RESEARCH DOSSIER
Ransomware Incident Frequency Severity and Victim Selection
Abstract. Ransomware imposes a distribution of financial and operational consequences on organizations exposed to malicious digital extortion. Frequency, victim selection, and severity are mutually dependent objects of insurance research because the observed victim population emerges through exposure, compromise, detection, disclosure, contract eligibility, and claim development. This monograph critically compares primary public statistics, insurance market reporting, specimen wording, security guidance, and original empirical studies. It distinguishes population incidence from complaint counts, the conditional probability of payment from positive payment severity, and economic damage from covered and ultimately indemnified loss. Explicit actuarial examples demonstrate reporting bias, portfolio composition effects, truncation, policy-layer transformations, and dependence. Evidence from the United States, United Kingdom, European Union, and Dutch incident research is assessed within its observation frame. The synthesis establishes why a reliable insurer estimate requires jointly specified exposure denominators, loss components, policy terms, observation mechanisms, and uncertainty; readily available aggregate indicators cannot by themselves identify global ransomware incidence or causal selection effects. No newly estimated empirical distribution or original named theory is asserted.
Keywords: Ransomware; cyber insurance; incident frequency; victim selection; severity distribution; extortion; exposure denominators; reporting bias; hurdle models; business interruption; indemnity; censoring; accumulation; cyber risk; actuarial science.
Introduction
Ransomware incident frequency measures the occurrence of a defined event during a specified period of exposure. Severity measures the magnitude of a specified consequence conditional on an event. Victim selection concerns the processes through which organizations enter the affected population and the subsequent samples through which researchers observe that population. The three objects meet in insurance because indemnity depends on what happens, to whom, and under which contractual conditions. A ransomware attack that causes substantial operational interruption can generate a large claim even when no extortion payment is made. An incident that generates no covered loss may remain absent from claims statistics while contributing to security incident frequency.
The parent Science, Cyber Insurance, establishes the broader subject as contractual transfer and financing of defined digital risks (Bogdanova, 2026). The present investigation concentrates on the empirical architecture of ransomware occurrence and financial consequences, leaving detailed negotiation procedures, sanctions-screening practice, and the construction of individual claim reserves to their respective dossiers. Its analytical unit is the organization at risk, with explicit distinctions between an enterprise, operating facility, policy, incident, claim, and common originating event. Those distinctions are necessary in a market in which several entities may share one digital dependency and in which insurance arrangements may comprise successive layers.
The study uses a structured, non-exhaustive review of accessible original research and institutional documents. The cutoff is November 18, 2026. Source authority varies by question: a regulatory market report supplies market measurements; a specimen clause illustrates contract mechanics; a clinical cohort study documents operational consequences; a police complaint system supplies recorded complaints. The material was examined at the level specified in the evidence register. Selected documents were available as full HTML, others as institutional abstracts, official releases, or extracted PDF text. The study does not claim systematic-review completeness, independent access to insurance microdata, or causal identification from reported associations. Numerical examples are hypothetical except where expressly attributed to published observations.
The research questions concern which frequencies can be identified from available evidence, how victim and reporting selection affect interpretations, and how severity changes when assessed as an economic, contractual, actuarial, or social quantity. The analysis proceeds from definitions to observation, mechanisms, loss measurement, institutional implications, and a research agenda.
Conceptual and Historical Foundations
Defining the event and the exposure
Ransomware is a form of digital extortion in which malicious interference with information systems or threats concerning data create demands for payment or other concessions. The empirical event definition for this study includes incidents identified as ransomware or related digital extortion according to the source's own classification, with a separate field to indicate encryption, exfiltration, disruption, and confirmed demand. Encryption and disclosure threats can coexist. Some malicious operations produce ransomware-like disruption without a reliably established financial objective. A classification should preserve uncertainty rather than retrospectively impose a single intention on all events.
An attempted intrusion is an adversarial event, a detected compromise is a security event, and a material disruption is an operational event. A reported incident is an administrative or statistical record; an insurance claim is a request under a contract. A paid claim represents a settlement or payment transaction, not a count of all affected organizations. Multiple claim payments may relate to one demand, and several policies may cover different consequences of the same incident. An organization can suffer multiple distinct incidents in a year; a campaign can affect many organizations.
The insured exposure unit is ordinarily a policy-year or entity-year, adjusted for time at risk, depending on the intended prediction. An entity covered for six months contributes one-half of an entity-year under a constant-exposure convention. Mergers, cancellations, coverage inception, temporary suspension, and changes in digital operations complicate this convention. A portfolio with fixed annual policy counts may nevertheless change in its concentration of insured revenue, shared providers, and limits. Exposure measurement therefore records both the count and the nature of the units.
Security data and insurance data answer different questions. The official updated ransomware guidance announcement (CISA et al., 2023) organizes risk reduction around prevention, detection, response, and restoration. Those activities help define observable stages of the incident process. They do not provide a statistical mapping from a control to an annual probability of insured loss. A security framework is evidence of a recommended practice, while a calibrated rate requires an observation design.
