RESEARCH DOSSIER
Marine Hull Frequency Severity and Inflation Models
Abstract. Marine hull insurance represents a physical asset exposure converted into contractual financial indemnity. The scientific measurement problem spans vessel time at risk, casualty occurrence, claims reporting, damage valuation, policy adjustment, repair-cost inflation, exchange-rate movement, and ultimate claim development. This monograph develops a source-critical international analysis of marine hull frequency, severity, inflation, and their joint influence on prospective loss cost. It integrates marine insurance law and policy wordings, generalized linear and compound aggregate loss models, conditional severity and heavy-tail assumptions, loss development and currency-adjusted repair economics. New 2026 data releases from IUMI, Cefor and EMSA are evaluated according to their different observational populations and definitions. Independently specified synthetic calculations demonstrate the consequences of frequency–severity covariance, deductible threshold migration, cost-basket inflation and foreign-exchange exposure. The analysis provides a coherent empirical measurement and validation program while preserving distinctions among observed incidents, paid insurance claims and industry-level financial statistics. It develops original synthesis and numerical counterexamples from established mathematical and legal principles. The research makes no claim of an empirically fitted worldwide marine hull model, an original named theory, access to confidential claims data or completed external peer review.
Keywords: Marine hull insurance; hull and machinery; insured vessel-year; exposure; casualty frequency; claim reporting; generalized linear models; negative binomial regression; conditional severity; deductible; claims inflation; repair cost; machinery damage; shipyard capacity; exchange rates; compound Poisson loss; extreme value risk; claims reserving; underwriting; stochastic loss cost; model validation; marine insurance law; IUMI; Cefor; EMSA.
Introduction
Statistical pricing of marine hull losses begins with a question of measurement. A ship is a moving capital asset, operates across technical and legal jurisdictions, and can be covered by several insuring interests and contract layers. An annual insurance claim count is not necessarily a count of maritime incidents. A paid loss amount is not necessarily the physical repair cost of the casualty that generated it. An average claims trend can change because the fleet grows, vessels age, deductibles change, claim recording accelerates, or shipyard prices rise. Each of these pathways requires a separate empirical definition.
The IUMI Stats Report 2026 reports USD 42.6 billion in global marine insurance premiums for 2025 and USD 10.5 billion in ocean hull premiums, with the hull amount up 9.4% nominally from 2024. The report connects the change to fleet growth, asset values, war premiums, competition and foreign-exchange effects. This is a sector-wide financial observation, not a direct measurement of hull claim frequency, repair inflation or risk-adjusted price per insured vessel-year.
An important contrast appears in the Cefor 2026 hull statistics: frequency in the USD 10,000–500,000 claims band continued to increase, while frequency above USD 500,000 flattened at an elevated level. The evidence concerns the specific participating NoMIS market and reported size categories. Meanwhile the EMSA 2026 casualty overview concerns marine incidents reported to a European safety investigation data system. These measurements differ in both numerator and denominator. Treating them as interchangeable observations of one global hull risk would contaminate the inference.
The central research questions concern the construction of insured hull exposure, the distinction between a physical occurrence and an insurer's recorded claim, the distribution of contract-adjusted repair losses, the attribution of nominal claims trends to inflation, and the robustness of technical premium indications. The method is legal and contractual source comparison, actuarial derivation, institutional statistical evidence appraisal, synthetic numerical demonstration, and explicit scientific-editorial quality control. The study covers international hull and machinery insurance and its interfaces with related marine coverages. The mathematics uses stated assumptions rather than invented insurer datasets.
Conceptual and Historical Foundations
Hull interest, peril and financial indemnity
The vessel, the policyholder, the insured interest, the operator and the risk carrier occupy different positions in the insurance arrangement. The vessel is the physical unit. The operator determines much of the voyage and maintenance profile. The insured interest and policy establish which financial consequences of damage pass to the carrier. Some vessels enter fleets in multiple placements or with distinct additional-value, war or machinery-related arrangements. Records for the same physical hull cannot be counted as independent insured physical vessels merely because multiple insurers appear.
The United Kingdom's Marine Insurance Act 1906 illustrates the distinction between insured peril and indemnity. Section 55 concerns proximate causation by a peril insured against. Sections 67–69 address the measure of indemnity for loss and, in section 69, partial loss of a ship, including the reasonable cost of repairs subject to policy provisions and applicable limits. This is original statutory authority within its legal scope; its effect in a particular dispute depends on applicable governing law, current amendments and policy terms. It establishes that the cost of a physical casualty and the amount ultimately payable by an insurer are conceptually different.
