RESEARCH DOSSIER
Offshore Captive Insurance Loss Experience
Abstract. Offshore captive insurance loss experience consists of the observed and retrospectively estimated claims arising from risks financed through captive insurance contracts domiciled outside the principal operating jurisdiction of their insured organizations. This monograph examines the relationship between underlying hazards, contractual risk selection, claim reporting, reserve development, reinsurance and reported underwriting outcomes. Its method combines comparative documentary analysis of Bermuda, Cayman Islands and Guernsey supervisory reporting with established actuarial methods, a review of professional standards, and explicitly synthetic numerical demonstrations. A five-year series reconstructed from successive Bermuda Monetary Authority reports gives published average loss ratios of 48%, 53%, 55%, 53% and 57% for 2019–2023. These observations establish a bounded description of reported experience within that reporting perimeter. They neither identify the loss cost of a standardized exposure nor establish a causal domicile effect. Cayman and Guernsey data demonstrate the importance of distinguishing entities, cells, risk location, financial aggregates, and claim-bearing populations. The analysis finds that comparable experience requires aligned claim layers, coverage triggers, valuation dates, currencies, exposure measures and premium denominators; reserve and reinsurance movements require a separate accounting reconciliation. The insurance significance of domicile resides in the institutional organization, financing and observation of risk. An analytically valid evaluation of captive performance preserves the distinction between the captive's legal balance sheet and the enterprise's consolidated economic burden. Publicly available aggregates support descriptive research and carefully framed institutional comparisons; risk-adjusted benchmarking and causal conclusions require additional claim-level and exposure data.
Keywords: Offshore insurance; captive insurance; loss experience; loss ratio; claims development; risk retention; reinsurance; actuarial credibility; Bermuda; Cayman Islands; Guernsey; insurance institutions; exposure normalization
Exposure, contracts, and the meaning of loss experience
The scientific object of this study is the realized and developing loss experience of offshore captive insurers. A captive is an insurance institution whose underwriting is materially organized around the risks of its owner, participating members, or associated organizations. Its domicile establishes the jurisdiction in which the insurer is constituted and supervised. The insured factory, hospital, vehicle fleet, employee population or digital infrastructure can be situated elsewhere. The same captive may receive risk from several jurisdictions and several members of an enterprise group. Offshore is therefore a relational designation in this study: the location of insurance organization differs from the principal location of the insured operations. It does not denote a particular moral, fiscal or risk-quality classification.
The IAIS (2015) application paper on captive supervision situates captives, insurance managers and cell companies within the supervisory problem. These institutional distinctions create separate units of observation. An insurer license describes a supervised entity. A cell may identify a segregated account or distinct risk-bearing arrangement. A policy identifies a contractual obligation; a claim identifies a demand arising under specified conditions. A single event may generate multiple claims or affect multiple cells. Statistical counts of these units answer different questions.
Loss experience begins with an occurrence that can generate a financial burden. An industrial accident can produce property damage, bodily injury, business interruption and litigation expenses. Whether these amounts become insurance losses depends on contractual triggers, limits, deductibles, exclusions, periods and applicable legal interpretation. A reserve then records an estimate of the insurer's outstanding obligations. Payment is a cash-flow event; it may occur years after the original occurrence. Each of these stages has its own temporal and measurement basis.
The analytical question is how observed claims can support reliable assessment of retained risk, underwriting results and the adequacy of historical experience for prospective decisions. The relevant comparisons are between consistent risk-bearing units. Claims arising from a hospital professional liability program and a short-tail property deductible program are not interchangeable observations even where the same captive administers them. The captive's location is informative about its institutional arrangements but gives no direct measurement of the frequency or severity of its insured events.
This study uses a purposive documentary corpus rather than a systematic census of all domiciles. Bermuda supplies a historical loss-ratio series; Cayman and Guernsey supply institutional denominators and the boundaries of published financial statistics. The original actuarial studies provide conceptual and methodological grounds, with access depth identified in the evidence ledger. All numerical examples outside the identified official observations are invented for demonstration. No confidential individual claims, regression estimates, actuarial opinion, comprehensive legal comparison or global market census is presented.