The changing economic context
Digital extortion is an adaptive enterprise. The loss mechanism evolves as security measures, restoration capacity, technology availability, incident reporting, law enforcement, and attacker behavior change. Historical samples consequently combine different technological and institutional regimes. The same label can refer to an isolated device incident or a multi-facility business disruption, and the financial distributions of the two differ.
The ENISA 2026 release documents the agency's review of events observed during calendar year 2025 and emphasizes the role of interconnected services and infrastructure (ENISA, 2026). Its source collection includes open information and anonymized contributions. This evidence supports an institutional concern with dependencies, while leaving the absolute population incidence outside its design. Event composition within a threat landscape should not be treated as the probability of compromise for a randomly selected business.
The history of cyber loss measurement shows the importance of the denominator. Romanosky's (2016) study of cyber incidents separates event volume, sector rates, litigation, and costs. An industry that produces many incidents in a database may simply contain many firms or have stronger disclosure practices. The older study addresses broader cyber risks and supplies a methodological antecedent rather than present-day ransomware calibration.
Units of aggregation
Let a common originating event be an incident identifier E; multiple organizations may be affected by E, and each organization may possess several policies. The frequency of origins, affected organizations, and payable claims can thus diverge. For enterprise risk management, the relevant quantity may be disruption episodes at the consolidated organization. For a primary insurer, the unit may be a covered claim under the wording. For a reinsurer, an occurrence definition or aggregation clause may organize losses under yet another unit.
Consider a hypothetical incident affecting three related companies, each with a primary policy and separate excess protection. An origin count of one, an entity count of three, and six insurance records can describe precisely the same underlying occurrence. Combining them without a linkage rule overstates event frequency; collapsing them into one record obscures portfolio exposure. A consistent dataset stores the origin identifier separately from the affected-entity identifier, coverage identifier, and payment identifier.
An annual claim-frequency table also needs a consistent clock. Attack initiation, organizational detection, notification to police, reporting to an insurer, coverage acceptance, and settlement may span different calendar periods. Recent incident cohorts are often incomplete; a calendar-year administrative count may contain earlier occurrences. This is a structural reason for divergence between changing claim counts and changes in the underlying threat process.
Primary Documents and Scope of Attack Frequency
Complaint observations
The FBI's 2025 IC3 Annual Report (2026) records 3,611 ransomware complaints in 2025, compared with 3,156 in 2024 and 2,825 in 2023. The 2025 section describes reported losses above USD 32 million and explains the omission of many categories of consequential costs, some unreported loss amounts, and incidents reported directly to field offices rather than via IC3. These are administrative data supplied through a particular reporting channel, with no complete denominator of organizations at risk.
The annual growth in the 2025 complaint count relative to 2024 is approximately 14.4%, calculated as 455 divided by 3,156. That is growth in recorded complaints. A proportional increase in true incidence would follow only if the effective recording probability and exposure remained constant and classification and duplication were sufficiently stable. The report establishes a useful indicator of reported activity and official workload, not a global estimate of annual organizational ransomware probability.
An organization that learns of a compromise after the attack occurred could report it in a later period. Another organization may report through an intermediary rather than the complaint channel. A record with missing loss data remains an event observation; replacing its loss by zero imposes a factual assertion unsupported by missingness. Both frequency and severity analyses must explicitly handle this distinction.
Survey observations
The British government breaches survey for 2025/2026 (DSIT and Home Office, 2026) reports that 43% of businesses identified some cyber security breach or attack in the preceding twelve months. Among those businesses that identified a breach or attack, 3% identified ransomware, compared with 6% in the preceding comparable survey. These proportions are conditional survey estimates. Multiplying the rounded 43% and 3% produces an approximate 1.29% of businesses in the eligible survey population; this is an algebraic illustration based on rounded values, not an official standalone estimate or a count of incidents.
The corresponding survey technical report states that quantitative fieldwork took place from August through December 2025 and included 2,112 businesses, with weighting intended to represent the eligible population. Sole traders, most public bodies, and businesses lacking an IT or online presence fall outside the specified sampling frame. The survey is closer to organization-level prevalence than a complaint count, but identification is still conditional on respondents' awareness and recollection. A business that experiences several events is counted once in an at-least-one-event prevalence measure.
If independent events follow a Poisson distribution with mean annual count lambda, the probability of at least one event is 1 minus exp(-lambda). A 1% prevalence would imply lambda = -ln(0.99), about 0.01005, under homogeneous exposure and complete detection. Heterogeneous intensities invalidate the direct inversion for the average rate: the average of exp(-lambda_i) is not exp(-average lambda_i). This illustrates why an empirical prevalence estimate cannot automatically calibrate a claim-frequency model.
Market and portfolio observations
The NAIC's 2026 insurance market report gives approximately USD 9.46 billion in U.S. direct written cyber premium for calendar 2025, including alien surplus lines, against USD 9.14 billion in 2024 (NAIC, 2026, p. 2). The series reflects premium volume across cyber insurance coverages. It provides market context but neither ransomware exposure-years nor ransomware-specific incident counts. An increase in premium volume can reflect changing prices, participation, limits, and business mix.
