RESEARCH DOSSIER
Cyber Catastrophe Event Footprints and Geographic Attribution
Abstract. Cyber catastrophes concentrate losses through common digital dependencies, transmission between organizations, and disruption of economic functions across territories. A defensible event footprint identifies affected units, locations, sectors, temporal intervals, mechanisms of impact, and the evidence supporting their inclusion. An insurance footprint adds contractual eligibility, monetary valuation, and allocation to particular insured interests. This monograph integrates historical event reconstruction, comparative incident reporting, insurance contract analysis, and actuarial dependence mathematics. Three documented episodes—WannaCry in 2017, NotPetya in 2017, and the CrowdStrike update failure in 2024—illustrate distinct mechanisms and the consequences of inconsistent observation boundaries. A synthetic portfolio demonstrates how geographic misallocation, missing dependency mappings, and contractual aggregation change the interpretation of an identical disruption. The analysis establishes that country of attack origin, country of corporate registration, location of affected operations, place of economic impact, and jurisdiction of insurance coverage answer different questions. Their conflation can distort concentration measurements, portfolio stress testing, and post-event adjustment. The conclusions support auditable, time-indexed event records with separate physical, economic, contractual, and reporting dimensions. The study uses public evidence and transparent illustrative arithmetic; it does not estimate industry-wide insured losses or fit a catastrophe frequency distribution.
Keywords: Cyber catastrophe; event footprint; geographic attribution; insurance accumulation; systemic cyber risk; business interruption; sectoral dependence; incident reporting; event aggregation; exposure data; reinsurance; uncertainty.
Introduction
Cyber catastrophe insurance requires a description of events that remains coherent from the first incident alert to the final settlement of claims. The subject of this study is the cross-peril footprint of a major digitally mediated disruption across territory, industry and time. An event footprint is a documented representation of the units affected by an event, the intensity and duration of their impacts, and the relations that connect the impacts to a common disturbance. A footprint can be expressed as a table, a network, a map or an aggregate distribution. Each representation selects a unit of analysis and a measurement objective; a valid footprint preserves these choices when data are transferred between representations.
The principal questions concern how evidence establishes geography, how industrial functions generate monetary loss from technical failures, and how divergent reporting thresholds alter what researchers and insurers can observe. The study treats accidental system failures and malicious attacks as alternative initiating mechanisms, while leaving their financial coverage to the terms of the specific contract. Deliberate state operations, cloud provider concentration, detailed reinsurance wording and model calibration each have separate research objects. Here they appear only at their points of intersection with cross-peril event identification and geographic allocation.
The investigation proceeds through structured comparative case analysis. Official incident accounts and audit findings establish observed manifestations; issuer filings establish firm-specific financial estimates; an appellate opinion illustrates the specificity of insurance terms; statutory extracts and guidance establish selected reporting criteria; actuarial literature provides the mathematical context. The cases are deliberately selected for mechanism diversity rather than statistical representativeness. The legal comparison draws on extracted text of selected EU provisions, and the 2025 threat landscape comparison draws on the report's methodology and executive summary. The 2026 edition is used only for its indexed observation-window description. Two published specimen policies are examined as illustrations of contract structure; neither proves the terms of any issued policy. The scope of each documentary check is recorded separately. No confidential claims register, complete exposure schedule, forensic image, treaty wording, or global incident census was available. All numerical examples designated synthetic are constructed for explanation, not fitted to observations.
Conceptual and Historical Foundations
The unit of observation
An event has an initiating disturbance and a set of associated manifestations. The initiating disturbance might be a malicious release, a compromised service, a faulty update, or the failure of an infrastructure component. Manifestations include the direct impairment of a system, interruptions experienced by dependent organizations, and subsequent expenses or liability. One disturbance may generate many manifestations; one manifestation may also have several contributing causes. Causal membership therefore requires evidence beyond temporal proximity.
A device is a technical unit; an establishment is an operational unit; an enterprise or legal entity is an organizational unit; a contract is an insurance unit; and a claim or loss item is a financial record. A count of one type cannot be transferred to another using an implicit conversion factor. A global company may have thousands of endpoints and a single master policy, multiple local policies, and several covered interests. Equally, one failed server may affect many independent businesses. A count of impacted endpoints can be valuable for technical response and simultaneously be a poor estimator of insured liabilities.
A footprint for a specified event and valuation date should retain the original measurement type, time interval, inclusion criterion, and denominator. For a device footprint, the denominator might be endpoints eligible to receive an update. For an enterprise footprint, it might be unique firms with a documented dependency. For an underwriting footprint, it might be in-force policies with the relevant exposures. Each denominator defines a distinct analytical population. A country share computed from insured sums measures exposure concentration; the same share computed from assessed claims measures loss concentration. They need not agree.
Dependence and catastrophe
A driver of catastrophe is correlated impact, whether through shared technology, common environmental conditions, or interorganizational transmission. Awiszus et al. (2023), in their survey of modeling and pricing cyber insurance, distinguish individually arising risk, shared risk drivers, and network-mediated systemic risk. Their discussion motivates separate treatment of common susceptibility and actual propagation. A shared vulnerable technology can generate nearly synchronous impairment across firms without any transmission between those firms. Network contagion, in contrast, creates sequential impairment through links that may extend far beyond the initial set.
