral. Presenting million as an actual premium or an unconditional average would be an actuarial category error.
The calculation shows a second important property. Even if an upper tail extends indefinitely at the ground-up level, the fixed -million layer payment cannot exceed million per qualifying claim. The expected layer amount depends mainly on probabilities around its attachment and exhaustion points, while an uncapped event-severity mean remains sensitive to much more distant hypothetical outcomes. The appropriate estimation target must therefore be matched to the decision: attachment probability, expected recovery, exhaustion probability, aggregate annual layer loss, or ultimate ground-up liability.
Tail dependence, accumulation, and event-level aggregation
The classical aggregate representation is useful only when the joint structure of the is understood. The same ground occurrence can damage several insured aircraft and injure multiple persons. A common maintenance error, aircraft component defect, airport infrastructure disruption, or airspace event can link otherwise dispersed risks. A portfolio may have a low marginal frequency per aircraft and still suffer a high annual event loss if the co-affected units have substantial insured values. Independent severities would materially understate the probability of such aggregate outcomes.
An event-based specification first generates event characteristics : location, time, physical mechanism, flight or ground status, affected operators, aircraft and persons, and applicable legal circumstances. It then maps an event to multiple policy responses and aggregates them at the correct occurrence or treaty level. The same claimant must not be duplicated as separate economic losses merely because more than one policy or market participant carries part of the obligation. Conversely, genuine multiple interests and cross-policy payments require their own allocation analysis. Aggregation is a legal as well as a probabilistic operation.
The Actuarial Standards Board's catastrophe modeling guidance supplies general expectations for understanding catastrophe model components, applicability, validation, documentation, and reliance on experts. Aviation accumulation scenarios need comparable governance even when the underlying peril differs from the familiar natural-catastrophe applications of that standard. Scenario assumptions should identify physical and legal common causes, exposure concentration, plausible protective controls, and the potential failure of historical independence assumptions.
Claims development and reserving interfaces
Observed paid claims are censored by time as well as by cover. A case reserve may include an adjuster's best estimate of unsettled liability but still miss future legal expenses, reopened issues, and additional claimant emergence. The reserve for incurred but not reported claims is especially difficult when an event is known but the full claim structure and ultimate allocation are not. Large aviation claims often warrant event-specific scenario reserving coupled with a separate frequency and severity model for future unreported events.
Mack's chain-ladder uncertainty analysis derives a distribution-free standard error under specified chain-ladder assumptions; England and Verrall's stochastic claims reserving review examines a broader stochastic and Bayesian reserving toolkit. Neither solves the central problem of a triangle with very few, nonexchangeable, exceptionally large claims. One event can determine an entire development factor, and the presence of new liability regimes can change settlement patterns. In such portfolios, reserve uncertainty should be decomposed into the ultimate values of identified events, unreported event counts, coverage and litigation uncertainty, and tail assumptions, rather than attributed entirely to a single fitted triangle parameter.
Reserve development must also be reconciled with prospective severity estimation. A claim appearing at million incurred and later settling for million contributes one occurrence, one ultimate severity, and a development path; it is not two separate claims. Every retrospective severity dataset needs a valuation basis and record linkage capable of connecting versions of the same event. The quality of those links can dominate the apparent statistical credibility of a loss portfolio.
Expert Calibration: Bayesian Credibility and Structured Judgment
Credibility begins with relevance
A sparse aviation portfolio is rarely devoid of information. Relevant evidence may include its operating exposure, component-specific maintenance history, known near misses, recent ground-damage incidents, engineering and safety controls, worldwide accident experience, comparable insured losses, treaty wordings, claims-adjuster assessments, and the limits of the reinsurance program. The difficulty lies in transporting those observations to a specific insurance estimand without assuming that unlike populations are exchangeable.
The Actuarial Standards Board's credibility procedures standard discusses the predictive value of subject experience and the selection of relevant external experience. Its practical implication for aviation loss estimation is that information weighting requires a documented basis. Ten thousand operational records do not make an excess-liability severity model fully credible when they contain no loss amount, policy attachment, or litigation outcome. Conversely, several well-documented insured losses may carry substantial information about claim development and wording-specific adjustment while providing little information about a portfolio-wide frequency.