In contrast, Coalition's (2026) cyber claims report announcement describes observations from its insured business. In its 2025 results, the initial ransomware demands increased by 47% year over year, while overall global cyber claim severity fell by 19%. Ransomware had a reported mean claim loss of GBP 202,000 in the UK announcement. These statements refer to different metrics in a selected insurance portfolio. Their simultaneous occurrence establishes that trends in demands and all-peril insurance severity need not run in the same direction; it does not establish a national ransom-payment rate or market-wide ransomware severity.
Portfolio estimates require exposure periods, consistent peril assignments, attachment points, and a record of changes in underwriting. A carrier that reduces high-risk business or modifies its deductible can record lower observed claim frequency without a corresponding change in population incident hazards. Premium and policy counts should be analyzed with these transformations in mind.
Statistical observation
Suppose N is the unobserved number of qualifying incidents and R the number observed using a defined channel. In a homogeneous hypothetical setting, N is Poisson with mean e times lambda, where e is exposure, and each incident is recorded independently with probability q. Then R has mean e times lambda times q. The observable rate estimates the product lambda q; without an additional source of information, lambda and q are not separately identified.
For example, take 10,000 organization-years and 80 recorded incidents. The observed rate is 0.008 per organization-year. If q is 0.4, the implied underlying rate is 0.02; if q is 0.8, it is 0.01. Both reproduce the expected record count. The scenario is useful as sensitivity analysis and not as an estimate of actual reporting probabilities. If q varies by loss size, sector, contractual terms, and period, the relationship must be integrated over those characteristics.
A frequency model using a Poisson distribution can include log exposure as an offset and covariates representing measured characteristics. A negative binomial specification can accommodate overdispersion from heterogeneity, provided its interpretation is consistent with the observation process. Recurrent-event methods allow time-varying covariates and repeated incidents. A mixture or random-effects formulation allows entity-specific variation. None repairs an absent denominator or selective recording automatically.
For incomplete recent periods, counts should be organized by occurrence and reporting dates. Reporting-delay models estimate the fraction expected to emerge by a given valuation date. Delayed notifications and revised event classification require versioned data. Model performance should be examined across out-of-time validation windows, because adaptation may make a historical calibration inappropriate even when the historical fit is excellent.
Mechanisms and Evidence for Victim Selection
Selection into incidents and records
Victim selection is the observed outcome of adversarial exposure, protective capability, compromise, operational consequences, and measurement. A research design may distinguish exposure to adversarial activity from the probability of successful compromise and the probability of detecting that compromise. Subsequent observation is mediated by organizations' choices and obligations. Selection into insurance adds a further stage: coverage is purchased or declined, and an insurer accepts, modifies, or rejects the risk. The insured population is therefore produced by economic and institutional decisions before the incident occurs.
The study treats these processes as distinct scientific questions. Describing the distribution of observed victims by size or sector does not identify the adversary's preferences. A large sector may generate more cases simply because it contains more exposed organizations. Within a sector, greater detection resources and formal reporting can increase observation probability. Severity-dependent notification also means that the observed case mix may overrepresent disruptive incidents.
The NIST Cybersecurity Framework 2.0 (Pascoe, Quinn, and Scarfone, 2024) provides an outcome-oriented taxonomy for assessing and communicating cybersecurity risk. Its role here is to support comparable measurement of protective and recovery capability. A binary assertion that an organization follows the framework is insufficient evidence that any specified outcome is implemented or effective. An actuarial dataset should use validated observations of controls, dates, and scope, and should distinguish missing information from documented absence.
Denominator and composition effects
Consider two hypothetical groups, A and B. Group A contributes 8,000 organization-years and records 80 incidents; group B contributes 2,000 organization-years and records 40. A generates twice as many observed incidents, yet the observed rates are 1% and 2%, respectively. If A records half of all qualifying events and B records all of them, the corresponding illustrative underlying rates both equal 2%. The apparent difference arose from observation rather than incidence.
A different composition can produce an aggregate rate change with unchanged group rates. If two risk classes have constant incident rates of 1% and 3%, a portfolio with 80% in the first class has a mean rate of 1.4%. With an equal split it has a mean rate of 2%. A time series without class weights would show an apparent 42.9% increase. Standardization to common exposures would separate composition from within-class rate changes.
These examples show why underwriters and researchers should distinguish crude from standardized rates. For classes indexed by j, a standardized rate is the sum of class-specific estimated rates multiplied by fixed reference weights that sum to one. It is meaningful only to the extent that definitions, observation, and exposure measurement are comparable. Where reporting propensities differ and remain unidentified, statistical standardization is insufficient; sensitivity analysis is required.
Case evidence and ascertainment
Neprash et al. (2022), in their health care ransomware cohort, identified 374 U.S. health care delivery organization attacks during 2016–2021. Their multi-source ascertainment was broader than a single federal breach dataset. Table 1 records 166 incidents with disrupted care and 198 affecting multiple facilities. Of the 374 cases, 290 appeared in the HHS reporting data and 203 were reported late according to the study's classification. The article was corrected in 2023; the current corrected HTML is used here.