A separate original contractual framework is supplied by the Nordic Marine Insurance Plan, whose sections specify insurable values, marine and war perils, causation, duties, coverage periods and hull adjustment. The Nordic Plan is a policy framework where contractually adopted, whereas the British Act is legislation. They can represent different settlement structures for apparently similar physical damage. This distinction matters to statistical modeling because the contract acts as a transformation between the damage event and the measured claim payment.
Historical comparison of marine losses must also distinguish hull cover from cargo insurance, protection-and-indemnity liabilities, offshore operations and charterer's liability. A grounding can generate vessel repair cost, cargo damage, navigation liabilities, salvage expense and crew claims. These are correlated consequences of one casualty but involve different insuring interests, limits, counterparties and data-generating rules. A hull claim severity response should capture the hull coverage selected for the study and treat other consequences as related risk or recoveries with documented definitions.
Event, file, payment and accounting year
The initial unit is a defined physical occurrence on a particular vessel at a particular time. The event may be notified after a delay, accepted as a registered claim, surveyed in stages, and eventually settled through multiple transactions. Claim accounting may include paid loss, current outstanding case reserve, incurred total, adjustment expenses, subrogation recoveries, deductibles and reinsurance shares. Event count, number of claims files and number of payment lines must retain distinct identifiers.
A basic causal sequence is: physical event; report; coverage and liability adjudication; inspection and damage estimation; repair or total-loss decision; claim reserve; expenditure and recovery; ultimate settlement. The sequence is crucial for inflation: a loss occurring in one year can be repaired in another and paid in a third. If a study uses the payment date as a surrogate for casualty date, changes in reporting and settlement speed will distort the apparent frequency or inflation trend.
The IUMI marine reporting-scope guidance documents differing premium-year and loss-year conventions. A nominal loss ratio assembled from aggregated accounting-year premiums and underwriting-year loss triangles cannot by itself identify an expected claim count, the severity of a fixed vessel population or pure inflation. Its interpretive purpose is market monitoring rather than complete actuarial reconstruction.
Actuarial foundations
Foundational statistical methods are provided by Renshaw's covariate claims analysis, by Ohlsson and Johansson's GLM text, and by de Jong and Heller's insurance models. Their work supports conditional count and positive-severity regression with exposure adjustments and uncertainty assessment. They do not furnish a universal coefficient for marine hull events. Actuarial models become marine-hull models only when the observation units, vessel characteristics, policy provisions and geographic claim-cost processes have been specified and tested.
Hull Exposure Units and Frequency
The denominator precedes the count likelihood
Let represent the fraction of year during which vessel has the specified hull cover, measured under an explicit day-count convention. Full-year insured service contributes approximately one vessel-year. A fleet with ten full-year ships and ten half-year ships contributes about fifteen vessel-years, not twenty. Entries, departures, cancellations, restricted cover, vessel replacement and participation shares require distinct treatment. Counting fleet membership only at year-end can bias the denominator when the fleet changes during the year.
For a general annual hull peril, vessel-years provide a defensible starting exposure base. For particular causes, risk days underway, port calls, machinery hours, docking episodes and passages through a defined hazard region may improve explanation. These operational measures do not automatically replace insured risk time: an insured vessel in port may face fire or contact damage, while a vessel laid up can face different machinery and weather hazards. The correct exposure must match the event to be counted and the contract under which it is observed.
A reproducible study needs unique vessel ID, claim ID and occurrence ID, vessel type and size, construction and machinery attributes, operator and class history, policy dates, deductible, insured value, accident and reporting dates, peril and cause code, all reserve snapshots, payment and recovery transactions, repair geography and payment currency. It must record changes in data extraction and vessel identification. Claims subscribed by co-insurers should be counted as one physical casualty for occurrence-frequency analysis, while each insurer's net share remains relevant for loss cost and capital.
Frequency regression and heterogeneity
For claim events and vessel-year exposure , a baseline count model is
The exposure offset implies expected claim count scales proportionally with at-risk time after conditioning on the risk factors. This restriction should be tested for machinery claims when sea-time intensity or docking changes within the period. The conditional Poisson zero probability is . At , it is about 0.90484, so even a perfectly ordinary count process can produce more than 90% zero-claim observations.
