Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Cyber Incident Exposure Denominators and Claims Data Bias

Abstract. Cyber insurance statistics are generated by a sequence of observation, insurance selection, contractual eligibility, reporting, adjudication and financial settlement. Exposure measurement must specify which population contributes time at risk and which event or claim increment belongs to that population. This monograph examines the consequences of using organizations, contracts, policy-years, digital assets, revenue, limits and shared technology dependencies as denominators. It integrates documentary comparison of insurer specimen wording, supervisory reports, incident surveys, cybersecurity reporting instruments and actuarial standards with algebraic analysis of selection, severity-dependent reporting, delayed development and correlated loss. United States insurance filings, United Kingdom survey estimates, European regulatory reporting and Australian and international incident-reporting evidence illustrate different selection mechanisms. A change in portfolio composition can alter aggregate claims frequency even when stratum-specific risk remains unchanged. A decrease in reported claims per policy in force can coexist with rising incident counts and higher economic losses. Neither a representative survey of recognized attacks nor a large collection of paid claims identifies the complete loss distribution. The scientific result is an operationally precise interpretation of complementary evidence: incident frequency, insured claim frequency, reported claim frequency, payment frequency and economic loss rates require different but reconcilable denominators. The analysis specifies conditions for credible comparison, loss development, pricing and accumulation measurement. It does not estimate a global incident rate or a fitted cyber loss distribution from unavailable microdata.

Keywords: Cyber insurance; exposure base; earned policy-years; incident prevalence; claims frequency; reporting bias; sample selection; truncation; censoring; loss development; cyber accumulation; actuarial data quality.

Introduction

The empirical study of cyber insurance begins with the relationship between a count and its opportunity for occurrence. A database containing ten thousand claims may support a study of the composition of observed claims, yet its size alone supplies no estimate of how frequently a covered organization experiences an incident. The missing quantity is not simply a technical inconvenience. It determines the scientific meaning of every inference about hazard, premium adequacy, prevention, protection gaps and market growth.

The research object encompasses three connected questions. First, what unit of exposure corresponds to the economic or contractual event under investigation? Second, through what mechanisms do detection, reporting, selection and settlement alter the distribution of recorded events? Third, what evidence and assumptions permit comparisons of different populations or periods? The object is narrower than general cyber threat intelligence and broader than a calculation of paid claims per policy. It is the construction and interpretation of the statistical population underlying insured cyber loss.

The study is an integrative documentary and analytical investigation. Sources were selected for relevance to the research questions: original supervisory reporting, official statistical methodology, published insurance wording, original regulatory texts, technical guidance and empirical scholarship. Public report summaries and indexed statutory passages are used only for the statements they establish; a gated report or unavailable microdata is not treated as inspected. The research is not an exhaustive systematic review, and no estimation from individual confidential claim files is asserted. The numerical scenarios explicitly labeled illustrative are independently reproducible calculations, not observations. United States, United Kingdom, European Union and Australian evidence is used to explain distinct sampling and legal mechanisms rather than to estimate a cross-national ranking.

The parent discipline is Cyber Insurance. The relevant unit of analysis is the risk borne by policyholders and transferred through coverage, including contingent liabilities and first-party disruption. Cybersecurity incidents occurring within an insurer's own operations belong to its operational risk unless a contractual or financial relationship makes them relevant to underwriting exposure. This distinction prevents an insurance supervisory indicator from being silently converted into a count of insured events.

Conceptual and Historical Foundations

A sequence of distinct events

An attack is adversarial activity directed toward a system or person. A security event is an occurrence relevant to a security process, potentially including an attempted or unsuccessful attack. A security incident is an event or connected sequence meeting an applicable incident definition. A breach is a narrower class under a particular technical, legal or empirical definition. Economic loss is damage valued in money within an identified accounting boundary. A claim is a request, notification or demand recorded under the applicable insurance process. Indemnity is a contractual payment, whose amount and eligibility depend on the policy and adjudication.

These categories are related without being interchangeable. An attempted intrusion may create only investigation expense. A service interruption may produce lost output without unauthorized disclosure. A privacy breach may create response expense immediately and civil liability years later. One compromised provider may generate hundreds of separate affected entities, with each generating several coverage transactions. Conversely, multiple malicious actions may be allocated to one claim under a contractual related-events clause. Event identifiers and claims identifiers therefore represent different objects.