Both mechanisms create accumulation even when policyholders are spread across different sovereign territories. A portfolio diversified by headquarters country can be highly concentrated by technology or service dependency. Regional distance can remain relevant through local operating hours, access to recovery staff, offline backups, distinct regulatory duties, and physical logistics. Its value is conditional on the pathway by which the initiating event becomes loss.
A conventional natural catastrophe footprint often uses a physical intensity field: wind speed, shaking, inundation depth. A cyber event may support a geographic representation, but its causal topology is frequently discontinuous. Disconnected sites can be impaired by the same software distribution channel; adjacent sites can experience different outcomes because they use different versions, configurations, or recovery arrangements. A spatial interpolation between neighboring observations is consequently a hypothesis requiring a physical or organizational mechanism.
Historical development and evidence
The three cases investigated here span self-propagating ransomware, a destructive operation with cross-border effects, and an accidental update failure. They demonstrate that large technical disruption can arise through several mechanisms, while economic and insured loss are separate observational layers. The chronology also matters: information about a disaster changes as investigators identify affected systems, accounts are finalized and contracts are interpreted. The retrospective reconstruction should preserve early estimates and their revision dates instead of silently substituting the last published number.
Intensities can be defined as binary technical impairment, downtime, fraction of capacity unavailable, records affected, or a monetary measure. Binary data are useful for membership; they omit severity. Downtime conveys persistence; it omits exposure to high-value activity. Revenue at risk indicates a financial flow, while actual business interruption loss requires counterfactual production or earnings and attention to avoided expenses and mitigation. Insurance then introduces distinct coverage conditions.
Event Geography
Geographic dimensions
An initiating actor's location answers a forensic or geopolitical question. An attack infrastructure location identifies a technical resource. The place where impairment occurred identifies an operational exposure. The place of a dependent activity identifies a possible economic loss. Corporate domicile identifies an organizational registration. The policy's territorial scope, governing law and risk location identify contractual or regulatory attributes. None of these locations is automatically a proxy for the others.
For each manifestation, an investigator should store the geography as typed information: country and, when justified, subnational area; the type of geography; evidence date; source; identification confidence; and an unknown category. Multiple geographic attributes can coexist in the same observation. The same entity may be registered in one country, operate a distribution center in another, and process data in a third. Assigning its entire disruption to the registration country produces a map of legal entities, not a map of affected operations.
Geographic allocation also depends on additive units. A single endpoint can be associated with an establishment and an enterprise, but counting all three as separate impacted organizations would inflate totals. An expense incurred centrally to restore many branches must be linked to its affected operations without duplicating the underlying value. A loss allocation across territories requires explicit weights, normalized over the selected loss item, and independent preservation of the original amount.
WannaCry and the geography of health care disruption
The English NHS case gives unusually informative denominator evidence. The National Audit Office's (2017) published investigation findings reported at least 81 affected trusts in a population of 236. The lower-bound share equals 81 divided by 236, approximately 34.3%. This figure applies to English trusts within the audit scope, rather than to all hospitals worldwide or to the English population. The audit separately reports infections in primary care and other organizations, and acknowledges incomplete observation of disruption caused by shared systems.
The direct infection map and the clinical interruption map therefore require separate denominators. When a healthy system cannot exchange information with an impaired partner, the affected clinical activity may lie outside the infection footprint. Precautionary isolation can also interrupt communication without local infection. In an insurance investigation the causal explanation for this interruption may matter as much as its duration.
A sector footprint can include service volume. The audit records 6,912 identified appointment cancellations and an estimate exceeding 19,000 overall. These measures distinguish recorded observations from an extrapolation. The report does not establish a comprehensive economic cost. Multiplying the estimated appointment count by an arbitrary procedure price would create a financial estimate with unsupported assumptions about clinical activity, marginal cost, rescheduling, and compensation. Even a carefully estimated public-service loss would require additional evidence before it could be classified as insured loss.
NotPetya and cross-border economic exposure
The UK's (2018) public NotPetya attribution identifies a destructive event directed primarily at Ukrainian sectors, with effects extending to businesses elsewhere. The attribution is a governmental assessment; it does not constitute a coverage ruling. The U.S. Department of Justice's (2020) announcement of the GRU indictment describes alleged impacts on identified victims in health care, logistics and pharmaceuticals. These records support cross-sector and international reach, while supplying different types of attribution and differing observation scopes.
The FedEx (2017) quarterly report to the SEC is particularly instructive. The issuer estimated an adverse first-quarter FY2018 earnings effect of USD 300 million, associated with decreased TNT shipment activity and restoration costs. It also stated that it had no insurance covering this attack. This historical disclosure supplies an example of substantial company-level economic impact without corresponding insurance coverage reported by that company. The amount cannot be added mechanically to an alleged combined victim total from the indictment announcement because the populations may overlap.
The appropriate geographic record associates the initiating event with Ukraine, records political attribution as a distinct sourced field, and maps affected operational locations wherever documentary evidence establishes them. The issuer-level financial estimate is retained as a consolidated amount until an auditable allocation across branches becomes available. An inability to obtain such a breakdown is represented as unknown geography for monetary allocation, not as absence of international loss.