The architecture of gross and net claims
Consider an enterprise that purchases primary insurance from a fronting company. The fronting company issues a policy to an operating affiliate and assumes the contractual duty specified by that policy. A captive assumes an agreed portion through reinsurance. The captive may in turn buy excess-of-loss protection. The operational event remains one occurrence, even though several contracts and several financial statements respond. A loss ledger organized by company without contract identifiers risks counting the same economic event more than once.
Gross loss must be defined from the perspective of the reporting entity. For a captive reinsurer, its gross assumed claim is the amount owed under its inward treaty before its outward reinsurance. It need not equal the original insured's total loss or the fronting insurer's gross direct claim. Net retained loss subtracts the contractual outward recoveries on the same incurred or paid basis. Expected recovery, billed recovery and cash received are different stages. A reserve net of estimated recovery describes an expected exposure; a cash forecast must accommodate settlement lags, counterparty failure and disputes.
For a simplified per-occurrence layer with attachment a and limit l, the layer responds to an eligible underlying loss x by the amount min(max(x − a, 0), l). This arithmetic assumes a single occurrence, no aggregate exhaustion, no additional expenses or reinstatement provisions and an eligible claim. Actual wording controls. The economic total above that layer belongs to another party or remains uninsured, depending on the contract structure.
Suppose an eligible event costs $2.4 million. A captive assumes the layer from $0.5 million through $2 million, a limit of $1.5 million. The captive gross amount is $1.5 million. Assume a further treaty for the captive reimburses $0.6 million. Its net retained amount is $0.9 million. The insurer's gross, ceded and net figures describe one contractual allocation, not three separate losses. An enterprise may still bear the first $0.5 million, depending on its agreements, and a different insurer or the enterprise may bear the $0.4 million above the captive layer.
A valid experience study must attach an event identifier to the underlying occurrence and a contract identifier to each allocation. Its gross-to-net reconciliation must track amount, status, currency, applicable share, expenses, recovery and date. When risk transfer is proportional, aggregation of matched shares may be straightforward; where protections are nonproportional or apply to annual aggregates, claim-by-claim subtraction may not reconstruct the actual net exposure. Treaty recovery depends on the whole contractual loss population.
A captive's entity-level underwriting profit is real within its financial reporting perimeter. Consolidated group economics are a second analysis. An internal premium may move funding between subsidiaries, while indemnity distributes losses within the group. The consolidated burden depends on external cover costs and recoveries, claims, administration, capital opportunity costs and applicable taxes. Treating all captive surplus as savings would ignore the costs and liabilities that support it. Conversely, assigning all gross claims to the captive's owner would disregard external insurance that actually responds.
Frequency, severity, and sparse observations
Frequency is the number of eligible claims divided by a meaningful exposure volume. For vehicle risks, vehicle-years or distance traveled may be appropriate; for workers' compensation, payroll or employee-hours; for medical liability, procedures or insured practitioner-years. Each denominator represents a different mechanism. A count of insurers measures organizational population, not the amount of risk to which those insurers are exposed.
Severity describes claim amount conditional on the defined claim population. For a fixed valuation, total measured claim amount equals count multiplied by average measured severity. Consider two synthetic captive programs with 1,000 vehicle-years each. Program A has 30 claims totaling $450,000; program B has 20 claims totaling $500,000. A records 3 claims per 100 vehicle-years and B 2. Average amounts are $15,000 and $25,000. A has higher frequency but lower observed loss cost: $450 per vehicle-year versus $500. Claims control decisions therefore require the two components to be separated.
The policy determines which events enter the dataset. Raising a deductible can remove minor events from the captive's claims records even if the underlying number of incidents remains unchanged. A recorded reduction in captive frequency following such a change is partly a filtering effect. A severity increase can also be caused by that filter because the remaining claims are concentrated higher in the loss distribution. A risk-control evaluation needs underlying incident information including events below the deductible, subject to consistent recording.
The empirical distribution of claim amounts contains information about skewness, large events, and concentration. Its sample mean is vulnerable to one expensive claim when exposures are small. Independence of occurrences cannot simply be assumed when facilities share a catastrophe footprint or when multiple claimants arise from one event. For independent claim counts N and identically distributed positive claim severities X, with count independent of the severities and finite second moments, aggregate S = sum of X has expectation E[N]E[X] and variance E[N]Var(X) + Var(N)(E[X])². Under a Poisson count with mean λ, this becomes λE[X²]. The conditional assumptions matter more than the convenience of the formula.
