Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Cyber Loss Event Taxonomy and Data Interoperability

Abstract. Cyber loss measurement requires a coherent relationship between the technical occurrence, its effects on an exposed entity, the contractual determination of insurance liability, and the resulting financial transactions. Contemporary cyber incident taxonomies, regulatory notification standards, threat-intelligence formats, operational risk classifications, and insurance claims records represent distinct observational purposes. Their codes cannot be treated as interchangeable simply because their labels resemble one another. This monograph examines international primary standards, insurance policy specimens, official market and incident statistics, and empirical research to determine the conditions under which cyber loss data can be compared and exchanged without material semantic distortion. It analyzes event granularity, causal uncertainty, coverage triggers, financial measurement bases, exposure denominators, dependence, versioned crosswalks, and governance. Synthetic calculations demonstrate the consequences of duplicated records, incomplete observation, incompatible waiting periods, and common-shock accumulation. The central analytical finding is that interoperability must preserve the distinctions that determine both statistical inference and contractual liability. A harmonized analytical representation should retain source identity, time, entity, provenance, classification uncertainty, contract-specific attributes, and loss valuation rather than force different kinds of records into a single incident count. Existing standards supply complementary components; their boundaries delimit what can be inferred from each. The research is documentary and analytical, not a proprietary claims-data estimation exercise.

Keywords: Cyber insurance; cyber loss event; incident taxonomy; data interoperability; actuarial science; claims classification; operational risk; incident reporting; FIRE; VERIS; systemic cyber risk; exposure denominators; loss aggregation; semantic mapping; data quality.

Introduction

The cyber insurance market transforms disruptions of digital systems into contingent financial obligations under contracts. To measure such obligations, an insurer must determine what happened, to whom, when, through which operational dependencies, under which wording, and with what ultimate financial consequences. Each question has a different observation unit. One software update may interrupt a thousand enterprises, one enterprise may encounter several technical incidents during a single interruption, and one insurance claim may include several categories of expense. A single number of cyber events, without a specified unit, cannot identify the frequency of covered loss.

This study asks how systems built for cybersecurity, financial supervision, operational risk management, and insurance can exchange useful data while retaining the meaning of their original records. Its scientific scope comprises event classification, common coding, and cross-source comparability. Its immediate disciplinary home is Cyber Insurance, with intersections with actuarial science, insurance contract law, operational resilience, statistics, and data engineering. The problem is international because a shared provider may create losses in numerous markets while legal notification and contract interpretation remain jurisdiction- and wording-specific.

The study uses comparative documentary analysis. Primary materials include normative standards, official statistical publications, and publicly available policy specimens; scholarly articles supply evidence of historical data constraints. The inclusion criterion is a demonstrable relationship to at least one of the three scientific dimensions. Sources are examined at the level accessible on November 18, 2026; an accessible summary is used only for propositions that it supports. Selected policy clauses are illustrations of contractual heterogeneity, not a market-representative sample or legal advice. Numerical examples are independently constructed and explicitly synthetic. They are algebraic sensitivity analyses, not empirical estimates. No primary claims microdata, confidential operational records, or licensed codebook is assumed available. No systematic-review exhaustiveness or newly established authorial theory is claimed.

The progression from definitions to mappings and then statistical consequences is deliberate. Technical equivalence of a data format alone cannot establish whether two observations describe the same economic exposure or legal occurrence. Even perfect transmission may preserve the wrong unit of analysis. Scientific interoperability begins by identifying that unit before normalizing its representation.

Conceptual and Historical Foundations

The scientific object of classification

A classification assigns observations to categories in accordance with stated characteristics. A taxonomy structures categories and their relationships; a codebook specifies the conditions for applying the categories. A vocabulary provides terms, a schema specifies fields and constraints, and an exchange format determines how representations can move between systems. These tasks overlap but have different success conditions. A syntactically valid record can be scientifically misleading if the code assigned to it has changed meaning or if its denominator is undefined.

The Financial Stability Board (FSB, 2023a) distinguishes observable cyber events, adverse cyber incidents and malicious cyber attacks in its updated Cyber Lexicon. The distinction matters because an unplanned system failure can be an incident without a hostile attacker. A detected but blocked attempt may be useful security telemetry without producing an adverse outcome or insured loss. An attack can also succeed without material financial damages. Conversely, an operational interruption may be economically severe even if confidentiality was never compromised.

For insurance analysis, an incident is ordinarily an observation about an adverse effect on a system or organization under a specified definition. A loss event is an occurrence, or grouping of occurrences, to which financial consequences are attributed under a stated accounting or actuarial convention. An insured event is an occurrence satisfying the applicable contract's operative conditions, as determined on evidence. A claim is a record of a request or assertion under insurance, with a particular legal or administrative identity. Coverage is the contractually evaluated entitlement; a payment is a transaction. The terms can coincide in a simple case but their units are distinct. These working definitions are explanatory conventions for this research, not universal legal definitions or a new theoretical system.

From security logs to financial datasets

Security telemetry originated around observations of systems, users, and network behavior. Insurance claims systems were organized around contracts, insured entities, case files, and financial obligations. Supervisory frameworks developed their own reporting priorities, notably continuity of critical services and financial stability. The growth of digital dependencies increased the significance of the relationship among these traditions. Each dataset reflects decisions about what its collecting organization needs to observe. Those decisions affect the resulting distribution.

The ENISA (2017) report on risk assessment language in cyber insurance addressed common terminology and data-sharing constraints. The significance of this work is not that a universal insurance codebook was established; it is that the absence of common semantics was recognized as a market issue. FSB (2023b) subsequently addressed convergence in cyber incident reporting in the financial sector. Its objective concerned fragmented reporting, not a harmonized adjudication of insurance claims. The development of incident-response guidance also followed the institutional need to integrate technical evidence with risk management. Nelson et al. (2025), in NIST incident response recommendations, place incident response within a broader cybersecurity risk-management process.

The historical pattern is an expansion of the number of uses for incident data. It creates demand for interoperability, but also makes uniformity harder. An event reporting system intended to generate a prompt supervisory alert has incentives different from a claims system that may take months to determine covered business interruption. Collapsing them into one table hides these time-dependent differences.