The cohort illustrates selection on visibility and multiple affected settings. Counts by organizational setting can overlap when an incident affects several kinds of facilities. A proportion of observed cases associated with clinics or hospitals is accordingly not a comparison of population hazards. The dataset has inherent ascertainment limits; the incomplete duration and reporting information also demonstrate why severity analysis needs explicit missing-data categories.
A credible design for victim-selection research starts with a population frame covering both affected and unaffected organizations. It specifies the observation window, organizational linkage, baseline covariates measured before incidents, and notification probabilities or proxies. A case-control analysis can estimate relative associations under a suitable sampling design, but it needs sampling weights or population incidence information to recover absolute risks. Time-dependent covariates should avoid the leakage of post-incident information into predictors.
Insurance selection and causal questions
Adverse selection concerns differences between the underlying risk distributions of organizations purchasing different insurance arrangements; underwriting selection concerns the insurer's own choices. Moral hazard concerns behavioral changes induced by contract protection. Adversarial selection is a separate phenomenon arising from external malicious action. An association between insurance and ransomware outcomes can reflect any combination of these processes, along with differences in organization size, risk controls, reporting, contractual terms, and access to professional response.
Meurs et al. (2024) present a two-stage ransom payment analysis based on 382 attacks reported to Dutch police or handled by an incident response company. Their institutional abstract describes a hurdle model distinguishing the decision to pay from the positive payment amount. It reports associations involving backups and response services in the first stage and insurance and other characteristics in positive payment amounts. The sample is selected on known incidents; the published abstract does not establish a causal effect of insurance on incidence. No regression coefficients are transferred into a universal prediction model here.
To identify an insurance effect, observational research must clarify which pre-contract factors influence both enrollment and potential outcomes. Policy inception may coincide with organizational growth, security investment, or regulatory changes. Conditioning only on observed incidents can itself create bias if both insurance and severity influence observation. A credible comparison could use changes in coverage rules with a documented comparison group, repeated pre-treatment measurements, and sensitivity to differential reporting. Any causal claim would require assumptions and evidence beyond an association in a selected sample.
The purpose of studying victim selection is to improve risk measurement, defense, and equitable insurance assessment. The results do not furnish target lists or actionable attacker selection rankings. Aggregate public evidence should be interpreted in its defensive and actuarial context.
Comparative Measurement of Severity Distribution
The components of loss
Define economic loss from an organization's perspective as the incremental resources consumed and economic output forgone due to the event, measured against a specified counterfactual. For a given incident, components can include emergency investigation, restoration, interruption, extra expense, extortion payment if made, legal defense, notification, and liability settlements. Some harms accrue to others and are outside the organization's accounts. An insurer's payment is a transfer between parties, not an additional physical loss to society.
Ransom demand D and ransom paid P are distinct. D may be unobserved for events without confirmed extortion; P may be zero because the organization did not pay. It may be missing when the fact of payment is undisclosed. Zero, missing, and below a reporting threshold must be coded separately. The gross loss G and insurance indemnity I are separate again. For an event without ransom payment, G can still be positive and substantial, and I can be positive under interruption or response coverage.
Claims records should preserve currency, loss date, valuation date, paid and case-estimated amounts, recoveries and adjustment expense conventions. Currency conversion at a stated valuation time enables nominal comparison; inflation adjustment requires an appropriate cost index. A wider change in technology, service prices, or the type of losses can survive inflation adjustment. Accordingly no single real-price conversion creates a homogeneous cross-country severity distribution.
Payment versus total severity
The hurdle specification distinguishes the probability p of making a payment from the conditional distribution F of amounts given payment. For a nonnegative P with no positives among non-payers, E[P] = p E[P | P > 0]. The unconditional distribution has a point mass at zero. The underlying assumptions include complete identification of the payment status and comparability of the positive amounts. Missing payment status should not be treated as nonpayment.
In a purely hypothetical population of 100 incidents, 20 involve a payment averaging USD 150,000. The unconditional average payment is USD 30,000 per incident. Suppose all incidents incur USD 80,000 in other costs on average. The mean gross loss is USD 110,000 if categories are mutually exclusive and additive. If payment probability declines to 10% while the positive mean doubles to USD 300,000 and other costs remain unchanged, the unconditional mean is identical. Payment frequency and positive severity tell distinct stories even when their product is constant.
The insurance interpretation differs from the economic interpretation. A deductible, sublimit, or coverage condition can reduce paid amounts without changing the direct costs that occur. Interruption may persist even after a successful system restoration because customers, suppliers, and employees require time to restore normal operations. A ransom payment provides no mathematical guarantee of loss reduction. A sound claims model therefore treats payment as one component and considers the joint distribution of all components.