CrowdStrike and the geometry of shared technology
The CrowdStrike (2024) technical account of the update describes a Windows configuration release at 04:09 UTC on July 19 and remediation at 05:27 UTC. The 78-minute interval identifies the eligibility window for receiving the defective configuration under the account's conditions. It does not measure the individual outage or recovery period. The vendor identifies a logic error rather than a malicious attack.
Weston (2024), writing for Microsoft, supplied an estimated device count of 8.5 million, less than one percent of Windows machines. The stated denominator is the Windows device population, not Windows users, global companies, or cyber insurance buyers. The estimate establishes scale without locating each affected site. A narrow technical exposure window and prolonged operational impairment can coexist, because distribution is relatively centralized while restoration is geographically and organizationally distributed.
A policyholder may have systems susceptible to the update but recover quickly through redundancy. Another may operate relatively few critical devices whose unavailability interrupts a large production process. Counts of endpoints consequently do not measure loss intensity uniformly. The relevant exposure for insurance is the part of the activity that depends on these endpoints and the conditions governing recovery and indemnification.
The later CrowdStrike (2024) root cause executive summary attributes the failure to a mismatch between expected and supplied input fields and the ensuing out-of-bounds memory read. Its reported recovery indicator approaches the pre-update online-sensor baseline; this measures a technical population and is not a measure of restored economic activity or indemnified loss. The proposed changes to validation and staged deployment deserve separate representation in prospective exposure data. A release cohort receiving a change before a stop decision has a different susceptibility from a later cohort. Remediation availability and completion at each endpoint are different times. A global count combined with a common release timestamp cannot therefore define a global business interruption duration.
Spatial and temporal reconstruction
A defensible timeline distinguishes initiation, exposure opportunity, first observed effect, discovery, notification, containment, restored service, resumption of normal production, and claims valuation. Times should be normalized to UTC while retaining local offsets and local working calendars. A disruption of identical length can affect markedly different amounts of activity on a working morning, overnight, or during a scheduled closure.
Spatial membership and temporal membership interact. A global update may reach regions during different business hours; a dependent supplier may transmit economic effects only after inventories are depleted. The event map at one hour and the map at one week describe different states. A cumulative footprint includes all verified manifestations during the specified observation window; a snapshot includes the current active impairment. Neither determines the contractual occurrence by itself.
Sector Impact
From technical damage to economic damage
The economic consequence of an interruption depends on what the affected service does, how rapidly substitutes become available, how much production can be rescheduled, and the cost of restoration. An affected hospital may face patient diversion and cancellations; a logistics service may face unprocessed shipments and backlog; a manufacturer may face production stoppages or loss of access to operational systems. The mechanisms differ even where the technical cause is identical.
A sector code is an imperfect proxy for function. A diversified enterprise with retail, manufacturing and logistics activities may require multiple operating-unit classifications. Assigning the whole enterprise to a single industry is justifiable for a legal entity count, but less informative for a dependency footprint. When mapping business activity, exposure records should preserve the primary classification and the specific function affected. A crosswalk between classifications should retain its version and mapping ambiguity.
Dependencies can be represented as directed links: a consuming organization relies on a supplier or internal service for a particular function. The link is material when loss of that service constrains production, payment, or delivery beyond available mitigation. A network edge can carry duration sensitivity, substitution availability and service capacity. These attributes matter because a backup supplier that becomes useful after three days offers little protection against an interruption with a two-hour tolerance but substantial protection against a month-long disruption.
The sectoral modeling discussion in Awiszus et al. (2023) supports the use of connected industry structures as research proxies. Input-output relationships describe economic exchange, however, and should not be interpreted as verified technological dependency at firm level. Direct mapping of software, network services and operational processes remains important when estimating contractual exposure.
Financial transformation and coverage
Economic loss is the difference between a realized financial trajectory and an appropriately specified counterfactual attributable to the event, including relevant mitigation and restorative costs. For commercial activity, a simplified interruption calculation may use lost contribution margin after avoided variable costs, rather than gross revenue. Actual contract wording may define gross profit, extra expense or other measures differently. Recovery expenditure should be counted only once when constructing an aggregate economic amount.
An insured loss is a contractually eligible monetary consequence after applying the relevant coverage terms. The process may involve establishing an insured interest, trigger, causation, territorial scope, waiting period, deductible, exclusion, sublimit, and limit. These features depend on actual wording and applicable law. A model can represent their economic implications but must preserve the distinction between estimated eligibility and adjudicated coverage.
The 2023 New Jersey Appellate Division opinion in Merck v. ACE American illustrates the boundary between public actor attribution and the legal construction of an exclusion. The dispute concerned the hostile or warlike action exclusion in all-risk property coverage for NotPetya losses. The appellate court affirmed the partial summary judgment on that exclusion. This ruling concerns the wording and facts in the litigation; it supplies neither a worldwide rule for cyber war exclusions nor a conclusion about all other policies. The analysis does not rely on a subsequent settlement or infer the outcome of every coverage issue.
In contrast to property-specific wording, a standalone cyber contract may distinguish security failure and system failure, or provide dependent interruption coverage under separately defined conditions. These are illustrative categories, not assertions that every market form contains them. An accidental outage can enter a technical catastrophe inventory while remaining outside a policy triggered solely by specified attacks. Direct and dependent interruption must be recorded as different causation relationships, with any applicable supplier restrictions and waiting periods.