The modern reporting problem reflects the development of several overlapping institutions. Technical teams classify compromise and respond to disruptions. Public authorities receive reports for regulatory supervision, consumer protection or law enforcement. Insurers administer claims and reimburse eligible losses. Researchers aggregate datasets to investigate trends. The origin and incentives of each institution affect what is observed. Nelson et al. (2025), in the official incident response guidance, locate detection and response within broader cybersecurity risk management. Increasing detection capability can consequently alter the observable incident rate even before the underlying exposure changes.

The estimand determines the exposure base

Let U designate a specified population, T the observation interval, N the number of events of the selected type, and E the corresponding measured exposure. A frequency estimate is N/E. Its physical and statistical units matter: 10 claims per 1,000 earned policy-years is a count rate, whereas 10 organizations with at least one breach per 1,000 organizations observed over a year is a prevalence proportion. Recurrent incidents make the difference material.

Under an illustrative homogeneous Poisson event model with annual intensity lambda, the expected annual event count per organization equals lambda, while the probability of at least one event equals 1 - exp(-lambda). At lambda = 0.5, the probability is about 39.35%, not 50%. For heterogeneous risks or correlated event processes, the simple conversion does not generally apply. A survey prevalence cannot be converted into claim frequency merely by renaming the denominator.

The ASOP No. 53 definition of exposure base explicitly permits variation across the cost components under investigation (Actuarial Standards Board, 2017, section 2.2). A suitable denominator supplies a measure of the opportunity for the particular outcome, remains consistent with the numerator, and can be observed with sufficient reliability. The unit that performs best in an insurer's pricing system may differ from the unit useful for security engineering or social welfare analysis.

Exposure units and their meanings

Exposure unit Appropriate outcome Interpretation and important limitation
Organization-year Incidents affecting organizations Captures population time; needs consistent entity identity and activity status.
Earned policy-year Contract-level reported or ultimate claims Requires coverage dates and trigger alignment; multiple policies may cover one entity.
Coverage-year Claims under a specified coverage Distinguishes coverage modules, endorsements and exclusions; requires effective terms.
Device-year or endpoint-year Device-associated event counts Depends on inventory and monitoring coverage; shared events are not independent.
Service-hour or workload-hour Interruption or service failure Relevant to operational availability, but ownership and dependencies matter.
Revenue-year Monetary loss relative to economic scale Normalization rather than entity event frequency; sensitive to currency and accounting.
Limit-year Insurance exposure scaled to contracted capacity Useful for certain risk analyses; limit is not expected loss or a physical hazard unit.
Provider-dependent entity-year Events affecting dependent entities Reveals common causes; exposure attributes require dependency mapping.

A policy-year is the integral of active covered time for a contract, expressed in years. Six months of otherwise comparable coverage contributes approximately one-half policy-year; cancellation and midterm adjustments change exposure. Policies in force at a reporting date are a stock, which cannot generally replace an integrated flow of earned exposure. With a stable book, the stock can approximate earned units, but rapid expansion, churn and policy layering invalidate that approximation.

An organization identifier should survive a change of insurer, broker or policy number without implying identical coverage. A contract identifier should survive accounting transactions without creating additional contracts. A coverage identifier should change when the corresponding insuring agreement or material terms change. A technical asset identifier must distinguish reused names, externally hosted workloads and retired systems. The denominator follows these histories rather than the number of rows in the latest inventory.

Romanosky's (2016) industry incident-rate analysis illustrates the interpretive importance of normalization: industry rankings by number of recorded incidents can differ from rankings by incidents divided by firms in the industry. Such a normalized rate still inherits the observation limits of the incident database. The distinction between numerator scaling and numerator completeness is fundamental.

Primary Documents and Scope of Exposure Unit Definition

Contract boundaries and at-risk time

Insurance exposure is legally structured. A risk-bearing entity may have several policies covering different perils, layer positions and periods. An insured group may contain several subsidiaries; a master policy may extend protection to specific additional insureds. An endorsement may provide restricted cyber coverage within another line. A policy count thus describes contracts, not necessarily independent insured organizations or equivalent units of protection.