Taxonomy as a measurement instrument

A measured frequency is a function of underlying occurrences, exposure, detection, definition, and recording practices. If one insurer codes ransomware as a distinct loss type and another codes it as business interruption whenever production stops, their individual categories reflect both incident composition and administrative policy. Their totals may be equal while the component distributions differ. A reclassification can therefore alter a reported trend without changing any underlying risk.

The same is true of severity. A source recording gross economic harm may include lost production or response expenses never submitted to an insurer. A claims system may include only payments or estimates of contractually covered amounts. A supervisory report may record financial impact in a band before damages are known. An analyst who averages these numbers as though they represent the same random variable has introduced a measurement error, not eliminated a data gap.

Classification is also affected by uncertainty. An incident initially recorded as unauthorized access may later be linked to compromised credentials or a software flaw. Evidence of exfiltration may become available after containment. The insurer may determine an exclusion only after investigating the cause or contractual language. A record should support changes to its asserted properties without rewriting the earlier observation history. As-of dates, evidence states, source revisions, and the difference between unknown and absent are necessary to that purpose.

Scope and granularity

Four levels are particularly important: shared originating episode, affected organization, claim or policy, and financial component. An originating episode is a candidate common cause, not necessarily a legally aggregated occurrence. An organization-level incident describes effects on one reporting or affected entity. Claim-level records relate to a policy, insured interest, or administrative case. Financial records describe valued expenses, liabilities, reserves, recoveries, or payments. Further subdivisions may be necessary: location, business service, asset, legal claimant, coverage section, or reinsurance layer.

This distinction leads to a practical principle of measurement: counts must be additive only within a stated partition. Organization incidents can be added over disjoint entity sets; claim payments can be added over disjoint financial entries on the same accounting basis; shared episodes can be counted by the adopted linkage rules. Adding across levels does not create a meaningful exposure statistic. The principle follows from elementary counting, not from a proprietary classification system.

Event Classification

Technical, operational, and economic dimensions

A useful event description records several complementary dimensions: suspected or established cause; action or failure mechanism; affected assets and services; confidentiality, integrity, and availability effects; uncertainty; and economic consequences. For insurance, it also records contractual status and the financial components associated with those consequences.

The MITRE Center for Threat-Informed Defense (n.d.) documents how the four VERIS dimensions of actor, action, asset, and attribute relate to its VERIS Mapping Methodology. The account identifies limitations in mapping to adversarial techniques, including actions outside adversarial behavior. This illustrates why an attacker-technique vocabulary alone cannot represent all insured cyber losses. A configuration failure without an attacker is still potentially relevant to interruption coverage under appropriate wording. A credential compromise may involve several techniques and numerous affected resources. Its classification should not be forced into a single code merely to make a chart convenient.

Threat actor labels and attribution require particular caution. A reported state-linked actor, suspected criminal group, or unknown origin is an evidentiary claim. These labels should carry assertion source and confidence, not become immutable attributes of the event. Even a technically plausible attribution does not establish the applicability of an insurance war exclusion or another contractual provision. Legal interpretation depends on wording, governing law, and the evidence admissible in the relevant context.

Multiple causes and consequences

Events rarely conform to a single linear sequence. An initial credential compromise may lead to unauthorized access, privilege escalation, copying of data, encryption, extortion, and an availability interruption. One chain can contain multiple observable actions and multiple consequences. Whether these comprise one contractual occurrence, several loss events, or several claims is a separate question.

The FSB (2025a) Format for Incident Reporting Exchange includes distinct cause, event, and impact elements and accommodates changing information during a reporting lifecycle. Separating these attributes is valuable for cross-domain work because causes, mechanisms, and outcomes can otherwise be confused. An impact code for interruption is not a synonym for the attack technique that produced the interruption. Several incident types may be relevant to an episode, and an early explanation can change as investigation progresses.

Consider a synthetic population of 100 affected-organization incidents. Forty-five have ransomware characteristics, 30 involve exfiltration, and 20 exhibit both. The union contains 55 incidents: 45 + 30 - 20. A tabulation of multi-valued tags legitimately reports 75 tag assignments but cannot label them 75 distinct incidents. If a mutually exclusive principal-peril distribution is required, the coding policy must state the priority rule. Whether that rule prefers initial compromise, sometimes dominant damage mechanism, or largest estimated financial component changes the resulting distribution. A separate multi-label record permits the alternative distributions to be reconstructed.

Common-cause incidents and affected entities

A common origin and a shared contractual occurrence are separate propositions. A shared service outage can have one origin, hundreds of individual effects, and numerous insurance policies responding under different conditions. The components of a major cyber event published by Lloyd's and the Association of British Insurers Cyber Working Group (n.d.) recognizes the significance of event characteristics and footprint for insurance and reinsurance. The origin, propagation, affected counterparties, and financial footprint should therefore be preserved as related data, rather than treated as one indivisible row.

The July 2024 CrowdStrike incident provides a documented example of scale and granularity. CrowdStrike (2024) identified a routine configuration content update associated with Windows sensor failures in its Channel File 291 RCA announcement. Microsoft's contemporaneous outage impact estimate (Weston, 2024) was 8.5 million affected Windows devices, less than one percent of all Windows machines. That estimate measures devices in a particular technical population; it supplies neither the number of affected insured organizations nor the number of insurance claims. The vendor accounts employ their own dates and operational vantage points. Their observation boundaries are preserved here: no reconstruction of insured losses, policy attachment, or universal restoration duration follows from those reports.

The episode nevertheless exposes an important structural issue. A software dependency can connect otherwise unrelated insureds. Two portfolios with similar counts of enterprises may have very different exposure to this type of disruption if their critical services depend on different suppliers or implementations. Reinsurance aggregation requires its own occurrence definition and the application of the relevant treaty terms. The technical linkage is evidence for the analysis, not its legal result.