A captive with few claims has little direct statistical information about rare losses. A year of zero reported claims can result from a short observation period, reporting delay, high attachment, low exposure or favorable random variation. It does not establish an absence of hazard. The ASB's (2013) credibility procedures provide a professional foundation for weighing subject experience alongside relevant supplementary data. The standard belongs to U.S. actuarial practice; its analytical logic is useful internationally without implying legal force in each domicile.
In a hypothetical credibility calculation, an own-experience estimate of $600 per exposure unit blended with a comparable external estimate of $800 using an assumed own weight of 0.4 yields $720. This choice of weight is illustrative, not estimated. To implement a credibility procedure scientifically, the complement must correspond to the actual risk class, layer, exposure and valuation basis. An external estimate from primary low-deductible insurance is not an appropriate unadjusted complement for a high excess layer.
An improvement in experience can reflect improved underlying safety, changing insured population, altered reporting practices, random variation or contractual filtering. The frequency-severity decomposition identifies where a change appears. Determining its cause requires an additional measurement design.
Loss development and valuation uncertainty
The loss incurred by an accident cohort develops over time. As claims are notified, facts become available, case estimates change, litigation resolves and payments occur. A measure that incorporates case reserves remains an estimate at its valuation date. Ultimate losses estimate the eventual total of all covered amounts for the cohort; even a mature cohort may retain uncertainty when obligations remain open.
On a consistent undiscounted basis, let P be cumulative paid claims, C outstanding case reserves and B additional development including incurred-but-not-reported claims and, where applicable, incurred-but-not-enough-reported amounts. The estimated ultimate is P + C + B. The unpaid estimate is C + B. An incurred amount that includes only paid and case reserves is P + C. Definitions must be documented because practices differ in whether booked incurred includes bulk provisions and allocated claim expenses.
The ASB (2007) unpaid claim estimate standard addresses the estimation of unpaid property/casualty obligations. Its distinction between an estimate and the eventual outcome supports a necessary interpretation of captive liability data: the recorded claims balance reflects information and uncertainty at a specified date. A favorable future revision may mean earlier reserves were conservative, while adverse development may reflect information that could not have been known earlier or assumptions that were inadequate.
For a conventional cumulative paid-loss development triangle, each row identifies an accident cohort and each column the time since origin. A volume-weighted age-to-age factor is the sum of later cumulative amounts divided by the sum of earlier amounts over comparable observed rows. Multiplying the latest cumulative amount by remaining factors projects an ultimate. The open actuarial reserving text explains this method and its stochastic extension.
The following triangle is synthetic, in thousands of dollars, with no tail beyond 36 months solely to make the demonstration transparent.
| Accident year | 12 months | 24 months | 36 months |
|---|---|---|---|
| 2023 | 400 | 600 | 660 |
| 2024 | 500 | 700 | — |
| 2025 | 550 | — | — |
The 12-to-24 factor is (600 + 700)/(400 + 500) = 1.44444. The 24-to-36 factor is 660/600 = 1.10. Projected ultimate for 2024 is 770; for 2025 it is 550 × 1.44444 × 1.10 = 873.89. Total unpaid on those cohorts is 70 + 323.89 = 393.89. This calculation is a demonstration, not an acceptable reserve analysis of real data: the final development factor has one observation; no tail or statistical interval is estimated.
Mack (1993) derives a chain-ladder standard error under specified stochastic assumptions. The model avoids specifying a complete claim distribution but requires structural conditions, including independence of accident years and an appropriate conditional mean and variance structure. A common legal shock affecting several cohorts, changes in settlement policy or a change in retention may undermine those assumptions. Uncertainty between cohorts can also arise from shared estimated development factors. The covariance terms in aggregate uncertainty calculations therefore matter even when the model assumes independent underlying accident periods.
England and Verrall (2002), in stochastic claims reserving, discuss prediction error, tail extrapolation and the treatment of prior estimates within a Bornhuetter–Ferguson framework. A simplified paid Bornhuetter–Ferguson illustration uses an expected ultimate of 900 and an assumed paid proportion of 0.60. With actual paid amount 550, the estimated ultimate is 550 + 900 × 0.40 = 910. The unpaid estimate is 360. The difference from the chain-ladder illustration indicates sensitivity to the assumed emergence and prior, not a contest yielding one universally preferable method.