A weighted average of internal and external estimates can be useful when both estimates target the same quantity. In a conventional credibility expression, , the weight should reflect predictive uncertainty and relevance rather than a ceremonial claim count. External could refer to a market book with different aircraft, geography, insurance limits, or loss reporting. A credibility factor calibrated on an unrelated motor or property portfolio would provide a misleading appearance of statistical structure. The actuarial judgment lies in selecting comparable information and expressing the resulting uncertainty.
Hierarchical estimation and the limits of pooling
A hierarchical model can represent related airline, fleet, or risk segments while allowing different baseline rates and severity distributions. For example, let be the count of qualifying losses for class observed over a compatible exposure measure , and suppose
The covariates can record exposure-relevant operating conditions; a partially pooled effect can allow classes to depart from the overall market tendency. The statistical benefit comes from regularization: a class with no recorded event still receives a posterior distribution informed by the overall pattern, while a class with relevant evidence can modify that distribution. Partial pooling assumes a meaningful shared population. General aviation, business jets, large scheduled passenger airlines, and specialized rotorcraft may require separate levels, explicit covariates, or separate models rather than forced membership in one common group.
For severity, a related hierarchy may place prior distributions on log scale and tail shape, using observed ultimate claims only when their coverage and development bases can be reconciled. In very small portfolios, flexible multilevel models can be weakly identified: apparent heterogeneity may reflect true different risk classes, hidden changes in reporting, or a single abnormal claim. Reporting posterior means without prior-sensitivity ranges would hide this ambiguity. A large external sample improves estimation only to the extent that its statistical relationship to the target risk is justified.
A zero-event gamma-Poisson calculation
Consider a completely synthetic high-layer event process. Let be the unknown expected number of qualifying covered events per one million comparable exposure units. For purposes of illustration, set a gamma prior in shape-rate parameterization, , so the prior mean is event per million units. Suppose one million units of matching exposure produce no qualifying events. Under a Poisson likelihood with measured in millions, conjugacy gives
The zero observation reduces the prior expected frequency but does not make the estimated rate vanish. For one additional million comparable exposure units, the full posterior predictive probability of at least one event is
A plug-in Poisson calculation using the posterior mean alone gives $1-e^{-2/3}\approx0.4866$. The difference is a direct effect of mixing over rate uncertainty: the posterior predictive distribution contains additional dispersion, and the convexity of the zero-count likelihood changes event probabilities. Both numbers are specific to the stated hypothetical prior, exposure, and qualifying event definition. Neither is a real aviation rate. Under an alternative prior with a different mean or concentration, the posterior and predictive probabilities change. A complete report would show those alternative assumptions and, where possible, calibration to compatible external experience.
The calculation also distinguishes event probabilities from claim severity. Even an accurately estimated chance of one loss above a treaty attachment leaves uncertain the size of that loss, the number of payments from one event, and the treaty recovery. In the simplest independent compound-Poisson representation, expected aggregate loss equals only when the event count and severities are modeled consistently and the relevant expectations exist. Common-cause risk and dynamic coverage changes require richer representations.
Eliciting expert judgments without concealing assumptions
Where upper-layer payments have never been observed, relevant expertise can include claims counsel, airline operational specialists, maintenance and airworthiness engineers, hull adjusters, aviation underwriters, brokers, and reinsurance actuaries. Each expert observes only part of the causal chain. A safety engineer may assess probabilities of technical failures but have little direct information about legal settlement amounts. A claims lawyer may know the range of potential damages but lack denominators for operational event frequency. Their judgments should therefore be elicited against narrowly defined quantities.
An elicitation exercise can ask, for example, for the probability of at least one covered -million net retained loss over ten specified years of fleet exposure; the 50th, 90th, and 99th percentiles of ultimate bodily-injury settlement amounts conditional on an eligible event; or the plausible range of a claim's reporting delay under a named jurisdiction. Experts should receive a precise event and exposure definition, the contractual assumptions, the temporal basis, and any public comparator evidence. Elicitation may then proceed through separate judgments, structured discussion of evidence, and reconciliation of divergent assumptions. Persistently divergent judgments may be retained as model alternatives rather than compressed into an unjustified average.