Contract evidence and the location of covered interest
The AXIS (2020) cyber insurance specimen distinguishes first-party business interruption associated with the insured network from interruption involving a service provider. The latter agreement depends on an eligible provider network, measurable interruption to the insured's business, a duration threshold and discovery during the policy period, with applicable retention and limits. The provider definition refers to contractual service relationships and specified kinds of services. This makes the relationship between the affected provider and protected enterprise central to the coverage footprint. A geographical map of failed networks does not establish this relationship, and a list of insured headquarters does not reveal which providers serve their affected business operations.
The BOXX Insurance (2024) Canadian cyber policy specimen makes a different set of distinctions explicit in agreements B.5 through B.8: interruption arising from security events or system failure, and interruption involving the insured or outsourced provider. Corresponding waiting periods are specified separately in the declarations. Section IV.1.3 also provides that the insured bears losses during the waiting period and, for specified interruption agreements where the waiting period has been borne, need not separately pay the stated retention. This illustrates why waiting periods and monetary deductibles in an actuarial example cannot be treated as universal concurrent terms. The synthetic example below deliberately imposes both; it is not a reconstruction of this specimen.
Two enterprises with the same interruption time and income loss may therefore produce different eligible loss components. Establishing whether a provider's failure is directly covered, contingently covered, excluded, or subject to a separate limit is a contract-level task. The location of that provider may be different from the location of all affected insured operations. The declarations, endorsements and applicable law must be consulted before attributing loss to any particular contract. These two publicly accessible forms demonstrate possible structures rather than their market prevalence. Neither specimen supplies the wording of a contract actually implicated in the historical events.
Aon's (2024) CrowdStrike reinsurer briefing describes the relevance of waiting periods and dependent interruption provisions and distinguishes revenue delayed by temporary service loss from revenue ultimately lost. A canceled booking, deferred shipment and rescheduled appointment can have different economic effects even if initial downtime is identical. Lost income is measured against a counterfactual and may be mitigated later; additional costs can continue after operations resume. An initial sector count is thus a measure of operational exposure to disruption, while valuation and coverage assessment require later observations.
Portfolio exposure and aggregation
Portfolio event loss depends on both the set of impacted exposures and the transformation of each exposure into a covered amount. A policy protecting an enterprise group may aggregate units across regions. Conversely, a multinational group may have separate local policies and umbrella limits. A geographic report at insured-unit level can help estimate loss but must reconcile those results to the actual policy structure before summation.
For a fixed scenario e, let X_i(e) denote the gross covered amount for policy i after applicable terms. Total gross portfolio loss is L(e) = sum_i X_i(e). If exposure i refers instead to a site, the site loss components must first be combined under the appropriate policy limits. Applying a policy limit independently to each site and then adding the results can overstate portfolio loss; applying it globally without authority can understate coverage. The portfolio data model must record this hierarchy.
For uncertain event loss X_i with finite second moments, Var(sum_i X_i) equals the sum of individual variances plus twice the sum of pairwise covariances. The covariance terms expose the importance of common causes and transmission. Geographic separation reduces concentration only to the degree that separated exposures have genuinely less dependent losses. A portfolio with many countries may have little diversification against a globally distributed common technology failure.
An excess-of-loss treaty may contain a further definition of occurrence, attachment, limit, hours condition, exclusions, and allocation rules. The insurer's gross event footprint is therefore an input to analysis rather than a direct determination of recoverable reinsurance. The next section considers why observation and reporting boundaries matter even before these contractual transformations.
Reporting Boundaries
Different observation mandates
Incident disclosure, operational-resilience reporting, security telemetry, insurance notification and claims accounting produce distinct observational systems. Each selects events using a mandate and threshold. Security telemetry may count attempts or devices. Prudential reporting may focus on disruption of critical functions. Securities reporting selects information material to investors. Claims records reflect policy coverage, reporting behavior and valuation stages. Comparing unharmonized counts can yield an apparent trend that reflects selection rather than changing risk.
In the United States the SEC's (2023) incident disclosure compliance guide explains that domestic registrants generally make the relevant Form 8-K disclosure within four business days of determining materiality, not within four days of the initial disturbance. Foreign private issuer reporting has a distinct Form 6-K mechanism. The population is securities registrants covered by the relevant rules, with an investor-oriented materiality criterion. Incident date cannot reliably be reconstructed by subtracting four days from a filing date.
The EU's (2022) Digital Operational Resilience Act uses geographic spread among its ICT incident classification considerations under Article 18. This criterion operates alongside affected clients and transactions, duration, data loss, service criticality and economic impact. The Commission's (2024) incident classification technical standards specify in Article 4 that assessment considers effects on clients or financial counterparties, branches and group entities, and market infrastructures or third-party providers in other Member States. Article 9(4) sets a threshold for geographic spread at impact in at least two Member States. A multinational footprint can therefore be relevant even when a single installation initiates the disruption.