Consider the AXIS cyber policy specimen (form AXIS 1012561 0120). Its claims-made liability section conditions coverage on the timing of claims against insureds and subsequent reporting. Several first-party grants instead turn on discovery during the policy period. Business interruption has a waiting-period condition, and applicable limits and retention shape payment. This form is a concrete example of the distinction between event date, discovery date, claim date, notice date and payment date. It is not a representative survey of the entire international market.

A single time series can therefore contain multiple contractual clocks. Under an occurrence-style analytical definition, the relevant risk interval is the period when the specified event took place. For claims-made liability, the contract-defined first-made trigger and retroactive provisions may govern how liability becomes attributable to the policy, while additional reporting requirements may also apply. First-party coverage may involve discovery. Allocating losses to underwriting year, accident year, report year or payment year results in different cohorts. Where alternative policy wording applies, the corresponding rules must be extracted from that wording.

A data dictionary for coverage

A minimum research record for contract exposure contains the insured entity, policy start, end and cancellation dates; policy form and endorsement version; territory; limits, sublimits, deductible or retention; waiting period; retroactive date if present; and layer attachment. Exposure should be segmented at a change of these conditions. A policy with distinct coverage sections can contribute one contract-year and several coverage-years, but the latter cannot simply be added to the former.

Coverage frequency for a particular first-party peril can be computed only within the contracts that grant the relevant cover. Suppose a book includes 500 policies with business interruption coverage and 500 without it, all active for a year. Ten business interruption claims among the covered policies yield 2 per 100 covered policy-years. Using the entire book gives 1 per 100 policy-years, a different quantity that combines hazard with coverage penetration. If the purpose is estimating portfolio-wide outgo, the second rate can be useful. If the purpose is comparing underwriting risk for the coverage, the first is appropriate.

The numerator also requires a definition. Count distinct root events when estimating the common shock process. Count affected entities when estimating transmission to insureds. Count claims when modeling claim administration and indemnity. Count transactions when reconciling accounting data. Counts of claims reported with no payment remain meaningful for operational workload and coverage disputes; filtering them out creates another outcome.

Supervisory classifications

The NAIC 2025 market report documents a material reporting change effective for 2024 filings: the cyber reporting categories were restructured into primary, excess and endorsement coverage, and identity theft reporting was removed from that supplement's requirements. It also identifies claims, premium, loss and policy stock fields. This change illustrates why time series need form-version metadata and a bridging table. A taxonomy transition can create a discontinuity even if the insured population's security position is unchanged.

The appropriate data structure records source population and filing basis for each measure. Domestic insurer policy counts cannot be silently combined with a premium numerator that includes alien surplus lines. Market measures for direct insurance should not be mixed with ceded or net reinsurance values. Similarly, earned premiums are a financial exposure measure affected by market prices, unlike time-based policy exposure. A ratio of loss to earned premium estimates a different business statistic from claim count per policy-year.

Reporting regimes and their populations

European and American reporting requirements illustrate the legal construction of observational samples. Under GDPR Article 33, a controller generally notifies the competent authority of a personal data breach, where feasible within 72 hours of awareness, subject to the specified risk exception. The population consists of breaches within the regulation's scope and threshold, not every cybersecurity compromise. The obligation is associated with controllers and protected personal data.

Under NIS2 Article 23, essential and important entities report significant incidents through the designated authorities, with an incident notification generally required within 72 hours of awareness. Domestic implementation, scope and service-specific exceptions must be researched for a particular country and date. DORA Article 19 establishes reporting of major ICT-related incidents for financial entities within its own legal framework. A DORA report may describe an insurer's operational problem, not a claim against a cyber policy.

In the United States, the SEC's 2023 adopted cybersecurity disclosure rules generally require domestic registrants to file material cybersecurity incident information within four business days after determining materiality. This clock is distinct from an event-occurrence clock, a privacy-awareness clock and a claim-notification clock. These legal rules are compared at the level of the sourced provisions, without asserting national transposition details or providing compliance advice.

The same hypothetical breach could enter a privacy regulator's records, a listed entity's disclosure, a security-response center's file and the insurer's claim system. Deduplicating these records requires evidence of identity, timing, affected entities and event relationships. Treating every report as a distinct incident produces double counting. Treating all reports as interchangeable discards precisely the institutional information needed to study reporting selection.