Insurance trigger as a separate dimension

A policy trigger is determined by policy wording, facts, and the applicable interpretation. A cybersecurity classification offers facts relevant to a trigger, but it does not by itself establish one. The same technical impairment can fall within different coverages or outside coverage entirely depending on the contract.

The CyberAcuView LLC (2022) base policy specimen differentiates security breach and system failure in its business interruption provisions and addresses material interruption, discovery, waiting periods, and dependent interruption. The Victor Insurance (2021) Cyber Secure policy wording organizes lost profit and increased costs under specified conditions; its waiting-period provision specifically bears on lost profit. The Corvus Insurance (n.d.) Smart Cyber policy specimen distinguishes business interruption, dependent interruption, and system failure through separate provisions with their own enumerated circumstances. These are actual examples of policy drafting, not endorsements of the products or a representative global policy survey.

Analytical question Relevant technical field Additional insurance evidence
Did a security breach occur? Access and security-impact observations Contractual definition and evidence of the qualifying conduct
Did a system failure occur? Error or disruption mechanism Policy-specific cause, insured system, exclusion and trigger
Did interruption attach? Service downtime and restoration timeline Waiting-period terms, materiality conditions, covered period and valuation
Was a dependent provider involved? Service relationship and provider footprint Covered dependency definition, supplier eligibility, limits
Has liability arisen? Harm to third parties and relevant acts Allegation, legal duty, claim definition, applicable period and exclusions
What is payable? Impact evidence and recovery Adjustment, deductible, limits, other insurance and recoveries

This comparative reading supplies a substantive constraint on interoperability. Mapping an incident code directly to a paid-loss code discards the contractual determination. A common representation should preserve the original policy form and version, applicable coverage, assessment outcome, and reason for an unresolved or denied claim. A field indicating no indemnity is different from a field indicating that indemnity has not yet been adjudicated.

Time, cause, and claim development

An incident may have dates for onset, occurrence, discovery, interruption, notification, claim registration, coverage determination, restoration, and payment. Each date answers a distinct question. Occurrence-year, reporting-year, underwriting-year, accident-year, and calendar-year analyses cannot be substituted for one another without reconciliation.

For example, a compromise originating in December, discovered in January, and claimed in February can enter three different time series. The policy conditions determine the relevant insurance period; the statistical specification determines the relevant observation year. Persisting only the claim registration date destroys the ability to analyze discovery delay or distinguish apparent calendar trends from changes in reporting practices.

An incident's temporal boundaries should also carry uncertainty. A detection timestamp is often exact, but the earliest compromise may be known only within an interval. In an event cluster, one insured may recover before another. A common event identifier must not imply common recovery duration. That difference is central to business interruption severity.

Common Coding

Standardization by purpose

International standardization provides useful semantic resources but does not establish that all codes have the same extension. A standard's object, population, and intended use must be specified before its terms are translated.

Source family Principal represented object Value for cyber insurance Boundary
VERIS-oriented incident classification Actors, actions, assets and security attributes Mechanism and impact coding No automatic policy adjudication
Threat-intelligence exchange Threat objects and relationships External evidence and correlation Not a claims ledger
Security telemetry schema Technical observation Ingestion and correlation High-volume observations need incident construction
FIRE and supervisory reporting Reporting entity and incident update Cross-institutional reporting fields Not a direct coverage or payment definition
Operational risk taxonomy Operational event, cause and impact Financial-institution risk analysis Population differs from insured claims
Insurance administration Policy, claim, adjustment, payment Contract and indemnity measurement Often incomplete technical history

OASIS Open's (2021) STIX Version 2.1 standard specifies typed threat-intelligence objects and relationships. Its representations can be used to exchange intelligence relevant to an event, but the grouping of intelligence into shared context is not a legal determination of an insurance occurrence. The OCSF core schema repository describes structured cybersecurity telemetry through categories, classes, types and attributes (OCSF Project, n.d.). Its observational granularity may be orders of magnitude finer than claims granularity. This study assesses the documented purposes of these formats, not implementation conformance.

The FSB (2025b) publishes a FIRE taxonomy package with machine-oriented resources including a data dictionary and validation materials. Its published version information is an example of the metadata needed for reproducibility. FIRE itself provides an exchange format rather than independently imposing national reporting duties. A regulatory reporting template must be applied with the authority's relevant rules.

Operational risk frameworks also contribute a financial orientation. The ORX (n.d.) operational risk reference taxonomy separates the classification of operational events from their causes and effects. The public description of ORX (2026) Cyber Operational Risk Reporting Standards identifies event types, attack characteristics, actors, failed controls, and exposure indicators. The proprietary detailed codebook was not inspected, so this research makes no field-by-field equivalence claims for it.

Crosswalk semantics

A crosswalk is a documented relation between source categories and destination categories. Its correctness depends on the defined scope, inclusion criteria, granularity, and effective version of the categories. A label match alone is insufficient. Two distinct terms may be extensionally equivalent within a bounded application, while identical terms may cover different populations.

Five ordinary mapping outcomes are useful: exact equivalence; source narrower than target; source broader than target; partial overlap; and no known mapping. A sixth operational condition is uncertainty pending evidence. These distinctions are standard semantic considerations applied here to insurance data. They require the crosswalk to declare direction. A source subclass can often be aggregated into a broader target category, whereas the broader record cannot be uniquely disaggregated back to its subclasses.

Assume a source has a category for malicious encryption accompanied by extortion. A broader target category covers all extortion involving digital threats. The source category can enter that target under the stated definitions, but a target event cannot automatically be returned as encryption. Likewise a source category combining accidental digital disruption and hardware breakdown may overlap with, rather than exactly equal, a particular policy's system failure definition. Mapping through an intermediate common vocabulary is convenient but does not cure a lossy relation. Two individually plausible broad mappings may compose into a false equivalence.

Source description Target analytical dimension Possible relation Necessary condition
Confirmed malicious encryption with demand Extortion-related mechanism Narrower Demand established and target includes encryption extortion
Unknown mechanism, service unavailable Availability impact Direct at impact level Unavailability documented
Regulatory major ICT incident Covered interruption Partial or unknown Contract and loss adjustment evidence
Paid forensic invoice Incident-response expense Conditional correspondence Basis, payer and policy classification known
Shared provider incident Common-origin cluster Candidate association Evidence of shared causal connection

These examples are constructed crosswalk illustrations, not official mappings between named taxonomies. Their purpose is to establish the information that must survive a translation.