Operational and social dimensions
Neprash et al.'s cohort records disruptions to care and multi-facility effects. Such outcomes show that severity can be measured through interrupted activity and duration as well as money. The observed mean duration in their table applies to cases with duration information and should not be generalized to all incidents or all health care organizations.
Dameff et al. (2023), in their adjacent emergency department study, compare two unaffected emergency departments before, during, and after a ransomware disruption in another health system. Mean daily patient census increased from 218.4 before to 251.4 during the attack and recovery phase. The retrospective local comparison demonstrates an observed association with pressure on neighboring facilities. It is insufficient to establish a precise national causal effect, but it establishes the relevance of organizational boundaries when defining losses and their bearers.
A financial model may represent service outage as a trajectory, with lost contribution margin and incremental mitigation expense integrated over the time during which operations differ from a baseline. A care-delivery model may measure delays, transfers, and queue lengths. Some costs are measurable in monetary terms; others should retain their independent units. Adding monetary and nonmonetary consequences requires an explicitly justified valuation scheme and an avoidance of double counting.
Distributional estimation and censoring
Empirical losses are often heterogeneous, with many smaller cases and relatively few large ones. Candidate continuous severity families include lognormal and gamma distributions, and an excess-over-threshold generalized Pareto distribution may be examined where extreme-value assumptions are defensible. These are standard statistical tools, not findings that a particular ransomware population follows any one distribution. Fit must be assessed using out-of-sample predictive accuracy, tail diagnostics, sensitivity to thresholds, and uncertainty intervals.
Limits censor claims from above while deductibles can remove small losses from the observed paid-claim sample. When the observed amount is min(G, L) under a simplified limit L, amounts at L identify only that G is at least L. A likelihood for uncensored observations uses a density, while an observation censored at L contributes the survival probability beyond L. A dataset truncated below a deductible requires conditioning on exceeding that threshold. Both require correct records of the terms in force at loss time.
The survival function S(x) = Pr(G > x) provides a direct interpretation of loss-layer expectations. For an ordinary deductible d and a payment limit L, the insurer pays min(max(G-d, 0), L). Its expected value equals the integral of S(x) over x from d to d+L. This relationship follows by integrating the survival function of a nonnegative capped payment. When G is a covered aggregate amount, this is a simple one-layer model; it does not reproduce multiple coverage modules or separate sublimits.
For a hypothetical severity distribution, assume equally likely gross losses of USD 20,000, USD 120,000, and USD 600,000. The mean gross loss is USD 246,666.67. With a USD 50,000 deductible and a USD 200,000 payment limit, the respective payments are zero, USD 70,000, and USD 200,000. Mean indemnity is USD 90,000. Paid claims alone would have a mean of USD 135,000 because the zero payment is omitted. A changing policy distribution can thus change observed paid severity and paid frequency together.
Frequency and severity jointly
In a standard compound model, annual aggregate loss S is the sum of individual losses from the incident process. When N is independent of identically distributed loss amounts X, E[S] = E[N]E[X]. With finite second moments, Var(S) = E[N]Var(X) + Var(N)(E[X])². The simplified independence assumption is important; a shared economic or technological condition can influence both event intensity and the severity distribution. Selection can also create apparent dependence in the observed sample.
For a hypothetical environment with two equally probable regimes, let annual expected event counts be 10 and 30, and conditional mean severities USD 50,000 and USD 150,000. Expected aggregate loss conditional on the regimes is USD 500,000 and USD 4.5 million, giving USD 2.5 million overall. Multiplying unconditional means, 20 events and USD 100,000, gives only USD 2 million. The discrepancy is the covariance of regime-specific conditional means. The calculation assumes conditional independence within each regime and is an illustration of dependence, not a fitted cyber catastrophe forecast.
Severity must be measured at a defined maturity. A closed-claim sample may exclude unresolved large claims; a fixed valuation-date distribution may understate later payments. Recoveries, litigation, and coverage decisions can change loss estimates over time. Development analysis should distinguish ultimate gross loss, ultimate insured indemnity, and current paid loss, and should avoid treating closure as proof that all economic consequences are known.
Financial Legal and Institutional Consequences
Contract terms and insured outcomes
The financial consequence of an incident becomes insured loss through the applicable contract. The CyberAcuView model wording (2022) illustrates a business interruption coverage module in which waiting periods, restoration periods, retention, and proof of loss affect payment. It specifies alternative calculation bases involving net income with continuing fixed expenses or revenue net of relevant costs. The wording is an adaptable example, not a universally applicable policy or a statement of governing law.
A hypothetical event causes covered lost margin of USD 120,000, separately covered restoration cost of USD 80,000, and a USD 50,000 extortion payment. Assume a waiting period removes USD 20,000 of otherwise covered income loss; the extortion module has a USD 30,000 sublimit; a single USD 25,000 deductible applies after the component adjustments; and an overall USD 200,000 cap applies. The resulting amount is min(200,000, 100,000 + 80,000 + 30,000 - 25,000) = USD 185,000. Gross cost for the three categories is USD 250,000. The difference is USD 65,000. These assumptions are invented for the example and do not describe the CyberAcuView wording.