Mechanisms and Evidence for Reporting Bias

Observation and missing events

Let lambda(x,t) be an underlying event intensity per organization-time with attributes x. Let p_D be the conditional probability of discovery, p_E the conditional probability that a discovered event meets the selected insurance eligibility definition, p_R the conditional probability that an eligible discovered event is reported within the observation horizon, and p_K the conditional probability that it is retained by the database. For a population restricted to insured entities, the product q = p_D p_E p_R p_K represents a simplified observation probability, with each probability conditional on the previous stages. For a one-event-one-record convention, the expected observed count per unit exposure is lambda q. This is an application of standard conditional probability, not a fitted structural model.

A given observed rate is consistent with many values of lambda and q. For two homogeneous books with 5 observed claims per 1,000 organization-years, an observation probability of 0.5 implies 10 target events per 1,000 organization-years; a probability of 0.25 implies 20. Without external information about q, a point estimate of the underlying rate is not identified. When an incident can generate multiple claims, the formula needs a further conditional expected multiplicity term. If the dataset includes false positives, a separate false-positive intensity is also necessary.

This matters when evaluating security controls. An insurer may introduce endpoint monitoring and simultaneously encourage early notifications. Detected incidents may rise even if controls prevent some actual losses. Claim severity may fall as response begins earlier. Interpreting the increased observed count as evidence of worsening security, without measuring detection changes, can incorrectly penalize the intervention.

The UK breaches survey 2025/2026 explicitly limits its estimates to attacks and breaches that respondents recognize and disclose (DSIT and Home Office, 2026). This is observation bias within a survey frame, distinct from the question of whether surveyed organizations hold insurance. A probability sample can reduce sampling error while leaving under-detection unresolved.

Thresholds, censoring and the severity distribution

Insurance terms influence observed loss severity through economic incentives and payment boundaries. A loss below a retention may never enter the formal claims database; the researcher may see only cases for which an insured anticipates some reimbursement or assistance. A reported claim below the retention may nevertheless appear with zero indemnity. A paid-only extract discards it. These are three different mechanisms: unobserved cases, observed zero-paid cases and filtering by the analyst.

For a simplified indemnity contract, let gross eligible loss be X, deductible d, and payment limit l. Payment Y is min(max(X-d,0),l). The approximation assumes a single eligible dollar loss, no coinsurance, no expense treatment, no reinstatement and no overlapping coverage. A paid claim at the limit establishes X at least d+l under these assumptions; it does not establish the exact underlying loss.

For a continuous loss distribution F and density f, an extract restricted to payments above zero is conditionally truncated at d. For uncapped positive payments y less than l, the conditional density of Y given positive payment is f(y+d)/(1-F(d)). The mass at l equals P(X >= d+l)/P(X > d). For a complete exposure portfolio, zero payments can result both from no event and from events below or outside coverage. Estimating a ground-up severity distribution from conditional paid data therefore requires the original terms and a model of inclusion. Without them, tail exposure may be underestimated and observed averages distorted.

An illustrative sample has 100 equally weighted eligible loss events: 70 at $1,000 and 30 at $10,000. The average ground-up loss is $3,700. If d = $2,000 and l = $5,000, 70 events produce no payment and the other 30 each produce $5,000. The average payment over all events is $1,500; the average over paid events is $5,000. Both can be calculated exactly. Neither is the original gross severity. The example isolates deductible and limit effects; actual contracts may additionally include response services or claim expenses.

Timing and loss development

A recent accident cohort typically has had less time to mature than an old cohort. Let N(a,u) be the accumulated claims reported for accident period a by development age u. If reporting has a stable distribution G(u) and the eligible ultimate expected number is m(a), then E[N(a,u)] = m(a)G(u) under the simplifying assumptions of a common report-delay distribution and complete eventual reporting. A count development factor 1/G(u) can then convert an immature expected count to an ultimate count. It does not recover incidents that never generate a claim.