Hull claim variability can exceed the Poisson mean because of fleet heterogeneity, repeat incidents, operator behavior, common weather and correlated machinery. Gamma-mixed Poisson intensities yield a negative-binomial family; a common parameterization has . Vessel or operator random effects and cause-specific hierarchical models can represent repeated risk. These general methods derive from established insurance regression literature and require stable exposure definitions before a distribution choice becomes interpretable.
A zero-inflated Poisson construction supplies another possible distribution when extra zero outcomes remain after conditioning. Its fitted mixture probability should not automatically be interpreted as a proportion of perfectly safe ships. In particular, policyholders may choose not to report an accident, small losses can remain below deductible and incomplete claims registration can produce observed zero files. A mixture fitted to observed counts cannot uniquely identify those physical mechanisms.
From accidents to recorded insurance claims
Define as the underlying count of physical accidents within the relevant peril definition, and let if accident becomes a recognized claim in the available data. Then . Under a homogeneous Poisson accident rate and independent acceptance probability , observed claim count has Poisson mean . Counts alone identify the product , not the separate physical accident rate and reporting probability. A deductible change may reduce while leaving the accident hazard fixed.
Claims reporting thresholds can also change the distribution of severity bands. If cost inflation multiplies base repair cost by , a fixed threshold corresponds to base real cost . Higher nominal repair prices may move a fixed physical event into a higher reporting band even when no additional incident occurs. Cefor's mid-size and high-value count distinction should therefore be examined with real-cost bands and observation definitions before attributing all nominal movements to increased casualty incidence.
An incomplete recent accident year produces another kind of zero: events may be unreported, investigated, or reserved without final file creation. Accident-year reporting triangles, claims-lag adjustment and extraction-date freeze rules help distinguish underdevelopment from lower true frequency. Model selection should test both calibrated zero frequency and upper count quantiles on held-out cohorts, not rely on an in-sample likelihood advantage alone.
What external casualty and claims statistics measure
The EMSA 2026 overview is based on incidents reported by safety investigation authorities in the EU Member States, Iceland and Norway. It concerns a defined reporting system rather than every insured global hull. The IUMI Major Claims Database is an insurer-claims collection emphasizing selected major losses from contributors, not a complete worldwide vessel-year denominator. Cefor provides another participant-defined record of hull insurance experience. These datasets are independently informative, but rates are comparable only after differences in market scope, participation, size thresholds, event definition and extraction date are controlled.
If one study counts accident reports in European waters and another counts insurer-paid hull files around the world, a difference in annual rates cannot be interpreted as a difference in global ship safety or insurance pricing. The observational boundary must be established before any mathematical pooling, causal attribution or cross-market quotation.
Loss Severity and Contractual Adjustment
Gross damage and net insurer payout
Let be the gross economic cost of repairing or replacing the damage associated with casualty at a specified valuation date. Let contain the insuring clauses, deductible, insured value or limit, coverage conditions, co-insurance, expenses and recoveries. The claim paid by a particular insurer is a contractual transformation . Severity may refer to , to conditional admitted indemnity, to , or to the ultimate net amount of a reported claim. These are different statistical targets, and an average computed from positive-paid files is selected by the contract.
For illustration only, suppose one covered physical damage cost is USD 180,000, the deductible is USD 100,000, and the payment limit is high. Under the simplified payment rule the insurer pays USD 80,000. This operation must not be applied mechanically to the mean of a random gross damage distribution. In general, ; the complete distribution and deductible exceedance probability determine the net mean. Fixed deductibles introduce nonlinear sensitivity to repair inflation.
The statutory partial-loss indemnity provisions include reasonable cost of repair and associated limitations, subject to the wording. Total loss, partial repair, unrepaired depreciation and salvage can require distinct decisions. Agreed insured values influence a total-loss outcome differently from the invoice prices of repair materials. Cover may also distinguish machinery damage and particular excluded or restricted perils. A single raw 'claim amount' field cannot explain all these legal and commercial transformations.
Positive severity models and censored claims
Common positive-severity models include lognormal and gamma regression. For a gamma GLM with log link, and ; dispersion governs variation around the conditional mean. A lognormal model instead describes , and the fitted log mean must be retransformed with the relevant variance correction to predict the arithmetic mean. Vessel size, age, value, cause, damaged component, shipyard location, inflation period and deductible threshold can all alter the observed response.