Small captives often have more heterogeneous experience than the large portfolios from which conventional factors are estimated. Case-level projections, separate large claim analyses, external development patterns, and explicit scenario ranges may be more informative than a mechanically fitted triangle. A model should be tested by reconstructing earlier valuation diagonals and comparing its prior forecasts with later emergence, while preserving information available at the original estimation date. Such a test distinguishes predictive adequacy from the ability to fit a completed triangle retrospectively.
Bermuda: a longitudinal supervisory observation
Bermuda Monetary Authority reports provide dated observations of captive underwriting outcomes. The BMA (2025a) 2023 year-end captive report identifies its statutory financial-return basis and discusses general-business captive classes alongside long-term captive classes. The profitability chart presents the aggregate captive and general-business class groups. The published ratio label is retained rather than silently recasting it as an independently reconstructed accident-year or premium-weighted ratio.
The BMA (2023) 2021 year-end report supplies the 2019–2021 sequence. The BMA (2025b) 2023 Captive Report contains 2020–2022 observations, whereas the later report contains 2021–2023 observations. Overlapping values agree in the retrieved text. The report called 2023 Captive Report refers here to the 2022 year-end observation; the data year and upload year are deliberately distinguished.
| Year-end data year | Published average loss ratio | Published average combined ratio |
|---|---|---|
| 2019 | 48% | 60% |
| 2020 | 53% | 77% |
| 2021 | 55% | 79% |
| 2022 | 53% | 75% |
| 2023 | 57% | 79% |
Source: successive BMA reports, profitability sections, pp. 20–21. Values are rounded as published. This is an aligned transcription of the published series, not a re-estimation from insurer microdata.
Across these five observations, the loss-ratio range is 9 percentage points, from 48% to 57%. The arithmetic mean of the five published annual percentages is 53.2%; the combined-ratio mean on the same basis is 74.0%. Those means give equal weight to years and are not pooled insurance-wide ratios. A pooled ratio would require the relevant annual denominator and numerator, a consistent class perimeter and a documented treatment of missing observations. The 2022–2023 increase is 4 percentage points in both reported series.
The 2023 report's published class combined ratios are 75% for Class 1, 85% for Class 2 and 78% for Class 3. The equivalent 2022 values are 67%, 79% and 80%. Thus the year-on-year movements differ in sign by class: plus 8, plus 6 and minus 2 points. A single aggregate movement obscures this heterogeneity. Class figures refer to combined ratios, not class-specific loss ratios. It would be an error to substitute those published combined ratios into a loss-only ranking.
The same 2023 report describes a market including pure captives, insurers in runoff and segregated accounts. Its summary reports 67% pure captive structures and 14% in runoff. These are population descriptors, not observations of claims frequency. Runoff status is important because claims and reserve development can persist after active premium writing has stopped. A changing runoff share may influence aggregate ratio interpretation. No claim in this study attributes the observed Bermuda series to a particular runoff mechanism, as the data required for such a decomposition are absent.
A five-point series is enough to illustrate fluctuation and the need to preserve valuation context. It is insufficient for identifying a causal trend, fitting a dependable stochastic distribution of jurisdiction-wide future loss ratios or evaluating the effect of any particular reform. The text-retrieval access provided the relevant figures and overlap checks; separate visual PDF page retrieval was unavailable. The report pages are identified explicitly to support independent replication.
Cayman and Guernsey: the limits of comparative data
The Cayman Islands Monetary Authority discloses detailed international insurance statistics. In its Q2 2026 international insurance company tables, CIMA reports 721 licensees in the category table, including 293 pure captives and 142 group captives. The remaining categories include commercial insurers, reinsurance companies, segregated portfolio companies and special purpose vehicles. The pure and group categories sum to 435, or approximately 60.3% of that published total. This category share does not identify how many separate risk-bearing portfolios are represented or which segregated portfolios function as captive programs.
The risk-location table identifies 635 of 721 licenses under North America, approximately 88.1%. Its primary class table identifies 149 under workers' compensation and 131 under medical malpractice liability. These are classifications of international insurers rather than direct incident counts. A company assigned to a primary class can have heterogeneous exposures. The concentration in North American business also illustrates why offshore domicile cannot serve as a proxy for geographic exposure.