Relational structure and identifiers

A flat export can be useful for descriptive statistics, provided the analyst declares its grain. A relational representation is better suited to an event with multiple affected entities, coverage sections, and revisions. The following is a practical description of necessary record types for exchange, not a proposed new industry standard.

Record type Identity and time Essential relations
Source observation Source namespace, source record ID, version, observed time Linked to incident assertion
Originating episode Cluster ID, interval, confidence, evidence Linked to affected incidents
Affected-entity incident Incident ID, entity token, affected service, discovery and restoration Linked to source, episode and exposure
Exposure Policy or exposure token, period, insured entity, contract version Linked to incident and claim
Claim Claim token, registration date, status, valuation date Linked to policy, incident and coverage
Loss component Component ID, kind, basis, amount, currency, valuation date Linked to claim or economic incident
Financial movement Transaction ID, type, amount, period Linked to component where possible
Mapping assertion Source and target codes, mapping version, confidence, rationale Linked to original classification

The records should retain stable internal identifiers and the namespace from which any externally provided ID originates. Reusing a claim number from another carrier without its namespace can erroneously merge unrelated cases. Privacy-preserving tokens can support linkage, provided their issuance and rotation are governed and the linkage has a lawful purpose. Identifiers should be distinct from classifier labels, because a label can change while the represented incident persists.

Record merging also needs history. If an incident is later split into two, both the earlier assumption and the new relation must remain traceable. A source assertion that two events are the same is itself a versioned observation, not permission to erase one of the original records. The reverse is true for mistaken duplication. A de-duplication process should retain the mapping from original IDs to the resulting analytical identity.

Data states and financial meaning

Unknown values, unavailable values, inapplicable fields, not-yet-assessed fields, and explicit negative findings should have distinct machine-readable representations. A null monetary field must not become zero automatically. Zero can be a confirmed financial outcome; unknown is a limitation of evidence. Likewise an excluded loss component can be recorded as economically positive and covered amount zero without eliminating the component from the economic dataset.

Financial fields should state monetary unit, currency, valuation date, accounting basis, and whether recoveries or reinsurance are included. A snapshot of incurred claims may combine paid amounts and outstanding estimates. A transactional export records movements over time. Adding successive snapshots creates multiple counts of the same liability. An incurred estimate at one date is not an additional liability to be added to another estimate at a later date.

If an analyst needs a common-currency measure, the currency conversion source, exchange date, and convention must be stated. Converting a payment at its transaction date and converting the outstanding reserve at the valuation date can support a reporting objective, but it is not equivalent to applying a single annual average to all amounts. Inflation adjustment is another transformation, with its own basis. Original amounts should remain accessible for reproducibility.

Governance of schema evolution

Versioning should apply to the source vocabulary, local crosswalk, transformation implementation, reference data, and analytical extract. A revised crosswalk should have an effective date and a record of changes. Where historical outputs remain published, they should be clearly distinguished from restated results. Changes in classification should not silently create a revised historical series.

Conformance tests should validate syntax and semantics independently. A schema can reject an invalid currency code or an end date before a start date. It cannot establish that a business interruption amount has been calculated under a policy's correct definition. Evidence review and contract controls are therefore required above structural validation.

The exchange organization should minimize unnecessary personal and operationally sensitive information. Shared datasets can often use pseudonymous entity tokens, bounded timestamps, standardized attributes, and aggregated exposures instead of raw breach details. Access rights and retention rules should depend on the purpose of the data and applicable restrictions. Pseudonymization by itself is not an assertion of anonymity: combinations of sector, provider, and outage date may identify a firm. Linkage risk and scientific usefulness must be assessed together.

Cross-Source Comparability

Regulatory reporting populations

Regulatory classification is purpose-specific. The European Commission's (2024) Delegated Regulation (EU) 2024/1772 specifies criteria and materiality thresholds for classifying ICT-related incidents under DORA. Articles 8 and 9 combine critical-service impact with specified threshold conditions; Article 8 also addresses certain recurring incidents within six months that share an apparent root cause and collectively meet relevant criteria. The economic-impact threshold in Article 9 is part of that rule structure, not a universal cyber-loss threshold. An incident below or above a monetary threshold cannot be labeled covered or uncovered on that basis.

In the United States, the SEC's (2023) cybersecurity disclosure rule announcement explains that domestic registrants' Item 1.05 Form 8-K disclosure is generally due four business days after materiality is determined, subject to a specified national-security or public-safety delay procedure. The observation unit is a registrant's material cybersecurity incident. The disclosure date and materiality determination date are not interchangeable with onset or discovery.

Australia's APRA (2018) Prudential Standard CPS 234 establishes information-security requirements for specified regulated entities, including insurers, and includes incident notification obligations. It addresses regulated institutional resilience. An insurer reporting its own disruption under a prudential standard and an insurer reporting an insured customer's loss have different roles and populations. They may even refer to connected aspects of the same systemic episode, but counting both records as separate insured claims would be an error.

The comparison establishes why a common report format cannot remove differences in legal scope. The population of obligated entities, triggering conditions, assessment clock, permitted reporting exceptions, and version of the applicable law must accompany a regulatory export. Combining regulatory observations with insurer records is useful for linkage and contextual analysis, not for assuming that all records have an equal probability of inclusion.

Official statistics and denominator discipline

The UK Department for Science, Innovation and Technology and Home Office (2026a) Cyber Security Breaches Survey 2025/2026 reports that 43 percent of businesses and 28 percent of charities in its relevant survey populations identified a breach or attack during the preceding twelve months. The survey's sampling frame, timing and definitions structure these results. According to the accompanying survey technical report (2026b), the study uses survey sampling and weighting; perceived costs and measurement changes require careful interpretation. Respondent detection and recall influence the observations. The reported prevalence is not insurer claim frequency, and a population-weighted proportion is not the same statistic as an event count per organization.