Cyber delays can arise from postponed discovery, legal proceedings, evolving business interruption assessments, or the time to identify related affected organizations. The AXIS specimen's notice of circumstances clause gives an example of a later claim attributed to an earlier reporting context under contractual conditions. A delay triangle should therefore store the event, discovery and claim dates separately. An underwriting-year triangle may develop differently from an incident-year triangle; changes in reporting behavior or coverage can invalidate historical factors.

For every valuation date, distinguish reported but unsettled claims from incurred but not reported claims. Payments to date and case estimates measure financial development; additions to claim counts measure reporting development. Reopened claims and revised event classifications require versioned transactions. A historical dataset reconstructed from the most recent claim state will leak future information if used to simulate an earlier underwriting decision. Prospective back-testing requires as-of snapshots.

Error, duplication and reporting incentives

Claims may be recorded for security-response assistance, reassurance or precaution, even where final coverage is uncertain. In contrast, organizations may prefer internally financed recovery when expected payment is low or when disclosure carries legal or reputational costs. Both mechanisms alter the composition of reported cases. A change in claim usage may reflect contractual service changes and insured preferences as well as any shift in physical cyber hazard.

The 2024 data in NAIC Figure 9 distinguish 9,941 claims closed with payment from 28,555 closed without payment. These counts concern the closed-claim population and do not equal ultimate claims from a matched 2024 exposure cohort. Zero payment is a genuine administrative outcome rather than proof that no incident occurred. It may result from coverage, layer attachment, withdrawal or another administrative process; the aggregate figures do not identify the distribution of those reasons.

The correct analysis preserves original status codes and defines a mutually exclusive terminal classification only after reconciliation. Unknown payment status must remain unknown rather than be converted to zero. A withdrawn notice must remain distinguishable from a denied claim. Reducing the database to paid cases may support conditional payment modeling but cannot support unbiased inference about reporting probability without additional data.

Comparative Measurement of Selection Effects

Distinct observation frames

Evidence sources have different sampling frames, institutional objectives and measurement limits. A dataset can be authoritative within its frame while unsuitable for estimating risk outside it.

Evidence source Observational unit Appropriate use Missing denominator or constraint
NAIC Cyber Supplement Insurer-reported contracts and claims U.S. reported insurance experience Earned exposure and consistent entity identity need supplementary records.
UK Breaches Survey Surveyed organizations Weighted recognized-attack prevalence Unknown latent attacks; not a contract claims census.
Verizon DBIR Contributed incidents and breaches Characteristics of observed events Comparable base of all organizations generating cases not established by overview.
Coalition claims overview Carrier portfolio claims Study of that portfolio Full earned exposure, selection rules and definitions not accessible here.
NetDiligence issuer announcement Contributed claim cases Case-count and sample-scope description Nonclaim exposure frame and full microdata unavailable.
Australian reporting records Disclosed cybercrime and incidents Reporting-system evidence Unreported cases and exposure base.
European insurance dashboard Prudential and other indicators Sector risk monitoring Not a cohort of covered cyber incidents.

Verizon's (2026) DBIR publication overview identifies a multi-contributor collection and an incident window extending from November 2024 through October 2025. The different contributing institutions mean the analyst must establish source and inclusion rules before generalizing. Coalition's (2026) claims report overview describes claims across more than 100,000 policyholders in calendar 2025. That disclosure identifies the portfolio scale; it does not itself supply a precise earned policy-year count. NetDiligence's (2026) claims study announcement describes 10,309 claims relating to incidents in 2021-2025. That is a case collection, not the associated population of insureds at risk.

ENISA's (2026) Threat Landscape publication summary specifies observed incidents and events during calendar 2025. The underlying report could not be directly retrieved; only that observation window is used. EIOPA's (2026) April Insurance Risk Dashboard draws on prudential returns and indicators with different reference dates, including quarterly end-2025 measures. Its population is supervised insurance entities and indicator sources. Australian Signals Directorate (2026) explicitly acknowledges incomplete threat representation in its annual report notes. The latter evidence is limited to the official indexed excerpt available.

These sources can triangulate qualitative mechanisms and constrain what an analyst may conclude. Their counts must not be added. Populations overlap and may contain insured and uninsured victims, organizations and individuals, different event categories, repeated observations and disparate time intervals.