A dataset containing only claims above threshold represents a truncated sample. For continuously distributed damage with density , its observed exceedance density is for . Claims capped at policy limit may be censored; for an observation reported as exhausted limit, the likelihood contribution can be the survival mass rather than the point density at . Confusing truncation, censoring and actual exact amount can understate the upper tail and distort pricing of high-value vessel exposures.
A flexible mixture can separate attritional machinery claims, collisions, fires, groundings and total losses by causal category. For extreme claims, generalized Pareto tail estimation or other heavy-tail methods need enough threshold exceedances, sensitivity to tail cutoffs and credible uncertainty bounds. A pooled severity estimate may fit the median while materially underestimating rare insured-value and salvage-intensive losses. Spatially concentrated weather or a common fleet operator can generate dependence among losses rather than independent draws.
Joint frequency and severity
If is an aggregate compound claim amount and count is independent of iid positive severities, then . For a Poisson count with mean , and when the second moment exists. The variance formula demonstrates why an unchanged expected claim count can coexist with substantially larger aggregate uncertainty when the severity tail grows.
When vessel hazard state affects both count intensity and mean severity, the correct expression is per vessel-year. It is not generally . A fully synthetic example clarifies the difference. Suppose 95% of vessel-years are in a low-risk condition with count intensity 0.12 and mean paid event severity USD 180,000, while 5% are in a high-risk condition with intensity 0.25 and mean severity USD 450,000. The correctly mixed expected loss is $0.95(0.12)(180000)+0.05(0.25)(450000)=USD 26,145$ per vessel-year. The vessel-weighted mean count is 0.1265 and the vessel-state mean severity is USD 193,500. Naively multiplying these latter means gives USD 24,477.75, an underestimate of USD 1,667.25. This difference reflects positive state-level covariance and is not an empirical estimate.
The Tweedie compound Poisson approach models aggregate loss with a mass at zero and a positive continuous component. It can provide a direct pure-premium response where underlying counts and positive claims are unavailable. Separate event frequency and severity models make physical and contractual drivers more transparent when detailed data permit. Neither approach substitutes for verified coverage definitions or credible data on heavy-tail claims.
Inflation Trends and Repair-Cost Translation
A hull-specific price concept
Marine hull repair prices depend on steel and components, specialized shipyard labor, dry-dock occupancy, machinery spares, transport, survey activity, local energy cost, salvage, timing and currency. Inflation in these inputs differs by repair geography and vessel design. The U.S. ship-repair producer index is an official unadjusted monthly series for nonmilitary ship repair in a defined American production sector. It is valuable within that economic scope and cannot be assumed to measure shipyard costs worldwide.
The IUMI Hull Inflation Index 2026 offers a current marine-industry attempt to account for macroeconomic repair drivers. Its official methodological notice identifies one remaining non-public input, Clarkson's Global Orderbook Forward Cover; a public spreadsheet with an alternative shipyard-capacity series; a public ten-year modeling horizon; and observed multicollinearity. The original proprietary calculations were not released. Consequently the index is a useful disclosed benchmark and research lead, while exact independent reconstruction and fleet-specific validation remain separate tasks. Quoting the public alternative as though it reproduced the unreleased original would misrepresent the evidence.
Occurrence date, estimate date and payment date
A hull event at accident date may be surveyed at , booked at a later estimate date , repaired in stages and paid at dates . Different material and labor items are purchased at different moments. An accident-year gross damage estimate and a future ultimate claims cash flow therefore need different price dates. If observed payments are compared without transforming them to a common purchasing-power basis, changes in repair duration or the distribution of open claims can imitate economic inflation.
Let base quantities of repair component be and its unit price at economic time be . An equivalent-scope target-date repair cost is . If a component index is justified, provides an explicit translation. The approximation assumes unchanged physical quantities and comparable supplier scope; a new-generation propulsion part or substitute-material redesign may violate both assumptions. Claims data with only one aggregate nominal payment cannot identify all component contributions without supplemental estimates, invoices or repair-location metadata.
Observed paid severity may increase because the damage distribution changes rather than the price of fixed repairs. An aging fleet, more complicated machinery, a higher share of large ships or longer yardside downtime can change claims composition. Claim-type and vessel-type standardization is needed to interpret a price trend. Inflation also modifies the cost of outstanding claims, so historical loss development and projected cost trends must be jointly reconsidered rather than independently multiplied without timing alignment.