Changes in a contract can affect the measured paid-claim frequency even when the number of events remains unchanged. Attachment points, exclusions, reporting clauses, definitions of an occurrence, and legal eligibility alter the population in the claims dataset. Data on rejected or withdrawn claims and notifications below retention are particularly valuable for disentangling this mechanism. Coverage analysis in any actual dispute would require the complete policy, endorsements, relevant facts, and applicable law.
Comparative institutional scope
The United States evidence in this dossier comprises federal complaint data, U.S. health organization research, an insurance supervisory market report, and an illustrative policy form offered through an industry repository. Each describes a different observation or contractual institution. The United Kingdom supplies a weighted national survey of eligible businesses and charities. The European Union evidence concerns aggregated threat observations. Dutch research provides a selected incident-level sample rather than a census. These are comparative settings, not equivalent longitudinal datasets.
| Evidence frame | Unit or denominator | Valid primary interpretation | Boundary |
|---|---|---|---|
| IC3 complaints | Recorded complaints | Administrative reporting trend | Unknown population reporting fraction |
| UK survey | Weighted eligible organizations | Identified-event prevalence | Awareness, recall, scope exclusions |
| ENISA release | Selected observed events | Threat landscape and dependencies | No organization-year denominator |
| NAIC | Direct written insurance premium | Market amount | Not ransomware incident rate |
| Coalition | Insurer observations | Portfolio claim and demand metrics | Underwriting and coverage selection |
| Health cohorts | Identified cases or selected facilities | Observed disruption and reporting | Limited sector and time scope |
| Dutch incident analysis | Police/response-company cases | Conditional payment associations | No population sampling frame |
Reporting obligations and legal liability differ across jurisdictions and organizational categories. A reporting threshold may create discontinuities in the observed case distribution. An institution with stronger reporting rules can produce more observed events even when effective preventive measures reduce harm. A legal duty to notify, a claim notification duty under a contract, and voluntary reporting to police are different events. This study deliberately limits comparative legal analysis to these structural distinctions; it does not certify the currently applicable statutory requirements of every jurisdiction or the lawfulness of any payment.
Underwriting and claims governance
Evidence of protective capacity can improve insurance risk assessment when collected consistently and tested against real outcomes. A framework such as NIST CSF supports the organization of control evidence, not a guarantee of effectiveness. Claims experience can test whether documented controls correlate with interruption duration or loss components, but such testing must adjust for portfolio mix and avoid post-event selection.
An organization that reports early may receive support and reduce harm. The same organization can generate more notifications, including events that never trigger payment. An insurer that incentivizes timely reporting might thus observe more incident records while suffering lower severity. Rewarding an apparent low incident count without examining reporting quality could undermine risk governance. Distinguishing notifications from payable claims provides an operational measurement of this problem.
Claim records should retain the reasons for coverage determinations, the policy version, dates of reporting and development, and gross and covered losses wherever available. Financial records should indicate amounts paid to vendors, policyholders, or third parties and the allocation of expenses. This supports both reserving and risk prevention. It also reduces the risk of interpreting an increase in professional response spending as an increase in underlying attack severity without examining whether the spending avoided greater loss.
Capital and accumulation
The individual-claim severity distribution is only one determinant of capital requirements. Shared digital infrastructure creates the possibility of multiple affected insureds from one common disturbance. The observed health care multi-facility cases and the institutional emphasis on dependencies illustrate why concentration requires explicit measurement. A carrier with many small policies could suffer a large aggregate event if those policies share a dependency.
For an equally exposed hypothetical portfolio of 1,000 entities with annual incident probability 1%, equal fixed insured loss USD 100,000 and independent Bernoulli indicators, expected loss is USD 1 million and variance of the incident count is 9.9. If all pairs instead have constant positive correlation 0.01, the count variance is 9.9 times (1 + 999 times 0.01), or 108.801. The mean is unchanged. This is a moment-based illustration; it does not specify a complete multivariate distribution or certify the feasibility of every other correlation pattern. It demonstrates how even modest dependence can substantially change volatility.
Reinsurance contracts may define event aggregation and attachment differently from primary policies. A capital stress analysis therefore needs a common-event representation, limits and retentions, contract terms, and uncertainty about cross-insured propagation. Historical loss data alone may have insufficient extreme events to estimate a credible return period. Scenario analysis can complement statistical estimation if assumptions and scope are explicit. Scenario outputs must remain distinct from empirical probabilities.
Risk governance and public consequences
Social losses can exceed the financial losses of the initially affected entity. The adjacent hospital case documents operational pressure at facilities outside the organization directly attacked. Contract compensation financed through insurance can improve recovery resources while leaving other harms outside contractual coverage. Public institutions have an interest in reducing interruption and restoring essential services. Their objectives include resilience and harm reduction, which may not be captured by insured severity.
Insurance data sharing can help public authorities and researchers if sensitive incident information is appropriately protected. Reporting should minimize unnecessary identifying technical details while preserving enough metadata to estimate rates and losses. The beneficial use of claims information depends on data governance, retention policies, the purpose of access, and appropriate legal safeguards, as well as the statistical validity of the resulting analyses.