The NAIC (2025) cybersecurity insurance market report documents changes to the 2024 insurance reporting categories, differentiating primary, excess, and endorsement policies in the newer collection. Such a revision creates potential structural breaks when comparing older categories. The NAIC (2026) 2026 cybersecurity insurance market report lists domestic direct written premium amounts of 7,082,935,371 dollars for 2024 and 7,458,668,491 dollars for 2025 in the cited series. Their difference is 375,733,120 dollars, approximately 5.30 percent of the earlier value. This is a change in that premium series, not an estimate of incident frequency, world market growth, or ultimate insured cyber losses. Premium includes the combined effects of price, quantity and mix.

These two sources illustrate different denominators. A survey proportion concerns eligible organizations that identify incidents. A direct written premium series measures financial volume written by a specified insurer population. A claim frequency estimate requires exposure units corresponding to the risk population and the event definition under study. Even when geographic labels match, the denominators do not.

Frequency and ascertainment

Let E denote the exposure in policy-years, lambda the underlying eligible-incident rate per policy-year, and p the probability that an incident enters the observed dataset. Under a homogeneous rate, independent observation conditional on occurrence, and no duplicate records, expected observed event count is E times lambda times p. Therefore the observed rate estimates lambda times p, rather than lambda alone. The formula is a conventional thinning calculation; it is not a new actuarial model.

In a synthetic example, consider 1,000 policy-years with 100 eligible incidents in each of two periods. If ascertainment increases from 0.50 to 0.80, observed counts increase from 50 to 80. Observed rates rise from 0.05 to 0.08 per policy-year, a relative increase of 60 percent; the underlying eligible-incident rate remains 0.10. Improved detection, reporting, or administrative capture can therefore appear as a deterioration in risk.

If 60 eligible incidents are observed in 1,000 policy-years and p is known only to lie between 0.40 and 0.80, the implied rate under the simplifying assumptions ranges from 0.075 to 0.15. These are sensitivity bounds conditional on the assumed observation probabilities, not statistical confidence intervals. Where p varies by size and severity, a single scalar correction is inadequate. If incidents with larger losses are more likely to be reported, the observed severity distribution is also distorted.

Exposure must match the numerator. Organization-years, policy-years, insured revenue, device-years, critical-service hours, and technology dependency exposures serve different research questions. A provider outage count divided by the number of customer policies is not an estimate of the provider's outage intensity. A policy-year claim count can exceed one per policy-year; an annual proportion of policies with any claim cannot exceed one. Where multiple policy layers cover the same entity, aggregate claims frequency may count layers unless the entity or event linkage is retained.

Peer-reviewed evidence and selection

Biener, Eling, and Wirfs (2015), in Insurability of Cyber Risk, analyze 994 cyber cases drawn from operational risk data and address insurability challenges, including data limitations and dependence. The sample is informative about the kinds of incidents represented in its source, not a census of contemporary insurance claims. Romanosky (2016), in Examining the Costs and Causes of Cyber Incidents, analyzes public data on more than 12,000 incidents over 2004–2015. The availability of cost measures limits which observations enter parts of the cost analysis. Publicly recorded events, operational risk reports, and insurance records represent different ascertainment mechanisms.

The common lesson is methodological: differences in sample composition may dominate differences in event labels. A mapping can align classification variables, but cannot make unequal selection mechanisms equal. Stratification or weighting may help only when relevant selection information is available. A regression on reported incidents cannot recover the distribution of unobserved ones merely by specifying many covariates. Nor does an association between incident categories and observed costs establish causal effects of those categories.

Loss development and reporting boundaries

A claim observed at early development may include a small initial reserve and subsequently become severe as interruption, litigation, or remediation develops. A closed claim can reopen. Severity comparisons therefore require common development horizons or suitable adjustment models, with assumptions disclosed. Paid and incurred triangles can support analysis when records have consistent accident and development periods. The use of triangles for cyber claims is constrained when definitions or coverage mix change sharply over time.

A multi-source comparison should define the valuation date, claim closure, treatment of open claims, deductible basis, limits, and recoveries. Economic loss datasets often lack the contractual and accounting state necessary for these adjustments. A policy-level analysis may not contain the gross underlying losses, particularly when claims fall below deductibles or are not reported. Censoring, truncation, and missingness are distinct: a limit may censor covered severity, a reporting threshold can truncate small events, and an unknown gross loss is missing. Each affects estimation differently.

Evidence Validation and Sensitivity Analysis

An annotated synthetic portfolio

The following entirely fictional portfolio illustrates four affected-entity incidents and their associated legal and financial records. All amounts are thousands of US dollars. Each entity has an exposure for the observation period. A and B are affected by the same accidental provider interruption; C experiences payment fraud; D experiences ransomware. A fifth enterprise encounters a blocked attempt without adverse impact and contributes exposure but no incident under the chosen adverse-effect definition. Assume that B's interruption fails its particular policy trigger and that it submits no claim. No universal rule of coverage is assumed.

Entity Originating episode Gross economic loss Eligible amount before deductible Deductible Applicable cap Indemnity
A Shared interruption 180 140 20 200 120
B Shared interruption 60 0 Not applicable Not applicable 0
C Fraud 100 90 10 100 80
D Ransomware 260 220 20 150 150
Total Three episodes 600 450 50 applied Not additive 350

The calculation assumes a deductible applied first and then a cap on indemnity, independently for each covered claim. Entity D's uncapped amount after deductible is 200, but its assumed limit reduces it to 150. Caps have not been added because they are not realized losses and may interact with other terms outside this example. Gross economic losses sum to 600; eligible losses sum to 450; indemnities sum to 350. The difference of 250 between gross economic loss and indemnity includes uncovered or ineligible loss, deductibles, and the effect of the D limit. These arithmetic categories are valid only under the stated assumptions.