The threshold is a component of the classification decision, not a self-sufficient major-incident determination. Article 8 combines impact on critical services with the specified data-loss threshold in Article 9(5)(b) or the required combination of other materiality criteria. Article 9 also supplies duration and economic-impact tests, including an economic threshold above EUR 100,000 and, for critical or important functions, service downtime above two hours. These criteria apply within the instrument's defined population and rules. A single geographic threshold cannot be transferred to an insurance occurrence definition. Recurring incidents may require combined assessment under Article 8(2), subject to its same-apparent-root-cause, time-window and applicability conditions. That regulatory aggregation serves the classification mandate; policy or treaty event aggregation remains governed by its own text.
Under Article 23 of the EU's (2022) NIS 2 Directive, a significant incident is assessed by serious operational or financial effects or considerable damage to other persons. The directive provides an early warning within 24 hours and an incident notification within 72 hours of awareness, with a specific 24-hour notification exception for relevant trust service providers. A final report generally follows within one month of incident notification, subject to the ongoing-incident provision. Cross-border effects are relevant to the information transmitted. Awareness, assessment, reporting and actual impact period remain distinct. The comparison is at directive level, not an assertion of uniform national implementation.
Threat landscape counts and insurance data
ENISA's (2025) Threat Landscape 2025 examines 4,875 selected incidents over July 2024–June 2025. Its methodology combines publicly available material with voluntarily shared information; delayed disclosure, incomplete records and ambiguous geographic and sector identification constrain interpretation. The indexed description for the 2026 threat landscape identifies a January–December 2025 observation period; the full 2026 report was not examined. The windows overlap but differ. Counts cannot be converted into claims frequencies without matched population, selection process and coverage evidence.
Selection also changes the relative frequencies of mechanisms. Incidents producing visible disruption are more readily recorded by open sources than events remaining undetected for long periods. Public claims of responsibility, service-provider notifications, and verified affected entities are distinct records. Insurer data are selected by coverage, reporting incentives, policy penetration and loss development. Comparing the two sources requires a common event definition and a denominator appropriate to each. Reweighting an incident sample by national gross domestic product does not make it an insurance exposure sample.
The EIOPA (2026) insurance risk dashboard separates insurers' own operational exposure from the cyber risks they underwrite and describes supervisory assessment and publicly reported incident frequency among its indicators. This illustrates why insurer resilience and insured accumulation merit separate reporting fields. The disruption of an insurer's claims platform may delay data on losses from an event simultaneously affecting its policyholders, creating a feedback between event severity and reporting timeliness.
Insured loss registers should distinguish notice-only records, open claims, denied claims, paid amounts, case estimates and incurred-but-not-reported estimates. A claim reported at one date may be reassessed later. Reporting by occurrence date answers a different question from reporting by notice date or accounting period. A footprint estimated as of month-end needs a valuation date attached to each claim measure.
Selection, missingness and geographic classification
Let A be the true set of manifestations under a declared definition and B the set observed through a particular reporting channel. Even if B is a subset of A, the observed share is not an unbiased estimator unless inclusion probabilities and denominators are known. Some observations are easier to obtain in large enterprises than in small firms; some sectors publish detailed assessments while others have restricted disclosure. A classification of absent reports as zero impact systematically favors jurisdictions with weak observation.
A reporting threshold can create left truncation: small losses are absent. A recent event may exhibit right censoring: restoration time is not yet fully observed. Monetary claims may also be immature because investigation is ongoing. These effects influence both geographic comparisons and model calibration. Event membership should use categories such as documented, plausible pending corroboration, contradicted, and unknown, with the inclusion criteria explicit.
If a researcher has a fixed observed lower-bound set and a larger independently established exposure universe, simple bounds may be defensible. For example, if 20 of 100 eligible organizations are observed affected and 30 are verified unaffected, then the affected share lies between 20% and 70%, assuming all 50 unclassified organizations belong to the same universe and no known classifications are erroneous. This interval represents ignorance about missing observations, not a statistical confidence interval.
Independent treatment of unknown location is equally important. If losses with a known value but unresolved location sum to U, the amount allocated to a given country can range from its confirmed allocation to that allocation plus U, absent tighter constraints. The width of this interval reveals how much a geographic conclusion depends on data incompleteness.
Evidence Validation and Sensitivity Analysis
Documentary verification and reconciliation
The forensic and financial records should be assembled with provenance: event identifier, source type, publication date, effective period, measurement unit, confidence, and claim-to-source pointer. Raw records remain unchanged; corrected interpretations are versioned. A source may be authoritative for one fact and weak for another. The vendor has direct knowledge of a software release but may lack independent information about customers' financial loss. An issuer knows its consolidated estimate but may not disclose its branch allocations. A court opinion establishes a ruling, not the complete insurance market.
The WannaCry observation demonstrates the need to distinguish direct infection and service disruption, while the NotPetya issuer disclosure distinguishes earnings impact from covered loss. The CrowdStrike technical window separates potential receipt of a defective configuration from actual downtime. These forms of evidence are complementary because they answer different questions. They are not independent estimates of the same latent loss.
A reproducible validation procedure begins by defining an event and as-of date, then identifies the exposure universe. It tests candidate manifestations for temporal and causal compatibility, reconciles entities and sites, classifies geography, and applies valuation and coverage information separately. Conflicting evidence remains visible with its date, type and scope. A revised event hypothesis may change membership without altering the original records.