Integrated Scientific Synthesis and Research Agenda
What the evidence establishes
Public sources establish the reality of changing reported ransomware activity, substantial financial consequences within particular insurance observations, and operational harms that extend across organizational boundaries. They also establish the incomparability of raw indicators without specification of units, periods, and observation mechanisms. The central research result is the identification of measurement conditions under which these sources become interpretable together.
IC3 complaint growth is a trend in one recorded stream. The UK survey supplies a weighted prevalence measure among eligible organizations and conditional proportions among respondents who identified incidents. NAIC market premiums describe market volume. Coalition observations concern claims and demands in its business. The clinical studies concern identified health care incidents and adjacent operational effects. Each answers a scientific question; none substitutes for the other. The empirical differences are consistent with a changing distribution of exposure, reporting, harm, and indemnification.
The examples in this monograph demonstrate mathematically that a constant observed rate can be compatible with different underlying rates; a change in portfolio mix can alter aggregate incidence; payment-probability and positive-payment severity changes can offset one another; and policy limits can transform both the observed probability of payment and the conditional paid severity. Dependence can increase aggregate volatility without increasing expected losses. Together these established statistical relationships provide an integrated actuarial interpretation rather than a new theory or an empirical fit.
Data structure and reproducibility
A minimum defensible research dataset would distinguish:
- An entity table with persistent pseudonymous identifier, organizational structure, sector and exposure intervals.
- A policy table recording coverage form, period, relevant limits and deductibles, endorsements and changes.
- An event table separating incident origin, detection and occurrence dates, event category, confidence and affected entities.
- A notification table recording channels, dates, distinct sources, and correction history.
- A loss-component table recording valuation date, currency, financial amount, missingness and accounting perspective.
- A claim table linking policies to incidents, status, covered amounts, expenses, indemnity, recoveries and closure.
- A dependency table recording relevant shared services using controlled access and aggregate disclosure.
Such a relational structure allows separate calculations without conflating levels. A population denominator may be obtained from a sampling frame or an insured exposure census, depending on the estimand. Deduplication should be documented with uncertainty when entities, dates, or identifiers conflict. Researchers should publish definitions, eligibility rules, extraction dates, code, summary statistics, and reproducible transformations to the degree confidentiality permits. A synthetic dataset is useful for testing software and formulas, but cannot replace representative observations.
A basic validation plan would first reproduce cohort counts and denominators; next reconcile missing dates, duplicates, and contracts; then compare frequency specifications, test positive severity and censoring, and validate aggregate dependence assumptions. Holdout windows should follow calendar order. Resampling should respect organizational or common-event clusters where observations are related. Parameter uncertainty, missingness uncertainty, and structural scenario uncertainty should be separated rather than combined into an apparently precise probability.
Sensitivity and partial identification
If an observed number of cases n is generated by a complete fixed exposure e apart from an unknown capture probability q bounded between q_min and q_max, then the corresponding intensity consistent with its expected count is bounded by n/(e q_max) and n/(e q_min). Sampling noise is additional. The earlier illustration with 80 cases yields 1% to 2% for q between 0.4 and 0.8. These bounds are conditional on the assumed capture range and not empirical confidence limits.
If recording varies materially by severity, the observed distribution is weighted by probability of observation. Even when the count can be adjusted, severity composition may remain biased. An observed severity distribution could be approximated with inverse-probability weighting when valid recording probabilities are known and positive. In the absence of those probabilities, a transparent range of alternative observation mechanisms is more scientifically defensible than a single purportedly corrected estimate.
Multiple sources can be linked by privacy-preserving entity and time identifiers. Capture-recapture methods may estimate missed cases under assumptions about overlap and inclusion probabilities. Reporting sources are often dependent because one notification triggers another, and heterogeneous capture probabilities can further bias the estimator. A two-source capture-recapture estimate under conditional independence is therefore a sensitivity exercise unless the assumptions are supported.
Research priorities
The first priority is longitudinal entity-year cohorts joining incident notifications, insurance exposure, and claims with consistent event definitions. The second is measurement of reporting behavior across channels, including losses without claims. The third is joint estimation of payment, restoration, business interruption, and liability with missingness and maturation explicitly treated. The fourth is the relationship between independently measured protective outcomes and observed frequency and severity within adequately controlled designs.
Further work should examine changes in policy terms and underwriting portfolios as sources of structural variation; analyze shared infrastructures and multi-organization loss development; and assess the effects of incident support using well-defined comparison settings. International studies need harmonized units with jurisdiction-specific reporting definitions and contractual applicability. They should avoid aggregating count series that describe distinct events and population frames.
The current evidence leaves no defensible basis for a single global ransomware frequency, a universal severity distribution, sector causal rankings, or an estimate of the causal effect of insurance on the probability of being attacked. A population study with appropriate microdata could produce more precise results. The scientific objective is a defensible estimate tied to a stated population and observation process.