Component basket and foreign exchange
For fixed base repair quantities and component weights summing to one, the local-currency relative cost index is , where is the component's price relative. For a synthetic basket, take materials at 30% weight with relative 1.20, labor at 45% with 1.10, yard services at 15% with 1.25, and transport at 10% with 1.05. The total is $0.30(1.20)+0.45(1.10)+0.15(1.25)+0.10(1.05)=1.1475$, or a 14.75% price increase at constant currency.
If 60% of the resulting cost is paid in a foreign currency that appreciates 10% against the insurer's reporting currency, and the other 40% is unchanged, the simplified reporting-currency relative becomes $1.1475[0.60(1.10)+0.40]=1.21635$, or 21.635%. The multiplication includes an interaction term. In an actual model, the foreign-currency share and the relevant forward or spot exchange rate must be assigned by item and payment date; an aggregate factor risks double-counting currency conversion already included in a vendor index.
Repair cost, insured value and nominal policy limits can change independently. Increasing vessel insured values may increase maximum total-loss exposure without raising the price of identical steel repairs by the same amount. A fixed nominal deductible can become economically smaller as repair prices rise, shifting more small repairs into positive paid claims. Reinsurance attachment points, if fixed in nominal currency, may also see more frequent exceedances, altering retained insurer severity without a change in underlying physical event incidence.
Threshold migration creates an observed frequency effect
Suppose comparable pre-inflation gross damage is , inflation relative is , and a fixed insurer deductible is . Then . As increases, the deductible's corresponding real-cost threshold falls, so more accidents generate positive payments even with a fixed physical accident intensity. Consequently the count of paid hull claims and their conditional average paid amount can both increase because of cost inflation. The event and observation rates must not be conflated.
The expected insurer payment before limits is . This expression cannot generally be reconstructed by inflating historical mean net payment by , because the deductible does not scale with the gross repair cost and the threshold-crossing mass changes. A properly specified model should simulate or integrate the gross severity distribution through the actual prospective deductible and limit. Where only net historic claims are accessible, state the lost identification and use bounds or justified gross-to-net reconstruction instead of a fictitious exact net-inflation rate.
Reported large-claim trends have a corresponding selection issue. The IUMI Major Claims Database focuses on major claims and its contributing institutions. If the nominal threshold is unchanged, inflation alone can affect the number of observations crossing it. Constant-purchasing-power bins, claim maturity adjustments and stable insurer contributions are needed for inference about underlying high-damage incidence.
Interpreting institutional market trends
The international 2026 IUMI report and Cefor's 2025 year-end hull statistics supply legitimate but different external checks. IUMI reports nominal global class-level market premiums, whereas NoMIS represents participating hull insurers and their claim experience. Neither directly supplies all variables required for a global repair inflation regression on insured vessel-years. Currency translation, selection and class composition must be identified before using published aggregates as input to a hull pure-premium trend.
Evidence Appraisal and Sensitivity
Distinct source types and estimands
An original statute supports a legal proposition within its territorial and temporal applicability. A plan clause supports an agreed contractual rule. An insurer's audited loss record supports the observed insurer payment and reserve at a specified valuation date. A safety authority's record supports its accident reporting universe. A market association's report supports the statistics collected from its contributing members. An actuarial paper supports properties of its mathematical model under stated assumptions. Exchanging one evidence type for another without a validated bridge creates an inference error.
The distinction is visible in contemporaneous institutional material. The EMSA report's European scope cannot supply the denominator for worldwide hull insurance claims. Cefor's 2026 claim-band findings depend on participating insurer records and specified nominal thresholds. IUMI's 2025 hull premium total contains changes in premium volume and insured-value composition. A ship-repair PPI measures producer prices in a specified national industry. Their combination provides external context, not a single observed joint dataset.
Source-to-claim control therefore requires recording each quotation or quantitative statement beside its directly identifying contextual link and full bibliography record. Source IDs are internal evidence pointers only; they carry no human-facing scientific meaning. A manuscript can use many short context-specific hyperlinks while keeping the visual line readable and the full source metadata in References and the evidence ledger.