A geographic record should undergo checks for duplicate legal entity identifiers, misplaced addresses, ambiguous service-region tags, duplicated multinational group totals, incompatible currencies and missing offsets. A financial record should be reconciled to ledger totals and policy-specific allocation rules. Model validation should test the sensitivity of the output to alternative assumptions about missing geography, propagation and recovery, with distinct parameter and structural uncertainties.
An illustrative multinational portfolio
Consider four distinct policy records, in thousands of USD, under a deliberately simplified first-party interruption coverage assumption. A common event affects their operations. The table shows registration country, actual affected operating country, covered economic basis before waiting or monetary terms, interruption duration, and monetary conditions. The values are synthetic. Each record has one affected operation and one policy; expenses are already net of avoided costs. There is no annual aggregate exhaustion and no currency conversion.
| Policy | Registration | Operation | Sector | Loss basis | Hours | Deductible | Limit |
|---|---|---|---|---|---|---|---|
| A | US | Germany | Logistics | 120 | 24 | 10 | 100 |
| B | US | US | Health | 80 | 12 | 10 | 70 |
| C | Germany | Poland | Manufacturing | 200 | 40 | 20 | 150 |
| D | Netherlands | Netherlands | Retail | 60 | 6 | 5 | 60 |
Assume an eight-hour waiting period operates as a strictly time-proportional deductible before the monetary deductible; loss accrual is uniform over the recorded hours. Denote loss basis by b, hours by h, waiting hours by w, monetary deductible by d, and limit by k. For this illustration, X = min(k, max(0, b times max(0, h-w)/h minus d)). This simplified mapping is mathematically explicit and is not a representation of a named policy form.
Policy A: 120 times 16/24 minus 10 = 70. Policy B: 80 times 4/12 minus 10 = 16.667. Policy C: 200 times 32/40 minus 20 = 140. Policy D has no eligible duration beyond the waiting period and produces zero. Total economic basis is 460; total modeled gross covered loss is 226.667. The difference is due solely to the stated illustrative assumptions and is not a real-world protection-gap estimate.
Attributing the 226.667 modeled covered loss by registration gives the US 86.667 (38.24%), Germany 140 (61.76%), and Netherlands zero. Attributing it by affected operations gives Germany 70 (30.88%), the US 16.667 (7.35%), Poland 140 (61.76%), and Netherlands zero. Rounding explains a 0.01 percentage-point difference in the total. The portfolio total is conserved in both maps, while the geographical concentration is materially different. A manager assessing US operational interruption would reach an erroneous conclusion from the registration map alone.
The country Herfindahl concentration index is the sum of squared country shares. The registration allocation yields approximately 0.528; the operational allocation yields approximately 0.482. This index describes the concentration of the selected loss allocation, not network independence, residual risk or geographic causation. Both values arise from the same event, policies and losses. The numerical contrast is therefore a pure attribution effect.
Alternative waiting periods and event definitions
Changing the waiting period to zero while preserving the monetary deductibles and limits yields A = 100, B = 70, C = 150, D = 55, for total 375. Changing it to 12 hours yields A = 50, B = 0, C = 120, D = 0, for total 170. These differences show why an event footprint with only a yes/no affected indicator cannot determine covered losses. A model must distinguish time-based deductible mechanics from franchise-style waiting conditions; the calculations here assume only the former.
Now suppose the original eight-hour losses are all contractually one occurrence under a hypothetical treaty with a 100 attachment and 100 occurrence limit, and no other terms. The recovery is min(100, max(0, 226.667-100)) = 100; the net retained gross amount is 126.667. Suppose instead that the same gross amounts are split by a valid two-occurrence interpretation: A plus B = 86.667, and C plus D = 140. The recovery becomes zero plus 40 = 40, leaving retention 186.667. No geographic or temporal split establishes a legal occurrence without the applicable contract. This illustration isolates aggregation mechanics rather than recommends any particular treaty wording.
If C's operational location is unknown, the confirmed operating-country allocations account for 86.667 out of 226.667, and the remaining 140 is unallocated. Germany's share then lies between 30.88% and 92.65%, without further information. Substituting C's headquarters country would select the upper bound without evidence. Holding an unresolved category makes the uncertainty transparent and preserves the portfolio total.
Sensitivity to allocation uncertainty
A useful validation procedure computes bounds before assigning unlocated observations. For each country let the lower bound equal verified allocated loss and the upper bound equal that loss plus unresolved loss potentially attributable to the country. Bounds for distinct countries cannot be summed as if all upper limits were simultaneously realizable. For multi-country operations, evidence of allocation shares, including production records and cost attribution, can narrow the feasible set while conserving aggregate loss. Corporate registration may serve as an explicitly labeled proxy only if a sensitivity comparison demonstrates the consequences of that assumption.
Consider a comparison of dates immediately following disruption and after partial claim development. The cohort of identified incident participants can be held fixed while claim values change; alternatively new entities can enter the event database. The first comparison estimates development on known observations, the second changes observation coverage. Neither independently measures the final incident population. Separate vintage fields permit analysts to reconstruct the state of information available at an earlier date and test whether a revised classification changes reported concentrations.