Suppose expert group A assigns an assumed annual large-loss frequency around , while group B assigns because it anticipates a wider coverage trigger and higher litigation costs. The difference may be primarily definitional rather than subjective. Before combining the numbers, analysts should harmonize the covered event, the exposure, the threshold, and the policy basis. Only after that should they consider weights or a mixture distribution. A probability mixture explicitly preserves competing views; it also tends to have a wider predictive distribution than an arbitrary single-parameter compromise. Weights require documented reliability grounds. This is a method of representing uncertainty, not an assertion that any particular set of experts is empirically calibrated.
The Actuarial Standards Board's modeling standard addresses assumptions, reliance on experts, validation, communication, and model risk. Its general requirements apply especially strongly when a numerical result could otherwise be mistaken for a fully data-driven estimate. Expert provenance should record the question, available material, expert role, independence, model version, alternative responses, and the reason for any adjustment. Correlated expert opinions drawing on the same accident record are not multiple independent samples.
Prior predictive and posterior predictive checks
A prior should be tested against observable consequences before the target data are analyzed. If a proposed prior implies a substantial chance of frequent multiple catastrophic airline losses despite decades of compatible operations, the discrepancy needs an explicit explanation. Equally, a prior that makes severe scenarios virtually impossible may encode the modeler's comfort rather than physical and legal evidence. In an upper layer with no historical attachments, prior predictive checks can examine the implied frequencies of all-layer events, near-attachment claims, total-limit exhaustion, and clustered losses.
After fitting, posterior predictive simulations should compare the observed count distribution, body severities, reporting delays, tail exceedance patterns, and event clustering with replicated data. A model can fit the observed zero high-layer claims while still performing poorly for small and medium losses. Observed zero events cannot validate a particular GPD shape or prove a chosen cross-aircraft dependence structure. Predictive assessment should use the closest available relevant phenomena and record which assertions remain assumptions because direct observations are absent.
Decision-making with a distribution of plausible loss models
An operational decision may involve selecting a deductible, allocating reinsurance, determining a claims reserve, or establishing a risk appetite. A posterior expected loss is one input, but decisions also depend on the uncertainty of the expected value, the severity tail, aggregate concentration, counterparty risk, and the consequences of estimation error. For example, the same expected annual ceded layer cost may arise from relatively frequent moderate losses or very rare near-total limits. These alternatives have different implications for liquidity, contract exhaustion, reinstatement, and capital.
When evidence is sparse, risk reporting should show the expected loss across plausible models alongside stress outcomes, attachment probabilities, and scenario losses. It should identify which uncertainties are reduced by future observation and which are structural. Better claims data may narrow uncertainty about settlement development. Better legal wording may narrow the coverage transformation. Neither can by itself eliminate uncertainty about a historically unobserved correlated fleet catastrophe.
Evidence Appraisal and Sensitivity
A hierarchy of evidence for the insured-loss question
The strongest source for a historical covered payment is normally the original claim record supported by policy wording, underwriting and exposure files, settlement documents, and valuation information, subject to confidentiality and applicable law. Such source material is usually unavailable in open aviation datasets. An accident report establishes aspects of the event, its operational setting, or the safety-investigation findings; a published annual safety review supplies defined counts and denominators; treaty texts and domestic regulations describe particular legal responsibilities; actuarial research informs the mathematical method. Each evidentiary class has its own competence and blind spots.
The National Transportation Safety Board's aviation data resources supply public accident and incident information under defined collection rules. Their public nature and incident-level detail make them valuable for event classification, historical reconstruction, and possible exposure-relevant covariates. They do not establish the policy terms or settlement amounts of private insurers. Similarly, the ICAO safety reporting program concerns safety monitoring; it is not a worldwide census of insured payouts. This distinction governs both statistical comparability and permissible conclusions.