Suppose D's eligible amount comprises 60 forensic expense, 100 restoration and 60 extortion expense. Its contract-level deductible and limit must be allocated for coverage-component reporting according to an explicit rule. If allocated proportionately solely for this illustration, the component indemnities are 150 times each eligible component divided by 220: approximately 40.909, 68.182 and 40.909, which sum to 150. Allocation under actual policy accounting could instead follow contractual terms or claim-handling practice.

The observations support different counts: three originating episodes, four affected-entity incidents, three submitted claims, and five eligible expense or loss components (one each for A and C and three for D). A deduplicated episode frequency is therefore not the same as affected-entity frequency or claims frequency. If a register counts claim-component rows as claims, it reports five rather than three. At fixed exposure, that overstates counts by two thirds. The ratio is an artifact of record granularity.

A further assumption can demonstrate the time dimension. If 200 of the total indemnity has been paid at valuation and 150 remains as an outstanding case estimate, reported incurred indemnity at that date is 350. The sum of the 350 incurred snapshot and 200 paid movement would double-count the paid component. A later incurred estimate of 360 establishes a 10 change relative to that snapshot, not another 360 of loss. The example excludes IBNR, expenses outside the stated indemnity, reinsurance and other recoveries; they would require separately specified bases.

Waiting-period sensitivity

A loss category labeled business interruption can encode materially different quantities. Consider a synthetic ten-hour failure with a constant covered margin loss of 10,000 dollars per hour and no other expenses, limits, or deductibles. Under an assumed eight-hour time deductible, losses after the first eight hours amount to 20,000 dollars. Under an assumed eight-hour franchise condition that pays from onset when the threshold is exceeded, the corresponding amount is 100,000 dollars. This fivefold difference follows solely from the assumed terms. Both claims could share a common technical outage code and the same measured interruption duration.

Actual forms can apply distinct waiting periods to different coverage components, specify qualifying interruption, or constrain start and end of the indemnity period. The specimen comparisons in the Event Classification chapter show why the financial measure needs a contract reference. Analysts who lack that information can preserve duration as a comparable technical attribute while withholding claims that the payable amounts are comparable. Coding should support partial comparability rather than require the impossible.

Classification uncertainty and reliability

Suppose a sample of 100 independently identified entity incidents has 30 confirmed malicious causes, 50 confirmed nonmalicious causes, and 20 unresolved causes. The fraction confirmed malicious is 30 percent. If unresolved cases could belong to either category, the malicious fraction is bounded between 30 and 50 percent under the assumed exhaustive dichotomy. Excluding unresolved records yields 30 divided by 80, or 37.5 percent among resolved cases. That value estimates the proportion among the resolved subset, not necessarily the whole sample. The bounds require no assumption about how missing cause labels are generated; a narrower estimate would require additional evidence.

Crosswalk quality can also be assessed by independently coded records. In a synthetic set of 100 records, two coders agree on positive classification for 40 and negative for 45; coder A alone assigns positive to 10, coder B alone to 5. Observed agreement is 85 percent. A's positive fraction is 0.50 and B's is 0.45. Expected agreement under independent marginal assignments is 0.50 times 0.45 plus 0.50 times 0.55, or 0.50. Cohen's kappa is (0.85 - 0.50) / (1 - 0.50) = 0.70. This does not establish accuracy against the true label. It measures chance-adjusted agreement under that convention. Low-frequency classifications and class imbalance warrant separate confusion matrices and adjudicated reference cases.

A useful validation protocol selects records from common categories, rare high-severity categories, and ambiguous boundaries, preserves the initial codes, performs independent review, and resolves discrepancies by documented rules. Aggregate metrics should not conceal financially critical mapping errors. A small misclassification rate among very large losses can be more consequential for aggregate modeling than a large rate among near-zero amounts. Count-weighted and loss-weighted diagnostics answer different questions.

Dependence and aggregation sensitivity

Let originating episodes follow a Poisson count with intensity lambda over a specified period. Assume independent identically distributed episode aggregate severities Y, independent of the episode count. The compound Poisson aggregate S has expectation lambda times E[Y] and variance lambda times E[Y squared]. This standard result follows from conditioning on the count and applying the law of total variance.

Compare two hypothetical portfolios with equal expected aggregate loss. In the first, each originating episode causes two simultaneous losses of 100 monetary units, so Y is fixed at 200 and the episode intensity is lambda. Expected aggregate is 200 lambda and variance is 40,000 lambda. In the second, losses of 100 arise independently with total intensity 2 lambda, giving expected aggregate 200 lambda and variance 20,000 lambda. The first portfolio's variance is twice the second's. The calculation relies on identical fixed severities and a simple independence structure; actual cyber losses may have heterogeneous severities, variable footprints and non-Poisson clustering.

The purpose is to demonstrate that losing the common-origin linkage can alter estimated aggregate dispersion even when observed total counts and expected losses are preserved. Dependence is especially relevant in reinsurance and solvency analysis. Common-origin coding alone cannot determine the degree of future dependence, because it is subject to incomplete linkage, evolving technologies and selection. A portfolio model also needs evidence of current dependency exposure and appropriate scenarios.

Technical and scientific validation

A reliable exchange requires separate tests for format integrity, semantic integrity, contractual meaning, and statistical fitness. These are different checks, not substitutes for one another.

Validation Example test Failure consequence
Format Required fields, dates, currency codes and types Record rejected or quarantined
Identity Unique namespace and source ID; duplicate linkage Counts become unreliable
Classification Multi-label completeness, version, crosswalk direction Category rates become incomparable
Contract Trigger, wording version, claim status and component Covered amount may be misrepresented
Finance Movement reconciles with snapshot on stated basis Aggregate may double count
Population Numerator and exposure use same scope Rates lack interpretable denominators
Provenance Original values and transformation lineage preserved Result cannot be reproduced
Uncertainty Unknown, estimated and confirmed remain distinct False precision and biased estimates

Validation should incorporate reconciliation across record levels. The number of component records can exceed the number of claims, but each allocated amount should reconcile to the claim amount on the stated basis. A shared episode can have many affected entities, but a duplicated source observation should not create an additional entity incident. Exceptions should be recorded rather than discarded. Data retained for operational investigation may legitimately contain records that are excluded from actuarial estimation.