Sensitivity to a shared cause
Assume two policies each incur a fixed loss of one unit when impacted and zero otherwise. Each has probability p = 0.1 of impact over a hypothetical period. Under independence, expected total loss is 0.2, variance 0.18, and probability that both are impacted is 0.01. Under perfect common-cause dependence, expected total loss remains 0.2, variance becomes 0.36, and probability that both are impacted becomes 0.1. The equality of mean loss does not imply equality of tail exposure.
These synthetic examples demonstrate why observed loss by jurisdiction and average frequency are insufficient for catastrophe capital assessment. Estimating actual tail probabilities requires an appropriate event catalog, exposure data and defensible parameterization. Extreme event scenarios can compare mechanisms and expose concentration blind spots, even when occurrence probabilities are presently weakly identified. Reporting scenario conditional loss separately from an assigned annual exceedance probability helps maintain that distinction.
Integrated Scientific Synthesis and Research Agenda
Geographic allocation as an evidentiary operation
Geography in cyber insurance serves multiple functions. It describes technical manifestation, economic consequence, legal exposure, portfolio concentration and regulatory impact. Each use selects an appropriate location attribute. The central analytical finding is that an attribution error can change the apparent concentration of a portfolio without changing its underlying event loss. Conversely, a technology dependency omitted from the exposure inventory can change the loss estimate itself. These errors require different remedies.
The first remedy is better classification and conservation of amounts. Every allocation should have a defined denominator and satisfy a balance check. Multiple nonadditive views can coexist, provided their relationships are documented and they are never summed as though independent. The second remedy is improved observation and modeling of causal exposure: dependencies, availability and recovery. Better geographic labels cannot compensate for a missing critical dependency.
The historical cases expose the importance of this distinction. A health system's observed count of affected organizations is sensitive to how indirect disruption is measured. A destructive international operation produces both insured and uninsured economic impacts depending on the individual entity and contract. A widely deployed update can generate broad technical eligibility but heterogeneous loss duration. These mechanisms determine which missing fields are material.
Implications for underwriting and claims administration
An underwriting exposure schedule should capture the insured's operational footprint, material technology dependencies, operating units, coverage territory, valuation basis and policy hierarchy. Data collection must be proportionate to the analytical need and protect sensitive infrastructure information. The purpose is effective concentration assessment rather than exhaustive identification of every device.
The high-level outcome taxonomy in Pascoe, Quinn and Scarfone's (2024) NIST Cybersecurity Framework 2.0 provides useful organizational context for assessing cybersecurity activities. Its abstract explicitly describes flexible outcomes rather than prescribed implementation. The framework can support structured dialogue about readiness, but a self-reported framework alignment should not be used as direct proof of a specific recovery time or coverage condition. Such estimates require operational evidence, for example tests, recovery records, or contract documentation.
Claims systems should associate each notice with an event identifier and separate direct technical effects from contingent effects. They should record the maturity and currency of each loss estimate, as well as original and revised geographic assignments. Event grouping can be preliminary while forensic work is in progress. Any later change to group membership must propagate through accounting and reinsurance reconciliation with a transparent audit history.
Insurer-level aggregate loss models should introduce common-cause event scenarios and propagation scenarios independently where practical, then consider interactions. Sensitivity to assumptions about undocumented dependencies and the correlation of recovery times belongs alongside sensitivity to severity parameters. Boundary assumptions for a catastrophe should be recorded before viewing portfolio outcomes, reducing the risk of choosing definitions to obtain preferred recoveries or risk measurements.
The methodological investigation by Zeller and Scherer (2024) into cyber accumulation risk considers the relation between underwriting information, claim-generating mechanisms and actuarial models for a shared vulnerability. Its results pertain to their modeled structures; they do not establish a universal bias for every insurer. Applied to geographic attribution, their approach suggests testing whether diversification represented in exposure classifications corresponds to diversification in the underlying mechanism. An insurer could use alternative common-vulnerability scenarios on the same fixed exposure inventory and assess the change in the conditional loss distribution. The appropriate benchmark is the difference in modeled exposure and claim pathways, rather than the number of countries alone.
Research priorities and validation criteria
The priority for empirical research is a matched exposure-and-claims dataset with consistent event identifiers, geographies by type, sector definitions, policy wording metadata, time stamps and valuation stages. Privacy-preserving linkage of incident records and insurer data could improve the assessment of geographic and sectoral loss distributions while protecting sensitive business information. Reporting harmonization should explicitly document differences in scope instead of compressing them into an apparently uniform international series.
A second priority is empirical measurement of dependency substitution and service recovery. Identical technical eligibility need not produce equal downtime. Recovery processes are shaped by shared resources, geographic access, and business functions. Research should compare the losses of directly impaired units and their verified dependent units, with attention to possible policy-selection and reporting bias.
A third priority is transparent testing of event grouping alternatives against actual policy and reinsurance provisions. Interdisciplinary work should join legal interpretation, forensic reconstruction, business interruption accounting and actuarial modeling without assigning one discipline's event definition to all the others. A model used for underwriting requires appropriate out-of-sample assessment, version control, and an explicit inventory of limitations. Synthetic tests establish arithmetic consistency, not empirical predictive performance.
The present investigation contributes a comparative interpretation and a worked analysis of attribution sensitivity. It introduces no named theory or new actuarial method. It identifies an important unresolved empirical problem: public event descriptions generally do not supply the contract-level data needed to reconstruct international insured loss distributions. This problem persists despite improvements in security reporting, and provides a precise direction for research.