CIMA's site separately publishes domestic audited insurance statistics for 2024, with earned premiums, paid claims and movements in reserves. These fields illustrate the components required to build a calendar-year loss measure. The domestic table must not be used as a proxy for international captive loss experience. This exclusion is analytically material: the source contains loss amounts alongside international license and premium data, but the tables describe different populations.
Guernsey's international insurance statistics provide another measurement distinction. At December 31, 2025, GFSC reports 196 non-cell captives and 126 captive cells; the respective preceding year-end counts were 197 and 122. Some earlier figures are restated in the source. The 2024 international insurer financial series reports £29.70 billion of gross assets, £12.62 billion of net worth and £5.19 billion of written premiums. These are broader international insurer aggregates, not a loss-experience table exclusively for captives. Adding cells and non-cell captive entries produces 322 recorded count units with dissimilar legal significance, not necessarily 322 standalone companies.
| Public source | Date of observation | Relevant observation | Comparison boundary |
|---|---|---|---|
| Bermuda BMA | 2019–2023 | Published average loss and combined ratios | Defined statutory report groups; weighting not independently reproduced |
| Cayman CIMA | Q2 2026 | License categories, primary classes, risk location | International classification, no matched captive numerator in inspected tables |
| Guernsey GFSC | 2025 year end; finances 2024 | Separate non-cell/cell counts and broad premium volume | Counts and financial aggregates have different scope |
The difference among these public disclosures is a substantive finding about observability. Bermuda offers a directly reported underwriting indicator; the other two sources, as inspected, support analysis of population composition. The comparison cannot supply three matched loss ratios. A researcher must preserve this missingness rather than construct a denominator by combining unrelated tables.
The research implication extends beyond the three examples. Better publication of a ratio improves the ability to observe that ratio; it does not itself establish superior insurance performance. A domicile with limited public detail may still maintain extensive confidential supervisory records. Data availability and underlying risk quality are distinct variables.
Benchmarking, selection, and causal identification
Industry benchmarks establish useful reference points for insurance organizations, provided their populations and measurement rules are known. AM Best's (2025) U.S. captive performance release reports a five-year combined ratio of 88.0 for its rated U.S. captive composite and 97.0 for the commercial casualty peer composite. It also reports a combined ratio of 98.5 in 2024, a deterioration of 9.6 percentage points. These are rated U.S. populations and combined ratios. They are not offshore captive loss ratios. The evidence supports the observation that a selected comparative population may perform differently over time and against another composite; it does not isolate the contribution of the captive structure.
Marsh's (2026) captive benchmarking report page identifies the practical objective of data-backed captive evaluation. A manager's customer portfolio can illustrate the programs it serves. Its representativeness of an entire domicile depends on coverage, inclusion criteria, client mix, entity definitions and treatment of terminated programs. This study uses the landing page to characterize the role of provider benchmarks, not as evidence of unpublished numerical findings.
A benchmark can be distorted by weighting. In a synthetic pool, captive A has $0.2 million in earned premium and $0.18 million incurred losses (90%); captive B has $1.8 million in earned premium and $0.72 million losses (40%). Their arithmetic mean loss ratio is 65%; the pooled premium-weighted ratio is 45%. Both are mathematically valid summaries of different questions. The first describes equally weighted entities; the second describes the total monetary loss relative to premium. A regulator concerned about the distribution of entity outcomes and a group evaluating total insured financing need different summaries.
Composition can also produce apparent improvement when performance within each category deteriorates. Suppose the initial high-risk segment has a 90% ratio on 80 units of premium and the low-risk segment has a 30% ratio on 20 units. The aggregate is 78%. In the next year, corresponding ratios worsen to 95% and 35%, but the weights reverse to 20 and 80 units. The aggregate falls to 47%. Both segments have deteriorated by 5 points while the total improves by 31. This is a synthetic demonstration of aggregation and changing weights, not an empirical result for a domicile.
Selection operates through formation, continuation, closure and reporting. Enterprises with sophisticated risk systems may be more inclined to form captives. Enterprises facing unusually expensive commercial coverage may retain different risks than those entering conventional markets. A dataset of active captives can exclude terminated or failed programs. Ratings select another population. If inclusion correlates with experience, the observed distribution differs systematically from the complete cohort.