Reproducible numerical sensitivity
Take a synthetic homogeneous portfolio with 100 fully exposed vessel-years, expected registered claim intensity per vessel-year, and independent mean net positive paid claim severity of USD 180,000. The expected number of claims is 12. Expected portfolio paid losses are $100(0.12)(180000)=USD 2,160,000$, or USD 21,600 per vessel-year. No measured insurer experience is implied by these inputs.
If the same conditional paid severity increases 12%, with count intensity fixed, pure premium becomes $21600(1.12)=USD 24,19221600(1.10)(1.12)=USD 26,611.20$. This illustrates multiplicative trend interaction. It is not automatically valid when the 12% refers to gross repairs while the USD 180,000 baseline is net of a fixed deductible: then the entire gross distribution must be converted again under the coverage terms.
A three-by-three risk input sensitivity grid can set frequency to 0.10, 0.12 or 0.15 and mean net severity to USD 150,000, USD 180,000 or USD 240,000. Under the independent compound mean formula, annual pure premium ranges from USD 15,000 to USD 36,000 per vessel-year. This is a deterministic scenario range, not a confidence interval. Predictive uncertainty also depends on count dispersion, severity variance, model coefficients, tail truncation, shared events, future inflation and random claim development. A good-looking narrow sensitivity range based only on two point means does not measure those uncertainties.
Fleet composition and selection
Let be the vessel-exposure proportion in cell for year , and let be its claim intensity. The actual pooled frequency is . To compare time trends at a fixed historical fleet mix , calculate . Their difference reveals the contribution from changes in proportions across risk cells. This is conventional direct standardization, applied to vessel count rates. A similar severity comparison requires weighting by claim-generating populations, not simply vessel exposures, because high-frequency vessel classes generate disproportionately many claims.
Vessel age, ship type, insured value, machinery, utilization and geographic region are candidate classification variables, but each can be imperfectly observed and correlated with insurer selection. Selection into insurance may change if war risk, fleet market conditions, premiums or deductibles change. An apparent historical frequency trend may thus reflect a new insured cohort. No public market statistic should be assigned a causal interpretation until the source population and observed exposure movement are reconciled.
Model comparison and statistical diagnostics
Compare Poisson with negative-binomial alternatives for overdispersion and dependence. Inspect conditional zero frequencies, count-tail quantiles, residuals by vessel type, and out-of-time prediction. Mixed-effects specifications may reflect operator and vessel heterogeneity but can overfit if ship histories are short. More flexible structures should be selected for meaningful predictive improvement after exposure and reporting effects are controlled. A zero-inflated model may fit excess zeros yet fail to distinguish no accident from an unreported accident.
For severity, use separate diagnostics for attritional losses, medium-size claims and large-loss tails. Check predictive likelihood for positive payments, limit exhaustion, deductible exceedance and out-of-time mean cost. Compare claims on a consistent inflation basis and ensure full claim development. The proper predictive scoring analysis of Gneiting and Raftery motivates evaluating the predictive distribution for the stated target, not just minimizing error in an in-sample point estimate. A nominal total-cost model can be well calibrated overall while badly miscalibrated for the ship classes that dominate the portfolio's capital risk.
Reserving and the calendar effect
Outstanding hull claims accumulate repair-price inflation, foreign-exchange movements, legal expense, survey revisions and settlement-delay effects after the original accident. An accident-year claim triangle with payments by development age is therefore sensitive to calendar-period changes. The stochastic claims reserving literature of England and Verrall offers alternative development models and predictive uncertainty treatment. Mack's 1993 chain-ladder result addresses standard errors under a particular chain-ladder setting. Neither is a substitute for testing whether payment development assumptions remain stable when yard congestion, invoice inflation or settlement delay shifts.
A robust validation plan would independently vary repair inflation, settlement speed and upper-tail event frequency, then reconcile differences between paid and incurred ultimates. For large total-loss or disputed files, claim-level technical reserves may be needed rather than a generic aggregated development factor. Claims model governance must preserve the date of every reserve snapshot and the exchange rate used to transform its original currency.
Repair index and empirical identification
Repair-cost inflation may be driven by multiple correlated macroeconomic factors. When shipyard activity, commodity prices, energy and labor costs move together, estimated regression coefficients can be unstable even if the in-sample fitted index appears smooth. IUMI's 2026 method warning expressly notes multicollinearity and non-public inputs. A rigorous reproduction should report which input series are original, which are substitutes, how missing and revised observations are handled, and how the model performs out of time. No calculated source in this dossier is represented as reproducing the proprietary IUMI series.