A quantitative pipeline should retain a versioned extraction specification: the source snapshot dates, inclusion criteria, join keys, time conventions, deduplication logic, mapping version, financial treatment, and population definition. The original dataset and a reproducible transformation should be accessible within the appropriate permissions. A public scientific report may disclose the specification and aggregate diagnostics without exposing confidential records. When reproducibility is constrained by privacy or licensing, the boundary should be explicit.

This study's own sensitivity analyses are reproducible from the numbers in the preceding sections. They do not estimate real-world probabilities or dependence; an empirical application requires the relevant claims and exposure records. The documentary comparisons establish differences in definition and purpose, while the examples demonstrate the mathematical consequences that can follow from those differences.

Integrated Scientific Synthesis and Research Agenda

Interoperability as preservation of meaning

The source comparisons and the numerical analyses yield a consistent conclusion. The central task is to preserve the distinctions that make cyber risk measurable. A security observation, the incident it supports, the affected entity, the insured relationship, the claim, and the payment are linked, but their meanings are not interchangeable. The analyst should aggregate along an expressly defined relationship and specify the unit governing the result.

The standards examined address complementary parts of this task. Threat-intelligence and telemetry representations provide detailed technical observations. Incident classifications provide structured information about effects and causation. Reporting frameworks address the needs of authorities and regulated institutions. Operational risk classifications relate events to business consequences. Insurance records supply contract-specific and financial distinctions. None of these uses establishes that a code in one system is a universal equivalent of a code in another.

The author's analytical contribution is the integration of these documentary distinctions with the statistical implications of grain, selection, timing, and dependence. This synthesis demonstrates why loss of semantics can generate apparent trends, misleading frequencies, inconsistent severity estimates, and materially different aggregate variances. The conclusions follow from comparative analysis and established statistical reasoning; no unique priority or new named theory is asserted.

Selecting an analytical unit for an insurance question

An underwriting study may start at the exposed organization or policy level and relate incident incidence to the characteristics of its digital dependencies. A claims study may use claim or coverage-component records and explicitly condition on observed claims. Reinsurance research may require a contractual occurrence or originating episode and then aggregate individual claims under treaty conditions. A resilience study may operate at service level and measure duration of disruption. Choosing one unit for every purpose would create avoidable distortions in at least some of these applications.

The exchange can support all four uses when it preserves the relations among records. For each published statistic, the provenance should make it possible to identify the relevant extract and selection rule. When a dataset cannot support a use, the correct result is a limitation of inference. It should not be repaired by assigning an invented denominator or by assuming that uncoded events did not occur.

An especially important distinction concerns coverage gaps. An incident with no insurance payment may reflect absence of insurance, an excluded event, a deductible, a pending decision, voluntary non-notification, or absence of economic harm. Each has different significance. Counting all unpaid incidents as denied claims would misstate insurance performance. Counting all events in a public security dataset as uncovered claims would misstate the insurance protection gap. Diagnosis requires policy and claims information in addition to the incident.

International comparability and legal pluralism

Cross-border analysis is possible without harmonizing national law. The exchange should preserve original jurisdiction-specific reporting codes and identify the common analytical fields that can validly be compared. Materiality requirements, deadlines, legal claims, insurability, and coverage terms remain attached to their legal contexts. For example, downtime in hours may be comparable across organizations under a common service definition even when reporting obligations and coverage triggers differ. A financial impact recorded in a regulatory band may support ordinal comparison but does not yield an exact amount for claims modeling.

An international aggregation should also distinguish currency conversion and nominal versus real values, group and subsidiary reporting, and jurisdiction of affected services versus policy placement. The resulting dataset can deliberately provide different levels of comparability: exact for well-defined attributes, bounded for partial mappings, and unresolved for incomparable ones. A scientifically defensible exchange admits all three outcomes instead of coercing all records into exact equivalence.

Operational implementation and incentives

Institutions that supply data have different incentives. A claims department needs reliable coverage and settlement records; an incident-response team needs clear technical evidence; a regulator needs timely and consistent reporting; an underwriter needs exposures and informative outcomes. An exchange that imposes excessive manual recoding can increase delays and error, while an automated crosswalk that suppresses context can create false confidence. Data specifications should exploit existing source fields where possible and add only information necessary for agreed analytical uses.

A governance process should assign ownership of definitions, version updates, quality exceptions, and access decisions. Review can focus on high-impact errors and categories where mapping uncertainty is greatest. Participating systems need a method to report corrections without overwriting the historical trail. The costs of standardization include engineering, training, data minimization, and maintenance. Its benefits include reduced repeated reconciliation, greater comparability, and clearer uncertainty. These are mechanisms to assess empirically, not numerical benefits estimated by this monograph.

Data linkage should be governed independently of widespread disclosure. Insurers may exchange bounded aggregate results or pseudonymous identifiers through trusted arrangements. The decision to share a record should take account of contractual confidentiality, privacy, security, and competition constraints. Broader visibility of confidential incident details is not required for every scientific question. The appropriate design depends on the minimum fields needed for a stated purpose, and the legal authority for processing them.

Priorities for empirical research

Future empirical studies should distinguish validation of the incident representation from validation of the contract representation. A first study could construct a permissioned sample of linked technical incidents and claim records, compare alternative coding schemes independently, and report disagreement by category and severity. A second could estimate how reporting probability varies with incident size, sector, coverage and detection capability, using an explicitly defined sampling frame. A third could examine event linkage at common suppliers and compare portfolio accumulation estimates using different validated linkage thresholds.

Longitudinal work should measure the impact of schema and reporting changes using overlapping dual-coded periods where available. Reclassification studies should publish both the original and restated series or disclose why reconstruction is impossible. Development studies should record ascertainment and valuation clocks, particularly for incidents with unresolved cause or claims. Where no ground truth exists, sensitivity analyses should show how estimates respond to plausible alternative assumptions.