Conclusion
Cyber catastrophe event footprints must connect the physical or logical manifestation of an event with the economic activities impaired, the duration of their impairment, the insured interests and the contractual mechanisms that transform loss into indemnity. Geography is integral to each connection. Its meaning depends on the object located and the purpose for which the information is used.
The comparison of WannaCry, NotPetya and CrowdStrike establishes the need to preserve distinct initiating mechanisms and observation units. The English health-sector audit supports a documented lower-bound organizational footprint. The NotPetya documentation establishes cross-sector consequences and illustrates why a firm-level financial impact cannot be assumed insured. The CrowdStrike account establishes a common technical exposure window whose duration is not the duration of each customer's loss.
Reporting mandates select different evidence. The US materiality framework, EU ICT incident classification and threat landscape reporting address different populations and clocks. The analytical consequences are explicit: reporting counts require defined denominators, and a missing observation is an unresolved state rather than an observed zero.
The synthetic portfolio demonstrates that geographic attribution changes concentration statistics, waiting periods change modeled gross covered losses, and occurrence grouping changes modeled reinsurance recovery even with fixed gross losses. These findings are ordinary deductions from specified assumptions. They establish the usefulness of preserving typed geographic information, time-indexed manifestations, and contract hierarchy. Empirical estimation of actual aggregate insured losses and their annual probabilities requires richer exposure, event and claims data. Selected legal provisions, two published specimen forms and the principal incident documents provide the documentary basis for the analysis. Limits remain on geographic claims data, issued contract wording, and the 2026 threat report, for which only the observation-window description is used.
References
National Audit Office. (2017, October 27). Investigation: WannaCry cyber attack and the NHS. HC 414, Session 2017–2019. Published findings. Source.
Foreign & Commonwealth Office and National Cyber Security Centre. (2018, February 15). Foreign Office Minister condemns Russia for NotPetya attacks. Source.
U.S. Department of Justice, U.S. Attorney's Office, Western District of Pennsylvania. (2020, October 19). Six Russian GRU officers charged in connection with worldwide deployment of destructive malware and other disruptive actions in cyberspace. Source.
FedEx Corporation. (2017). Quarterly report on Form 10-Q for the quarter ended August 31, 2017. Item 2, MD&A. U.S. Securities and Exchange Commission EDGAR. Source.
Merck & Co., Inc. v. ACE American Insurance Co. (2023, May 1). Superior Court of New Jersey, Appellate Division, 475 N.J. Super. 420, docket A-1879-21. Published opinion, reproduced by Justia. Source.
CrowdStrike. (2024, July 19). Technical details: Falcon update for Windows hosts. Source.
Weston, D. (2024, July 20). Helping our customers through the CrowdStrike outage. Microsoft Official Blog. Source.
Awiszus, K., Knispel, T., Penner, I., Svindland, G., Voß, A., & Weber, S. (2023). Modeling and pricing cyber insurance. European Actuarial Journal. DOI: 10.1007/s13385-023-00341-9. Source.
U.S. Securities and Exchange Commission. (2023, August 30). Cybersecurity risk management, strategy, governance, and incident disclosure: A small entity compliance guide. Source.
European Parliament and Council. (2022, December 14). Regulation (EU) 2022/2554 on digital operational resilience for the financial sector. Official Journal L 333, 27 December 2022, 1–79. Article 18. Source.
European Commission. (2024, March 13). Commission Delegated Regulation (EU) 2024/1772 supplementing Regulation (EU) 2022/2554 with regulatory technical standards on incident classification, materiality thresholds and details of reports. Official Journal L, 25 June 2024. Source.
European Parliament and Council. (2022, December 14). Directive (EU) 2022/2555 (NIS 2). Official Journal L 333, 27 December 2022, 80–152. Article 23. Source.
Pascoe, C., Quinn, S., & Scarfone, K. (2024, February 26). The NIST Cybersecurity Framework (CSF) 2.0. NIST CSWP 29. National Institute of Standards and Technology. DOI: 10.6028/NIST.CSWP.29. Source.
European Insurance and Occupational Pensions Authority. (2026, April). Insurance risk dashboard. EIOPA-BoS-26/170. Source.
European Union Agency for Cybersecurity. (2025, October). ENISA Threat Landscape 2025. Version 1.2. ENISA. Source.
European Union Agency for Cybersecurity. (2026). ENISA Threat Landscape 2026. Publication landing page. Source.
AXIS. (2020). AXIS Cyber Insurance Policy. AXIS 1012561 0120, specimen. Document.
BOXX Insurance. (2024, June 24). Cyber, Data and Privacy Insurance Policy: Cyberboxx Canada 5.0. BXC-CDPW-20240624, specimen. Document.
Zeller, G., & Scherer, M. (2024). Is accumulation risk in cyber methodically underestimated? European Actuarial Journal, 14, 711–748. DOI: 10.1007/s13385-024-00381-9. Document.
CrowdStrike. (2024, August). Executive summary: Root cause analysis, Channel File 291. Document.
Aon. (2024, July). CrowdStrike and Windows event briefing: Implications and initial findings for cyber reinsurers. Document.