The Actuarial Standards Board's data-quality standard addresses the suitability, review, and limitations of data used in actuarial services. Applied here, its lesson is concrete: source provenance must be recorded at the level of each field. For example, an accident severity label from a public safety record, an estimated aircraft value from a market source, and a settlement amount from a broker-reported anecdote possess different evidentiary status. Their merger into one row does not give them common statistical reliability. A reproducible research ledger should capture source authority, reference date, event identifier, coverage basis, observation window, currency, measurement definition, modifications, and residual uncertainty.
Missingness in aviation claims is often informative. Large settlements may be confidential; high layers may see only exceptional losses; insurers may retain different claim-reporting thresholds; and litigation can delay final valuation. Treating missing values as missing completely at random would therefore be difficult to defend. A defensible model specifies whether observation depends on amount, exposure, jurisdiction, or reporting age and whether the dependence can be estimated from available data. If it cannot, alternate selection mechanisms should enter the sensitivity analysis.
Exposure and contract standardization as measurement tasks
Before pooling severity records, the researcher needs an exposure and coverage dictionary. It should identify the aircraft or insured interest, operator and route profile, insurance class, currency, sum insured, policy period, attachment and limit, occurrence definition, and whether the amount is ground-up, insured gross, or retained net. Multiple aircraft policies with different limits are not exchangeable severity observations simply because they relate to the same accident category. The available exposure needs to be linked to the insured obligation, not merely to a convenient industry statistic.
For older losses the analyst should save original nominal figures and separately compute standardized amounts. Currency translation should identify both the source and target currencies and the relevant valuation date. Inflation adjustments should distinguish legal claims inflation, aircraft acquisition or repair costs, and general price trends. The analyst must retain the original sum insured and attachment structure because a layer with a historical -million attachment is economically different from a contemporary layer with the same nominal attachment. Some historical contract terms may be impossible to reconstruct; those records may be eligible only for broad scenario support.
A useful distinction is between a historical normalization and a counterfactual restatement. A normalization translates the same recorded obligation into a common reporting basis; a counterfactual asks how a past event would have been paid under a present contract or jurisdiction. The latter requires assumptions about legal outcomes and coverage, and therefore belongs in a modeled scenario with uncertainty. Conflating them can produce falsely smooth long-run severity trends.
Capital measures and aviation stress cases
The European Solvency II framework supplies an example of a prescriptive aviation stress within a larger regulatory capital architecture. Article 131 of Delegated Regulation 2015/35 specifies the aviation risk sub-module using the maximum, over relevant insured aircraft, of specified net sums insured; its text also addresses when reinsurance recoverables can be deducted and when such deduction would insufficiently capture the exposure. This is a regulatory stress construct, not an empirically fitted accident-severity probability distribution. It illustrates how supervisors can require deterministic stress calculations even when the private claim tail is not directly identified.
A pricing estimate, an accounting reserve, and a regulatory capital requirement can therefore differ substantially while remaining internally coherent because they answer different questions. Expected annual loss is a probability-weighted cost on a specified basis. An ultimate claims reserve concerns recognized events and future development. A prescribed stress concerns a defined adverse condition and its impact on the insurer's financial position. The data and model can be shared, but each result requires its own interpretation, assumptions, and validation.
An explicit sensitivity audit
For a real application, the minimum sensitivity exercise should separately vary event frequency, attachment probability, GPD or alternate-tail parameters, historical normalization, settlement development, insurance wording, and accumulation. The purpose is to discover which uncertainties change the decision, not merely to produce many Monte Carlo iterations. Increasing simulation runs narrows computational error around the selected assumptions; it does not narrow ignorance about the appropriateness of those assumptions.
Consider again the synthetic GPD model for losses exceeding million. Keep the threshold and scale at , million while varying shape across , , and . Conditional survival to million, derived directly from the GPD formula, becomes respectively , $2^{-4}=0.0625, and \3^{-2}\approx0.1111$. Even this small illustrative selection of shape values changes the probability of the remote event by more than threefold. None of these choices can be identified from the synthetic exercise itself. The result demonstrates why a single fitted tail index should not control all capital and treaty decisions without external support and model-risk disclosure.