An empirical denominator exercise

The 2026 NAIC market report states that domestic policies in force increased from 4,368,614 in 2024 to 5,843,175 in 2025, while reported claims increased by approximately 20% to 56,323. Dividing 56,323 by 5,843,175 gives about 9.64 reported claims per thousand policies in force. With a policy stock increase of approximately 33.75%, dividing the two reported relative growth factors, 1.20/1.3375, suggests a reduction of approximately 10.3% in the aggregate claims-to-policy-stock ratio relative to the prior year.

This calculation is expressly a stock-normalized indicator, not earned policy-year frequency. The claim growth percentage is rounded, so the change in the ratio is approximate. Rapid growth in endorsements, the changing share of primary and excess coverage, differences in policy inception times and report lag are plausible contributors. The data alone do not estimate their separate effects. An inference that technical hazards improved would require matched entity and coverage exposure, contractual terms and stable observation or adjustment for its variation.

The UK survey, by contrast, estimates that 43% of businesses and 28% of charities identified a breach or attack in the preceding 12 months. Its 2025/2026 technical report describes random-probability sampling and a change to business-region weights following regional oversampling. This is a weighted organization-level measure. Comparing it numerically with the NAIC ratio would confuse prevalence with claim counts, the UK population with the U.S. insured market, repeated events with distinct organizations, and attack detection with contract reporting.

Portfolio composition and standardization

Insurance purchase, risk acceptance and renewal are selective. Insured organizations may have higher awareness, more digital assets, better controls, greater exposure or greater incentives to obtain coverage than uninsured organizations. Insurer acceptance policies may further change the risk distribution. The direction of the resulting bias relative to the economic population cannot be assumed solely from insurance status.

Let f_s(t) be claims per earned unit in stratum s at time t, and w_s(t) the share of earned exposure in that stratum. The aggregate rate is the sum over s of w_s(t) f_s(t). The change between t = 0 and t = 1 can be written exactly as the sum of w_s(0)[f_s(1)-f_s(0)] plus the sum of [w_s(1)-w_s(0)]f_s(1). This decomposition uses base-period weights for the within-stratum change and end-period rates for the composition change; other reference choices redistribute the interaction term.

In a synthetic book, low-risk contracts yield 2 claims per 100 policy-years, while high-risk contracts yield 10. Initially the insurer writes 800 low-risk and 200 high-risk policy-years, yielding 16+20=36 expected claims, or 3.6 per 100. In the next period it writes 400 low-risk and 600 high-risk policy-years, yielding 8+60=68 expected claims, or 6.8 per 100. Stratum-specific frequencies are unchanged. Standardizing both years to the earlier 80:20 exposure weights restores the identical 3.6 rate. The increase is a change in composition, not deterioration of either stratum.

Observed differences may also reflect measured covariates that conceal unmeasured factors. A standardization approach assumes adequate comparability of strata. Inverse-probability weighting of insured samples to a target population requires estimable selection probabilities, positive overlap and a defensible conditional exchangeability assumption. For organizations outside the support of the insurer's underwriting book, weighting cannot recover empirical evidence.

Selection on a claim outcome

Suppose an unobserved event severity affects both the probability of insurance notification and the potential outcome, while insurance retention also affects notification. Conditioning on reported claims can induce an association between retention and loss even where retention has no causal effect on underlying severity. The effect is generated by inclusion in the dataset, not necessarily by behavior after insurance purchase.

Researchers investigating moral hazard or the effectiveness of insurer-provided mitigation must separate prebinding selection, postbinding changes in behavior, changes in detection and claims reporting. Restricting analysis to organizations with positive losses can obscure the effect of controls that prevent all losses. A cohort of eligible policyholders including zero claims is usually necessary for studying frequency. A study of conditional claim cost may remain useful with claim-only data if its scope is stated.

Sensitivity analysis should vary reporting and capture probabilities explicitly. If an observed one-record rate is 0.01 per organization-year and q is bounded by an independently justified interval [0.4,0.8], the underlying rate lies between 0.0125 and 0.025 under the simplified framework. The range is identified by the external probability assumptions, not by the insurance sample alone. Without a justified positive lower bound on q, no finite upper bound for lambda follows from the observed rate alone.