Inflation analysis should include local versus reporting currency results, cost composition, shipyard location and case-settlement horizon. It should explicitly test whether changing the deductible alters positive-paid claim frequency and average severity. If complete ground-up costs are unavailable, estimate sensitivity intervals for selection effects and identify them as assumptions. A direct positive-claim severity regression alone cannot recover sub-deductible losses absent external data.
Legal and professional scope of inference
The prospective-cost approach of ASOP No. 53 is applicable to actuaries within its specified U.S. professional scope. Its general lesson—define the intended cost and assess exposure and data appropriateness—has analytical utility internationally, but legal authority remains bounded. Similarly, the British statute and Nordic contractual plan cannot be cited as universal maritime insurance law. International comparison must label jurisdiction, document date and the contractual adoption condition.
Scientific Synthesis and Research Questions
Integrated analytical findings
Hull insurance frequency is a rate of events meeting a defined insurance counting rule over a defined insured exposure. A vessel-year is a natural time basis for annual hull accounts; additional operating intensity indicators can improve cause-specific modeling. Frequency of physical casualties, accepted claim files and positive payment claims are separate rates, linked by reporting and contract selection. An insurer may record higher paid claim frequency during inflation simply because more physical repair amounts exceed fixed nominal deductibles.
Hull severity is a distribution over a specified cost definition and valuation date. Gross repair expenditure, admitted contract loss, ultimate net payment and current accounting transaction are different monetary variables. Severity can be dominated by high-tail events, while the routine mid-size claims band may exhibit a distinct trend. Conditional event-frequency and severity dependence matters: the original two-state counterexample shows an explicit difference between a correct state-conditional aggregate loss and the product of independent vessel-weighted means.
Hull inflation is a multi-component repair price process translated into insurance contract and reporting currencies over time. It intersects with claim observation and reserving. The original cost-basket example produces a 14.75% fixed-currency increase and a 21.635% reporting-currency increase under stated foreign-currency assumptions, demonstrating why broad consumer inflation or exchange rates alone are insufficient. Historical hull profitability and technical price comparisons require standardization of fleet composition, contract terms, economic loss date and claims development.
The available institutional record confirms the importance of these distinctions. IUMI documents aggregate marine premium markets and its inflation index methodology, Cefor reports claims patterns for its participating hull market, EMSA compiles European maritime casualty reports, and BLS measures a U.S. repair sector. Each supports particular propositions; none alone identifies worldwide insured hull claim cost per vessel-year. The scientific contribution of the present synthesis is to connect these observational layers with existing actuarial mathematics and identify the exact additional data needed for an empirically identified hull rating model. It establishes no named new theory or doctrine.
Future empirical research
The first empirical test is whether the reported increase in mid-size hull claim frequency survives adjustment for insured vessel years, fleet age and type, fixed real-value loss bands, deductibles and reporting lag. A second test is decomposition of nominal positive severity changes by physical damage type, regional yard labor, machinery inputs, material costs, settlement timing and payment currency. A third test is whether vessel condition and external operating regimes induce a materially positive frequency–severity covariance in longitudinal marine hull data. A fourth concerns large-claim tail sensitivity to vessel values, cause mix and reinsurance attachment. A fifth concerns whether aggregate claim development can be separated reliably from calendar-year price inflation in the presence of long-running hull repairs.
Answering these questions requires anonymized contract-linked hull records with vessel identifiers, prospective exposure intervals, damage causes, gross repair estimates, deductible and limit histories, payment currencies, survey and reserve dates, insurer participation shares and consistent extraction dates. Institutional summary reports can provide external consistency checks, but the required joint exposure–claim observations are usually proprietary. Empirical inference should report these access constraints rather than reverse-engineer fictional microdata.
Boundaries of the completed research
The linked legal documents, institutional releases, statistical series and publisher-indexed actuarial works have been reviewed within the stated source-access scope. Some historical academic works were available through publisher metadata and abstract rather than full text. The IUMI inflation index's exact original private calculations were unavailable, and no insurer microdata or representative world fleet panel was available. All numerical illustrations are explicitly synthetic. These limitations restrict claims of empirical generalizability but do not invalidate the stated conditional mathematical derivations or verified institutional descriptions. External peer review was outside the assigned internal research cycle.