The calculation also reveals why sensitivity should be tied to the attachment layer. A change in tail shape affects extremely high unlimited loss measures more strongly than some bounded layers. One can compare the expected -million excess of -million payment under each shape while holding the threshold exceedance probability and contractual mapping fixed. Alternatively, one can vary the threshold exceedance frequency simultaneously. The latter test may be more realistic but must disclose that it changes two assumptions rather than isolating one parameter.
In a broader stress exercise, record at least the following assumptions in a reproducible table: the contract and calendar date; exposure denominator and forecast; event taxonomy; reporting and development basis; external-data relevance; body and tail distribution; threshold and number of exceedances; prior and its justification; assumed dependence; currency and inflation bases; treaty exhaustion or reinstatements; and uncertainty ranges. Each stress result should be accompanied by the probability or scenario interpretation that generated it. An estimated return period from a highly extrapolated tail model deserves a range and an explicit extrapolation warning rather than an unsupported exact recurrence date.
Validation when the decisive event has not happened
Classical backtesting is constrained when the target layer has no historical paid loss. A correct zero-event prediction can arise from many mutually incompatible models. Model validation should therefore be distributed across components: operational event frequencies; smaller and near-threshold loss counts; expert probability forecasts; exposure reconciliations; development of known claims; legal and coverage transformations; and event-dependence scenarios. The model can be rejected or revised when these components fail even if no catastrophic claim has yet materialized.
Prospective monitoring has value because future observations can be more informative than a retrospective sample whose reporting and coverage have changed. Underwriting and claims teams can agree on a minimum event data schema now, preserve source documents and event IDs, maintain development snapshots, and record both rejected and accepted claims. Such a design produces a future evidence base that can distinguish physical occurrence rates from insured attachment rates and coverage disputes.
The Actuarial Standards Board's future-cost estimation standard specifically addresses data, exposure bases, infrequent events, catastrophe treatment, reinsurance, and future-cost assumptions in applicable U.S. actuarial practice. Its sections support the wider methodological conclusion that rare losses are evaluated with attention to future exposure and risk transfer, rather than by mechanical extrapolation from a short claims history. The discussion here is comparative research and does not purport to issue a jurisdiction-specific actuarial opinion.
Epistemic limits and interpretive controls
No public accident count identifies an insurer's gross or net loss without policy and claims information. No observed zero identifies the upper-tail parameter. No professional standard supplies the missing empirical aviation claim distribution. No mathematical model can remove ambiguity from undetermined coverage or future legal interpretation. These are identifiable boundaries of the evidence, rather than technical problems that can be solved by a larger regression or more elaborate computational machinery.
The present research uses official public materials and specialist scholarly literature where their content or publisher abstracts are available. It does not claim access to confidential insurer claims, full paywalled actuarial texts beyond the verified available record, or an independently reconstructed global paid-claims dataset. The synthetic examples were constructed to make inference auditable and to expose uncertainty. They do not constitute a new statistical finding about accident rates, loss costs, premiums, solvency capital, or the insurance performance of a particular airline or territory.
Scientific Synthesis and Research Questions
Integrated findings
The investigation yields six connected analytical findings. First, the data-generating process relevant to aviation insured loss passes through several distinct gates: physical event, public classification, insurable interest, legal responsibility, contractual attachment, reporting, and ultimate settlement. At each gate an observed sample can change. Sparse data estimation fails when those gates remain implicit, because one may inadvertently fit a distribution to safety records while reporting conclusions about indemnity payments.
Second, scarcity has multiple meanings. A lack of severe events within adequate exposure is evidence about a particular event frequency under a particular model. A lack of disclosed reinsurance recoveries is often evidence about reporting and commercial access. A lack of developed ultimate amounts reflects valuation maturity. Each demands a different research strategy. The mere number of records is consequently a poor measure of their scientific value for an actuarial estimand.
Third, severity-tail estimates are decision-dependent. A distribution that broadly fits ordinary hull repair claims may offer weak predictions for extreme passenger liability. A conditional upper-tail model requires an independently established exceedance frequency. A bounded reinsurance layer can have finite expected cost even when an idealized uncapped ground-up distribution has problematic moments. The -million conditional layer illustration demonstrates how precise mathematical quantities can be calculated once assumptions are stated without conferring empirical validity on those assumptions.