A causal comparison of domiciles would require the potential experience of substantially comparable portfolios under alternative institutional arrangements. Switching domicile may coincide with changed coverage, fronting, reinsurance, management and reserve policies. Even a within-captive before-and-after study can be confounded by those contemporaneous changes. An observational design could use matched risk cells, prior experience and event timing; its identification assumptions would still require substantive assessment. No credible causal ranking can be derived from the reported descriptive figures here.
Catastrophes, aggregation, and retention changes
Offshore captives finance layers with widely differing loss distributions. A deductible reimbursement layer may have recurring relatively small claims, whereas a high excess layer can experience long periods without a covered loss. Risk concentration also differs between a diversified fleet and a single asset exposed to earthquake, windstorm or operational interruption.
The ASB (2017) prospective risk retention standard explicitly addresses infrequent events, relevant exposure adjustments and reinsurance assumptions. For a captive renewing its program, experience must be restated on the intended future terms. Historical incurred claims can be used to examine prior performance; expected future costs require updated assumptions. The difference matters when attaching at a different layer.
For independent outcomes, pooling can reduce relative uncertainty. A common event exposes several insured sites simultaneously, producing dependence that invalidates a simple independence-based calculation of total variance. If two exposures have variances of 100 and covariance of 60, their combined variance is 100 + 100 + 2 × 60 = 320, rather than 200 under independence. The units are squared loss units. This synthetic example shows why an international distribution of locations is not automatically evidence of diversification.
An aggregate stop-loss contract responds to total eligible losses over the period after the stated threshold, subject to its limit. Suppose a captive incurs three eligible claims of $0.7 million, $0.9 million and $0.8 million. With annual attachment of $1.8 million and a $0.4 million limit, the aggregate recovery is $0.4 million and net loss is $2 million. The equivalent recovery cannot be allocated by applying the aggregate attachment independently to each claim.
Catastrophe analysis must separate event identification, exposure at risk, policy response and claim development. Property losses may be visible rapidly while business interruption extends over months. Liability episodes can emerge slowly when injury attribution or litigation spans policy years. Cyber events may generate multiple claims across shared service providers. The event definition and common cause determine the analytical grouping; the insurance contract may define occurrence differently, and both groupings must be preserved.
Stress scenarios complement observed history, especially for rare losses. Scenarios should be transparent about the assumed event, exposed values, vulnerability, contractual provisions and correlations. A stress result is a conditional projection. It is not an empirical frequency estimate. Combining empirical attritional experience with explicitly modeled rare-event scenarios can support risk decisions without fabricating an observation for an event that has not occurred.
Claims governance and reporting integrity
Insurance governance converts claim information into financial oversight. The first requirement is to maintain a traceable link between insured events, claim notifications, third-party administrator records, payments, case reserves, actuarial estimates, reinsurance records and the general ledger. Each of these datasets may be collected by a different institution operating in a different jurisdiction. An offshore captive therefore involves a cross-border information chain as well as a contractual chain.
The ASB (2016) data quality standard identifies suitability for the intended analytical purpose as a central criterion and distinguishes data review from audit. This distinction is important in an offshore context. A reconciled claims ledger is an essential starting point for research, but it does not independently establish that every underlying event was recorded, that each claim is legally valid, or that unrelated parties supplied accurate estimates.
The claim record should preserve occurrence, report, coverage, payment and valuation dates, as well as a stable identifier and policy version. Exposure records should accompany claims. When a claim is transferred between systems or policies, the old identifier and mapping should remain discoverable. Duplicate identifiers, reopened claims, salvage and subrogation receipts, exchange-rate adjustments and negative reserve movements require explicit treatment. The inquiry is whether each entry represents a genuine correction, reimbursement, contractual transfer, or recording error.
Reconciliation should proceed from like to like. The sum of paid claim entries for a period should be compared with the corresponding claim expense or cash account after identifying timing, expenses, currency and classification differences. Closing outstanding amounts should be reconciled to the opening amount and movements attributable to new claims, payments, valuation changes and transfers. Reinsurance receivables should be reconciled separately. Equating a difference to fraud or negligence without further examination would exceed the evidence.
A captive's board and its professional advisers require information about both average results and adverse deviations. Appropriate indicators include outstanding claim development, severity concentration, aged recoveries, liquidity exposure and changes in expected ultimate. A year-end dashboard that reports only premium and paid claims can hide substantial development; a solvency dashboard that reports only reserve amounts can hide near-term cash-flow pressure.