Integrated Scientific Synthesis and Research Agenda

Reconciliation across levels of observation

The evidence supports an integrated interpretation based on several connected measurements. A threat process generates actions and service failures. Technical monitoring and reporting institutions determine observed incidents. Insurance purchase and underwriting specify the covered population. Contracts determine the incidents, discoveries or claims that can trigger coverage. Claim administration defines the recorded numerator, and settlement converts eligible amounts into financial transactions.

This synthesis has a concrete consequence: a valid estimate at one stage cannot be upgraded into a claim about another stage without a justified mapping. The UK recognized-attack prevalence, NAIC claims-to-policy-stock ratio and carrier claim-type composition each have recognizable empirical meaning. They remain informative precisely when their denominators and boundaries are retained. Their divergence can suggest hypotheses about selection and exposure composition, but does not by itself establish any specific causal mechanism.

A researcher assessing whether cybersecurity across an insured population has improved should therefore ask for three distinct data products: a longitudinal insured exposure table including zero claims; an incident table linking discovered events to eligible and ineligible insurance outcomes; and a claim and payment history organized by development age and contract version. For common shocks, an additional root event and dependency table is necessary. No universal numeric adjustment factor can replace missing links between these objects.

Minimum verification and reproducibility

A practical analytical dataset uses immutable source snapshots and records the unit, risk population, jurisdiction, interval, valuation date, event definition, coverage form, status and reporting origin. The exposure table must preserve entities entering and leaving the cohort. Event tables preserve source IDs, uncertain identity matches and root-event relationships. Claim tables maintain report date, adjudication status, estimated ultimate, payments, recoveries and appropriate allocation keys.

Before analyzing a rate, verify numerator units against exposure. A claim attached to a canceled policy may be valid under an extended reporting rule; this is a contract question, not necessarily a data error. A record associated with two active policies may represent valid layered cover rather than duplication. A record with no matching policy may indicate scope mismatch or data integration error. These categories should be adjudicated rather than deleted by generic deduplication.

Document missing data at the level at which missingness arises. Missing loss estimates differ from absent incidents. Missing exposure time differs from zero exposure. Records without a final decision differ from rejected coverage. A reproducible report lists the number of included and excluded records by reason, matches numerator and denominator on the relevant keys, retains an as-of snapshot and tests alternative reasonable assumptions.

A sample frame should be evaluated against independently collected information on the target population. Similar distribution of revenue or sector does not imply similar security controls or reporting behavior. Explicit comparison by insurer, policy structure, insured size and renewal cohort can reveal divergences. If a certain subgroup is entirely absent, report lack of support rather than obtaining spurious precision through extrapolation.

Research priorities

First, empirical exposure studies require acquisition of earned contract-time and organization-time, with harmonized entity identities and coverage history. Policy stocks should remain useful for market monitoring, while researchers seeking hazard rates should supplement them with transaction and exposure-flow data.

Second, detection and reporting should be studied jointly. A study could pair confidential incident histories with insurance notification records in a defined cohort, investigate reasons for nonnotification, and evaluate how retention levels and response services affect capture. Such a study would need appropriate governance and a design addressing consent and follow-up. This monograph has not obtained that dataset, and no estimate of latent reporting probability is asserted.

Third, loss development must be examined across policy and event cohorts, with coverage and risk changes recorded explicitly. Separating discovery delay from claim-report delay and financial settlement delay can reveal which assumptions transport across periods. A stable historical triangle is evidence about the book and practices comprising it, not a universal cyber development curve.

Fourth, shared-service exposures should support accumulation measurement at the root-event, affected-entity, contract and layer levels. Retaining common shock identifiers permits a scientifically meaningful distinction between frequency of initiating events and frequency of resulting claims. It also permits conditional stress scenarios in which coverage and concentration are explicitly specified.

Fifth, comparative policy research can connect disclosure threshold changes with reporting changes using suitable pre/post observations and comparison populations. Legal changes may be accompanied by shifts in cybersecurity, market participation, enforcement or classification, all of which can confound simple before-after comparisons. The design should test alternative explanations and avoid causal claims where the necessary assumptions are unsupported.