Fourth, partial pooling and expert calibration are indispensable in selected sparse-data settings but must be justified by transportability. A prior derived from genuinely comparable experience can regularize a zero-event portfolio. An unrelated market average can import bias. The gamma-Poisson example makes the implication concrete: zero recorded claims modify the posterior, while the forecast remains prior-dependent. Calibration quality rests on the pertinence of the prior and the explicit exposure unit rather than the elegance of the conjugate formula.
Fifth, legal and contractual state variables belong inside the severity model. Aviation liability rules, their temporal revisions, insurance minimums, and particular policies determine the amount potentially paid. Treating legal regimes merely as explanatory footnotes can generate an incorrectly mixed sample. The distinct roles of the Montreal liability framework, regulatory insurance minima, and a negotiated insurance payment function provide a verifiable example of this structural issue.
Sixth, model uncertainty should be decomposed instead of hidden in a single confidence interval. Sampling uncertainty concerns finite observations under an accepted model. Parameter uncertainty concerns the value of model coefficients. Structural uncertainty concerns model families, dependence, selection, exposure definitions, and contractual transformations. Legal and operational scenario uncertainty may also persist after excellent statistical estimation. Decision-makers need a transparent representation of which uncertainty source dominates each outcome.
Implications for research design
A rigorous future empirical aviation claims study would define a fixed observation population, contract basis, and evaluation date before extracting data. It would retain occurrence identifiers capable of joining accident, underwriting, adjuster, legal, and payment records, while protecting confidentiality. It would record both exposed units and claim reporting thresholds. It would separate ground-up economic costs from gross and net insurance payments, distinguish occurrences from claim files, and preserve the evolution from reported to ultimate values. The study would use public safety information to enrich event classification, while relying on appropriately documented financial records for insurance severity.
Model comparison should examine at least a parsimonious body distribution, a tail-specific distribution, and a conservative scenario approach for extreme layers. The analysis should use rolling or other temporally appropriate validation where data permit, and should identify low-information segments that depend on priors or expert judgments. Rather than removing difficult points as statistical outliers, researchers should investigate whether they represent data errors, unusual but insured events, a different coverage regime, or a genuinely relevant tail mechanism.
Data cooperation presents a significant institutional research opportunity. Shared, anonymized event-level insured loss information across insurers and reinsurers could improve the estimation of extreme claims, provided that policy terms, normalization, reporting thresholds, and event aggregation remain traceable. Pooling without these metadata might simply create a larger but less interpretable database. Governance arrangements must also manage confidentiality, antitrust obligations, data rights, and access controls under the jurisdictions involved.
Research questions remaining open
Further research should determine how much predictive information public safety occurrence data add to covered high-layer severity after controlling for fleet characteristics and insurance-contract selection. It should examine which hierarchical transport assumptions are defensible between commercial, business, and general aviation. It should test how different prior elicitation protocols affect tail-sensitive forecasts when genuine high-excess claims are absent. It should quantify the contribution of liability regime changes, claims inflation, and delayed settlement to apparent tail thickening. It should investigate dependence mechanisms linking aircraft and liability claims within an occurrence, and the effect of those mechanisms on reinsurance recoveries and capital. Finally, it should evaluate which data-sharing structures deliver real gains in tail identification while preserving the provenance of sensitive insurance records.
These questions follow from the documented limitations of existing source types and standard statistical methods. They specify empirical and methodological programs for subsequent work; they do not constitute a newly named doctrine, a new actuarial theory, or a claim of discovered aviation market parameters.
Conclusion
Aviation loss severity is an actuarial object constructed from a physical event, an insured interest, a legal obligation, a contract, an exposure period, and an ultimate claims valuation. Sparse observations of exceptional events offer genuine information about risk only when those elements have been made commensurable. The comparison of contemporary ICAO and IATA accident reporting, the separation of public safety and private claims evidence, the dated Montreal liability revisions, and the formal treatment of attachment layers all show why statistical estimation and documentary interpretation belong together.