Data confidentiality is an integral part of research governance. Medical liability, workers' compensation and employment claims can contain highly sensitive personal information. Research extracts should be proportionate to the question, use de-identification and access controls, and avoid disclosure of identifiable claim details. Cross-border analysis of data should respect applicable requirements. The institutional advantage of complete internal records does not depend upon making those individual records publicly available.
Research design for a comparable empirical study
A more conclusive empirical study would start with an explicitly enumerated population of captive programs and their reporting periods. It would define separate entity, cell, policy, treaty, event and claim tables, with a crosswalk to underlying economic exposure. Entry, exit and runoff history would remain visible. A fixed reporting cut-off would prevent mixing valuations drawn from materially different information sets.
A research extract should include at least the following dimensions.
| Dimension | Measurement requirement |
|---|---|
| Population | Domicile, entity/cell, formation, exit, runoff and reporting scope |
| Exposure | Risk location, class, exposure unit, volume and coverage period |
| Contract | Trigger, attachment, limit, aggregate terms and inward/outward share |
| Loss | Event identifier, paid amounts, case and bulk estimates, expenses |
| Valuation | Accident/report/calendar basis, development age and reporting date |
| Premium | Gross/net, earned/written, on-level adjustments and accounting basis |
| Currency | Original currency, conversion convention and inflation basis |
First, experience should be aggregated on a consistent exposure basis, stratifying by coverage class, layer and valuation age. Second, reporting and reserve movements should be reconciled against accounting records. Third, claim amounts should be developed, with sensitivity to alternative development assumptions and large claims. Fourth, comparable strata can be standardized to an explicit reference distribution of exposures. The analysis must retain both original and standardized results so that the consequences of weighting remain visible.
A fixed-mix direct standardization illustrates the operation. Using the second-year synthetic group ratios of 95% and 35% from the earlier example and the original premium weights of 80% and 20%, the standardized ratio is 83%, compared with an original 78%. The reported second-year aggregate of 47% is then attributable arithmetically to both the change of within-group ratios and their changed weights. A similar standardization on exposure units would estimate a different target quantity. The analyst should select the target in advance.
Uncertainty should be calculated at the level appropriate to the research question. Claim observations may be clustered by insured group, event, legal environment or reporting administrator. Aggregating individual claims as if they were independent can overstate precision. For an estimated ultimate, the variation from claim occurrence is compounded by parameter and process uncertainty during development, and by changes in coverage. An interval chosen from a model should state the assumptions and target. A confidence interval for an expected loss is not a prediction interval for the future realized total.
Where claim-level data remain inaccessible, the defensible alternative is a bounded descriptive study with explicit missing cells. A public population count can inform the institutional geography; a published ratio can inform the reported financial experience of its stated group. It is scientifically preferable to retain these two kinds of observations than to fill missing values through undocumented conversions.
Implications for offshore insurance institutions
Offshore insurance institutions redistribute the financial incidence of operational hazards across legal entities and contractual layers. They also redistribute the information needed to evaluate those hazards. The most significant research consequence is that the jurisdiction in which an insurer records a loss may be distant from the occurrence and from the enterprise ultimately affected by it. The interpretation of a reported loss therefore requires both the risk geography and the institutional geography.
Three distinct evaluations follow. An actuarial evaluation concerns the expected cost and variability of the obligations assumed by the captive. An entity-level financial evaluation concerns its underwriting, balance sheet and capacity to honor claims. An enterprise assessment concerns the costs and benefits of the total risk financing arrangement, including external risk transfer. These evaluations share data but answer distinct questions. A favorable entity ratio may coexist with important consolidated costs, while a high ratio may be consistent with the captive's intended financing function.
A comparison between domicile institutions remains useful at the level of licensing categories, availability of financial statistics and features of supervision. The present evidence identifies a difference in public observability between the selected jurisdictions. It does not establish a difference in the hazards or quality of the underlying portfolios. A regulatory evaluation requires its own evidence about supervision, compliance and outcomes.
The analytical contribution of this monograph is the joint interpretation of publicly reported underwriting indicators, the measurement boundaries that generate them, and established actuarial methods that could support more precise claims. It demonstrates how denominator shifts, class composition, claim development and contractual retention can change observed conclusions. These are applications and syntheses of established insurance and statistical reasoning, with no independent claim of a new theory or priority.