Claims-level validation of insurance terms remains an essential research agenda. A sufficiently diverse sample of policy forms, endorsements, governing jurisdictions and adjudication outcomes could identify which terminology supports reliable cross-market mapping. The three specimens examined here establish variation but cannot estimate its prevalence. Licensed operational risk standards or private claim repositories could support richer crosswalk tests, subject to access and permission. The present study does not infer their contents from marketing descriptions.

The appropriate scientific goal is a progressively more interpretable body of comparative data. Its value can be measured through reproducibility, reduced reconciliation errors, clearer distinctions between economic and insured losses, and improved capacity to quantify uncertainty. Such measurement requires empirical validation; it cannot be asserted solely from the adoption of a common format.

Conclusion

Cyber loss event taxonomy and data interoperability are integral to actuarial measurement and insurance governance. The fundamental requirement is to preserve the identity and meaning of the objects that enter an analysis. International technical and regulatory standards contribute complementary evidence about events, effects, reporting, and operational risk. Policy wording and claims administration determine additional facts indispensable to insurance loss measurement.

The examination of policy specimens shows that an identical technical incident and interruption duration can have different financial consequences under different contractual terms. Official survey and market statistics demonstrate that distinct source populations and measures must retain their denominators and accounting bases. Peer-reviewed empirical research identifies the relevance of data scarcity, selection, and dependence. Synthetic sensitivity analyses establish the consequences of counting tags or coverage components as incidents, confusing incomplete reporting with changing underlying frequency, and ignoring the linkage among losses produced by a shared cause.

Practical interoperability therefore requires documented source versions, record grain, time and identity, classification mappings, legal context, financial bases, and evidence states. The original values and relationships must remain available for authorized audit. Where categories only partially overlap, comparability remains partial. Where data are unavailable, uncertainty is preserved rather than converted to a fictional observation.

The result is an integrated scientific account of the conditions under which cyber event and loss datasets can support insurance and actuarial inference. It does not estimate a proprietary loss distribution or make claims about the market prevalence of individual policy clauses. Further empirical work requires linked, permissioned claims, exposure, and technical data. Within the documentary and analytical scope established here, the conclusions follow from the evidence examined and the explicitly stated mathematical assumptions.

References

  • Financial Stability Board. (2023a). Cyber Lexicon: Updated in 2023. April 13. Basel: FSB. Source document

  • Financial Stability Board. (2023b). Recommendations to Achieve Greater Convergence in Cyber Incident Reporting: Final Report. April 13. Basel: FSB. Source document

  • Financial Stability Board. (2025a). Format for Incident Reporting Exchange (FIRE): Final Report. April 15. Basel: FSB. Source document

  • Financial Stability Board. (2025b). Format for Incident Reporting Exchange (FIRE): Taxonomy Package. Basel: FSB. Source document

  • Nelson, A., Rekhi, S., Souppaya, M., & Scarfone, K. (2025). Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile. NIST SP 800-61r3. National Institute of Standards and Technology. doi:10.6028/NIST.SP.800-61r3. Source document

  • MITRE Center for Threat-Informed Defense. (n.d.). VERIS Mapping Methodology. Mappings Explorer. Retrieved November 18, 2026. Source document

  • OASIS Open. (2021). STIX Version 2.1. OASIS Standard. Source document

  • Open Cybersecurity Schema Framework Project. (n.d.). Open Cybersecurity Schema Framework: Core Schema Repository. Retrieved November 18, 2026. Source document

  • ORX. (2026). Cyber Operational Risk Reporting Standards (C-ORRS). February. ORX. Source document

  • ORX. (n.d.). ORX Operational Risk Reference Taxonomy. Retrieved November 18, 2026. Source document

  • European Union Agency for Network and Information Security. (2017). Commonality of Risk Assessment Language in Cyber Insurance. ENISA. Source document

  • European Commission. (2024). Commission Delegated Regulation (EU) 2024/1772 of 13 March 2024 supplementing Regulation (EU) 2022/2554 with regard to regulatory technical standards specifying the criteria for the classification of ICT-related incidents and cyber threats, setting out materiality thresholds and specifying the details of reports of major incidents. Official Journal of the European Union, June 25. Source document

  • U.S. Securities and Exchange Commission. (2023). SEC Adopts Rules on Cybersecurity Risk Management, Strategy, Governance, and Incident Disclosure by Public Companies. Press release 2023-139, July 26. Source document

  • Australian Prudential Regulation Authority. (2018). Prudential Standard CPS 234 Information Security. Determination No. 1 of 2018, November 30; effective July 1, 2019. Source document

  • CyberAcuView LLC. (2022). CyberAcuView Cyber Insurance Wording. Base policy specimen, reproduced by Lloyd's Market Association. Source document

  • Victor Insurance. (2021). Cyber Secure: Policy Wording. Version 3, July 1, 2021. UK specimen. Source document

  • Corvus Insurance. (n.d.). Smart Cyber Policy Form. Specimen CB-101-001. Retrieved November 18, 2026. Source document

  • Lloyd's and Association of British Insurers Cyber Working Group. (n.d.). Components of a Major Cyber Event: A (Re)insurance Approach. Retrieved November 18, 2026. Source document

  • CrowdStrike. (2024). Channel File 291 Incident: Root Cause Analysis Is Available. August 6. Source document

  • Weston, D. (2024). Helping Our Customers Through the CrowdStrike Outage. Microsoft, July 20. Source document

  • Department for Science, Innovation and Technology, & Home Office. (2026a). Cyber Security Breaches Survey 2025/2026. UK Government. Source document

  • Department for Science, Innovation and Technology, & Home Office. (2026b). Cyber Security Breaches Survey 2025/2026: Technical Report. UK Government. Source document

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

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

  • Biener, C., Eling, M., & Wirfs, J. H. (2015). Insurability of Cyber Risk: An Empirical Analysis. The Geneva Papers on Risk and Insurance—Issues and Practice, 40(1), 131–158. doi:10.1057/gpp.2014.19. Advance online publication 2014. Source document

  • Romanosky, S. (2016). Examining the Costs and Causes of Cyber Incidents. Journal of Cybersecurity, 2(2), 121–135. doi:10.1093/cybsec/tyw001. Source document