Extreme-value methods characterize plausible upper tails under stated conditions; credibility and Bayesian methods express how limited portfolio experience interacts with relevant external evidence; event-based approaches preserve accumulation and dependence; and reserving methods account for the time between occurrence, notification, and ultimate payment. Each method answers a specified question and introduces assumptions that must be inspected. The synthetic calculations demonstrate the central quantitative lessons: no observed rare events does not imply no future risk, and a conditional reinsurance-layer expectation cannot be interpreted as an unconditional portfolio loss cost.
The scientific conclusion is a methodological one. The credibility of an aviation severity estimate depends on the alignment of event definitions, exposure denominators, contractual payment functions, historical valuation, external-source transportability, and uncertainty diagnostics. Transparent choices at these interfaces make the estimate auditable and useful for further empirical investigation. The present study contributes an integrated critical synthesis and independently derivable examples within the established literature, with no claim of a new theory, proprietary dataset, or external peer-review certification.
References
Actuarial Standards Board. (2013). Actuarial Standard of Practice No. 25: Credibility Procedures* (effective May 1, 2014). https://www.actuarialstandardsboard.org/asops/credibility-procedures/
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality* (effective April 30, 2017). https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention* (effective August 1, 2018). https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling* (effective October 1, 2020). https://www.actuarialstandardsboard.org/asops/modeling-3/
Actuarial Standards Board. (2021). Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas)* (revised; effective December 1, 2021). https://www.actuarialstandardsboard.org/asops/actuarial-standard-of-practice-no-38-revised-edition/
Balkema, A. A., & de Haan, L. (1974). Residual life time at great age. The Annals of Probability, 2(5), 792–804. https://doi.org/10.1214/aop/1176996548
Canadian Transportation Agency. (2024). Notification to air carriers of upward revision of the limits of liability governed by the Montreal Convention (revisions effective December 28, 2024). https://otc-cta.gc.ca/eng/content/notification-air-carriers-upward-revision-limits-liability-governed-montreal-convention
England, P. D., & Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. https://doi.org/10.1017/S1357321700003809
European Parliament & Council of the European Union. (2004). Regulation (EC) No 785/2004 on insurance requirements for air carriers and aircraft operators (consolidated applicability and amendments to be checked at use date). https://eur-lex.europa.eu/eli/reg/2004/785/oj/eng
European Commission. (2015). Commission Delegated Regulation (EU) 2015/35, Article 131, Aviation risk sub-module (EIOPA rulebook). https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-5750_en?prefLang=ga
International Air Transport Association. (2026, March 9). IATA releases 2025 safety report. https://www.iata.org/en/pressroom/2026-releases/2026-03-09-01/
International Air Transport Association. (n.d.). Montreal Convention 1999. https://www.iata.org/en/programs/passenger/mc99/
International Civil Aviation Organization. (2026, July). International Civil Aviation Organization analysis reveals major priorities for air transport safety. https://www.icao.int/news/international-civil-aviation-organization-analysis-reveals-major-priorities-air-transport
International Civil Aviation Organization. (n.d.). ICAO API Data Service 2.0 (service description; data access requires authorization). https://dataservices.icao.int/
International Civil Aviation Organization. (n.d.). Safety reports. https://www.icao.int/safety-reports
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://doi.org/10.2143/AST.23.2.2005092
McNeil, A. J. (1997). Estimating the tails of loss severity distributions using extreme value theory. ASTIN Bulletin, 27(1), 117–137. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/estimating-the-tails-of-loss-severity-distributions-using-extreme-value-theory/745048C645338C953C522CB29C7D0FF9
National Transportation Safety Board. (n.d.). Aviation accident database and statistics. https://www.ntsb.gov/safety/data/pages/Data_Stats.aspx
United Kingdom Civil Aviation Authority. (n.d.). Insurance requirements and guidance. https://www.caa.co.uk/commercial-industry/airlines/licensing/requirements-and-guidance/insurance/
United States, Code of Federal Regulations. (n.d.). 49 CFR § 830.2, Definitions (current eCFR text as republished by Cornell Legal Information Institute). https://www.law.cornell.edu/cfr/text/49/830.2