The practical implication is to make comparable exposure and valuation documentation integral to captive benchmarking. An insurer that increases its retention changes both the distribution of retained claims and the future capital and cash requirements. An enterprise that expands its captive into unfamiliar lines must evaluate the relevance of its prior experience. A supervisor examining heterogeneous classes must preserve the distribution of outcomes rather than relying on a single market ratio. Each implication follows from the measurement problem described in this investigation.
Conclusion
Offshore captive insurance loss experience is a product of insured hazards, contract allocation, actuarial estimation, premium funding, accounting and reporting. The captive's institutional location influences the organization of those processes while the underlying events often occur in other jurisdictions. Consistent analysis begins by separating the event from its insurance allocation and the period of occurrence from the period of recognition and payment.
The official Bermuda series establishes published average loss ratios between 48% and 57% over 2019–2023. Its contemporaneous combined ratios and class distinctions demonstrate that aggregate underwriting outcomes fluctuate and conceal heterogeneity. Cayman and Guernsey disclose institutional populations and financial measures that cannot, in the inspected public tables, be transformed into like-for-like captive loss ratios. The resulting documentary comparison is valuable precisely within these defined limits.
Established actuarial practice enables the analysis to progress from simple loss ratios to exposure-based costs, claim development, credibility, uncertainty and risk-adjusted comparisons. Those methods require data suited to their intended application. Sparse histories, heterogeneous layers and correlated losses demand particular care; apparent precision gained from a model does not cure structural incomparability.
The evidence supports an institutional interpretation of offshore loss experience, a bounded empirical description of the selected published observations and explicit conditions for more conclusive future comparative research. It supports neither a ranking of domicile risk quality nor a global estimate of captive losses. Preserving these evidentiary boundaries gives the subject its scientific integrity and makes subsequent research reproducible.
References
Bermuda Monetary Authority (BMA). (2025a). Bermuda Captive Insight Report – 2023 Year-end Data. Hamilton: BMA, uploaded December 4, 2025, pp. 6, 21. Source
Bermuda Monetary Authority (BMA). (2025b). 2023 Captive Report. Hamilton: BMA, uploaded March 24, 2025, p. 21. Source
Bermuda Monetary Authority (BMA). (2023). BMA Captive Report [2021 year-end data]. Hamilton: BMA, December 2023 upload, p. 20. Source
Cayman Islands Monetary Authority (CIMA). (2026). Cayman Islands Insurance Statistics. International Insurance Company by Risk Location, Primary Class of Business, Category and License Class, Q2 2026; Domestic Insurance Companies Audited Statistics 2024. George Town: CIMA. Source
Guernsey Financial Services Commission (GFSC). (2026). Statistics: International Insurers. Annual license statistics for December 31, 2025 and international insurance financial data through 2024. St Peter Port: GFSC. Source
International Association of Insurance Supervisors (IAIS). (2015). Application Paper on the Regulation and Supervision of Captive Insurers. Financial Stability Board Compendium, November 12, 2015. Source
Actuarial Standards Board (ASB). (2016). Actuarial Standard of Practice No. 23: Data Quality. Washington, DC: ASB. Adopted December 2016, effective April 30, 2017. Source
Actuarial Standards Board (ASB). (2013). Actuarial Standard of Practice No. 25: Credibility Procedures. Washington, DC: ASB. Adopted December 2013, effective May 1, 2014. Source
Actuarial Standards Board (ASB). (2007). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Washington, DC: ASB. Adopted June 2007. Source
Actuarial Standards Board (ASB). (2017). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. Washington, DC: ASB. Source
Mack, T. (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
England, P. D., & Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. doi:10.1017/S1357321700003809. Source
Open Actuarial Textbooks. (n.d.). Loss Data Analytics, second edition. Chapter 14: Loss Reserving. OpenActTexts. Source
International Accounting Standards Board (IASB). (2017, as amended). IFRS 17 Insurance Contracts. London: IFRS Foundation. Official standard overview. Source
AM Best. (2025). Best’s Market Segment Report: Captive Insurance and Alternative Risk Entities Continue to Emerge and Excel. Press release, August 4, 2025. Oldwick, NJ: AM Best. Source
Marsh. (2026). 2026 Captive Benchmarking Report. Public report landing page. Marsh. Source