Conclusion
Marine hull claims are generated through connected physical, contractual, economic and administrative processes. Their statistical analysis requires an exposure base tied to insured vessels and peril periods, a stable casualty and reporting definition, a distinct model for gross damage and net indemnity, a temporal treatment of repair inflation and exchange rates, and a claims development process consistent with the intended future cost. The standard compound expected-loss relationship is reliable only when its frequency, severity and dependence assumptions match the observed data.
Contemporary source evidence supports the importance of model heterogeneity and inflation, but also demands restraint about what is measured. The IUMI global hull premium series measures market financial volume, Cefor's hull claim bands describe participating market loss experience, EMSA describes defined safety-reporting incidents, and BLS provides a localized producer-price index. The 2026 IUMI Hull Inflation Index carries documented data-access and collinearity limitations. No one of these sources supports an unqualified worldwide risk-adjusted hull frequency or repair-cost inflation coefficient.
The monograph's original analytical synthesis shows how contractual selection, statistical count models, severity distributions, repair-cost composition and claims development interact. Its synthetic arithmetic demonstrates economically consequential mistakes produced by conflating the levels. The resulting method is an application and critical integration of established law, industry documentation and actuarial statistics, with all empirical and jurisdictional uncertainties explicitly delimited.
References
International Union of Marine Insurance (IUMI). 2026. IUMI Stats Report 2026. October 23. Hamburg: IUMI. Source
International Union of Marine Insurance (IUMI). 2026. The IUMI Hull Inflation Index 2026. May 5. Hamburg: IUMI. Source
Nordic Association of Marine Insurers (Cefor). 2026. 2026 Mid-year Hull Report. Nordic Marine Insurance Statistics. Source
European Maritime Safety Agency (EMSA). 2026. Annual Overview of Marine Casualties and Incidents 2026. October 13. Lisbon: EMSA. Source
United Kingdom. 1906. Marine Insurance Act 1906, 6 Edw. 7 c. 41, sections 55 and 67–69. Legislation.gov.uk. Source
Nordic Marine Insurance Plan. 2023. Nordic Marine Insurance Plan of 2013, Version 2023. NordicPlan.org. Source
U.S. Bureau of Labor Statistics. 2026. Producer Price Index by Industry: Ship Building and Repairing: Ship Repair, Nonmilitary, PCU336611336611A. Federal Reserve Bank of St. Louis FRED. Source
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. Source
Ohlsson, Esbjörn, and Björn Johansson. 2010. Non-Life Insurance Pricing with Generalized Linear Models. Springer. DOI: 10.1007/978-3-642-10791-7. Source
de Jong, Piet, and Gillian Z. Heller. 2008. Generalized Linear Models for Insurance Data. Cambridge University Press. DOI: 10.1017/CBO9780511755408. Source
Renshaw, Arthur E. 1994. Modelling the Claims Process in the Presence of Covariates. ASTIN Bulletin 24 (2): 265–285. DOI: 10.2143/AST.24.2.2005070. Source
Jørgensen, Bent, and Marta C. Paes de Souza. 1994. Fitting Tweedie's Compound Poisson Model to Insurance Claims Data. Scandinavian Actuarial Journal 1994 (1): 69–93. DOI: 10.1080/03461238.1994.10413930. Source
Lambert, Diane. 1992. Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing. Technometrics 34 (1): 1–14. DOI: 10.1080/00401706.1992.10485228. Source
England, Peter D., and Richard J. Verrall. 2002. Stochastic Claims Reserving in General Insurance. British Actuarial Journal 8 (3): 443–518. DOI: 10.1017/S1357321700003809. Source
Mack, Thomas. 1993. Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates. ASTIN Bulletin 23 (2): 213–225. DOI: 10.2143/AST.23.2.2005092. Source
Gneiting, Tilmann, and Adrian E. Raftery. 2007. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association 102 (477): 359–378. DOI: 10.1198/016214506000001437. Source
International Union of Marine Insurance (IUMI). 2025. Major Claims Database 2025. Hamburg: IUMI. Source
International Union of Marine Insurance (IUMI). 2020. Scope of Marine Premium and Loss Ratio Reporting. Hamburg: IUMI. Source
Nordic Association of Marine Insurers (Cefor). 2026. 2025 Year-end Hull Reports. April 13. Nordic Marine Insurance Statistics. Source