Limits of this investigation
This inquiry is a source-based synthesis. No confidential policyholder records, raw payment transactions, or independently reconstructed attack dataset were obtained. The Dutch conference paper was assessed through its institutional abstract and metadata, the full ENISA report was inaccessible in direct retrieval, and the Coalition source is its official announcement. The retrieved NAIC and specimen-policy PDF text supported the narrow claims made; independent screenshots for those two documents failed. The FBI ransomware section was visually inspected. No legal opinion concerning actual payment eligibility or jurisdiction-specific current mandatory reporting deadlines is offered.
Published data describe unequal periods and populations. The research covers selected original sources rather than every available industry report or legal instrument. The synthetic examples are algebraically reproducible and deliberately do not derive actionable victim profiles. A further comprehensive scholarly review, proof of each external link at the final locked-source stage, and site-wide local technical validation are required before editorial release clearance.
Conclusion
Ransomware frequency is meaningful when the event, population, exposure period, and observation process are specified. Victim selection affects both the event population and the samples used to estimate risk. Severity comprises multiple economic and contractual consequences, and its distribution changes according to the loss measure, eligibility, policy structure, reporting, and development. A rigorous insurance analysis consequently preserves the distinct identities of an incident, demand, payment, economic loss, claim, and indemnity.
The primary evidence establishes that complaint trends, survey prevalence, insurer observations, and health care case studies convey complementary information. Their combined use supports a contextual understanding of ransomware risk but does not identify a universal global distribution. The actuarial illustrations establish the importance of selection, truncation, censoring, dependence, and temporal variation. The original contribution of the study is their integrated interpretation for insurance research and an explicit specification of the data needed to estimate the quantities responsibly.
A scientifically useful insurance risk assessment can be organized around population-specific exposure, independently documented event definitions, observed contract terms, multiple loss components, transparent uncertainty, and the dependencies that link affected organizations. That architecture enables defensible scientific comparison and supports institutional decisions about risk financing and prevention.
References
FBI (2026). 2025 IC3 Annual Report. Federal Bureau of Investigation, Internet Crime Complaint Center, ransomware section p. 14 and complaint-type table. Source
Department for Science, Innovation and Technology and Home Office (2026). Cyber security breaches survey 2025/2026. UK official statistics. Source
Department for Science, Innovation and Technology and Home Office (2026). Cyber security breaches survey 2025/2026: technical report. UK official statistics. Source
European Union Agency for Cybersecurity (2026). Exploring the evolution of the cyber threat landscape: How dependencies weaken our digital resilience. Institutional release accompanying ENISA Threat Landscape 2026. Source
National Association of Insurance Commissioners (2026). Report on the Cybersecurity Insurance Market. Section 1, p. 2. Source
Coalition (2026, March 6). Coalition’s 2026 Cyber Claims Report Finds Initial Ransom Demands Surged 47%. UK announcement. Source
Romanosky, S. (2016). Examining the costs and causes of cyber incidents. Journal of Cybersecurity, 2(2), 121–135. DOI: 10.1093/cybsec/tyw001. Source
Neprash, H. T., McGlave, C. C., Cross, D. A., Virnig, B. A., Puskarich, M. A., Huling, J. D., Rozenshtein, A. Z., & Nikpay, S. S. (2022). Trends in Ransomware Attacks on US Hospitals, Clinics, and Other Health Care Delivery Organizations, 2016–2021. JAMA Health Forum, 3(12), e224873. DOI: 10.1001/jamahealthforum.2022.4873. Corrected February 17, 2023. Source
Dameff, C., Tully, J., Chan, T. C., Castillo, E. M., Savage, S., Maysent, P., Hemmen, T. M., Clay, B. J., & Longhurst, C. A. (2023). Ransomware Attack Associated With Disruptions at Adjacent Emergency Departments in the US. JAMA Network Open, 6(5), e2312270. DOI: 10.1001/jamanetworkopen.2023.12270. Source
Meurs, T., Cartwright, E., Cartwright, A., Junger, M., Hoheisel, R., Tews, E., & Abhishta, A. (2024). Ransomware Economics: A Two-Step Approach To Model Ransom Paid. 2023 APWG Symposium on Electronic Crime Research (eCrime). IEEE. DOI: 10.1109/eCrime61234.2023.10485506. Source
Pascoe, C., Quinn, S., & Scarfone, K. (2024). The NIST Cybersecurity Framework (CSF) 2.0. NIST CSWP 29. DOI: 10.6028/NIST.CSWP.29. Source
Cybersecurity and Infrastructure Security Agency, FBI, NSA, and MS-ISAC (2023, May 23). CISA, FBI, NSA, MS-ISAC Publish Updated #StopRansomware Guide. Official agency bulletin. Source
CyberAcuView, LLC (2022). CyberAcuView Cyber Insurance Wording. Illustrative base wording, Business Interruption Coverage Module, IV, pp. 21–22. Source
Bogdanova, A. (2026). Cyber Insurance. Insurance by Bogdanova. Published by Aisentica Development. Science IBB-SCI-027. Source