Boundaries of the present synthesis

This investigation establishes measurement distinctions, contractual mechanisms and transparent mathematical implications. It does not fit a regression, estimate causal effects or provide a representative global dataset. Publicly disclosed aggregates and specimen forms are sufficient for the documentary and logical comparisons undertaken; carrier microdata, exact capture probabilities, country-specific implementation studies and full treaty texts would be required for additional empirical tasks. Source access varied, and claims from issuer summaries or statutory excerpts remain limited to their checked content. No named new theory or claim of priority is made.

Conclusion

Cyber incident exposure denominators determine the meaning of claims statistics. An insurance claim count is generated within an institutionally selected, contractually bounded and temporally developing population. A technical attack count describes a different observation process. Insurance coverage, monitoring, notification thresholds, retention, limits, event aggregation, administrative practice and data-source coverage shape what a researcher can observe.

The comparative evidence demonstrates why market expansion and reported claims must be analyzed together, and why even policy-stock normalization does not resolve earned exposure or changing coverage composition. The mathematical examples show how a stable stratum-specific rate can coexist with large aggregate changes, how truncation and payment caps transform severity, and how a single correlated shock can generate many claims. These are consequences of established actuarial and statistical reasoning applied to cyber insurance evidence.

Sound research connects each numerator to its population, observation interval, contractual trigger, reporting horizon and valuation basis. It states the selection assumptions required for extrapolation, preserves zero and unknown outcomes as separate categories, and distinguishes model-based uncertainty from uncertainty about unobserved cases. Market evidence, incident surveys and prudential reports have complementary roles when used within their actual scope. Integration requires explicit reconciliation, rather than substitution of a convenient measure for an unavailable one.

References

  • National Association of Insurance Commissioners (NAIC). 2025. Report on the Cybersecurity Insurance Market. 2024 reporting year. Source

  • National Association of Insurance Commissioners (NAIC). 2026. Report on the Cybersecurity Insurance Market. 2025 reporting year. Source

  • Department for Science, Innovation and Technology and Home Office. 2026. Cyber Security Breaches Survey 2025/2026. Source

  • Department for Science, Innovation and Technology and Home Office. 2026. Cyber Security Breaches Survey 2025/2026: Technical Report. Source

  • Romanosky, Sasha. 2016. Examining the Costs and Causes of Cyber Incidents. Journal of Cybersecurity 2 (2): 121-135. DOI: 10.1093/cybsec/tyw001. Source

  • AXIS. 2020 form edition. AXIS Cyber Insurance Policy. Specimen form AXIS 1012561 0120. Source

  • Actuarial Standards Board. 2016. Actuarial Standard of Practice No. 23: Data Quality. December 2016 revision. 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

  • European Parliament and Council. 2016. Regulation (EU) 2016/679 (General Data Protection Regulation). OJ L 119, May 4, 2016, 1-88. Source

  • European Parliament and Council. 2022. Directive (EU) 2022/2555 (NIS2 Directive). OJ L 333, December 27, 2022, 80-152. Source

  • European Parliament and Council. 2022. Regulation (EU) 2022/2554 on Digital Operational Resilience for the Financial Sector (DORA). OJ L 333, December 27, 2022, 1-79. Source

  • U.S. Securities and Exchange Commission. 2023. Cybersecurity Risk Management, Strategy, Governance, and Incident Disclosure by Public Companies. Press release 2023-139, July 26, 2023. Source

  • Nelson, Alexander, Sanjay Rekhi, Murugiah Souppaya, and Karen Scarfone. 2025. Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile. NIST SP 800-61 Rev. 3. DOI: 10.6028/NIST.SP.800-61r3. Source

  • Verizon. 2026. Data Breach Investigations Report: Official Publication Overview, 2026 edition. Source

  • Coalition. 2026. 2026 Cyber Claims Report: Official Report Overview. Source

  • European Insurance and Occupational Pensions Authority (EIOPA). 2026. April 2026 Insurance Risk Dashboard. EIOPA-BoS-26-170. Source

  • Australian Signals Directorate, Australian Cyber Security Centre. 2026. Annual Cyber Threat Report 2025-26: Notes. Source

  • NetDiligence. 2026. NetDiligence Publishes Sixteenth Annual Cyber Claims Study. September 15, 2026. Source

  • European Union Agency for Cybersecurity (ENISA). 2026. ENISA Threat Landscape 2026. September 22, 2026. Source