RESEARCH DOSSIER
Cyber Catastrophe Probable Maximum Loss and Capital Modeling
Abstract. Cyber catastrophe probable maximum loss (PML) is a conditional extreme-loss measure that links digital dependencies, event mechanisms, insurance coverage and portfolio aggregation. This monograph investigates the scientific meaning of PML, occurrence and aggregate exceedance distributions, extreme-tail uncertainty, shared-service and propagation scenarios, contractual loss transformation, reinsurance and insurer capital. Primary supervisory documents, documented operational failures, a policy specimen, industry scenarios and peer-reviewed actuarial literature establish the evidentiary basis. Reproducible synthetic calculations demonstrate how identical marginal losses can generate different portfolio tails, how probability shifts alter loss quantiles, and how recovery uncertainty affects Expected Shortfall without changing Value at Risk. Comparative examination distinguishes European Union, United States, United Kingdom and international group-level prudential contexts. The original analytical synthesis identifies the necessary connections between event evidence, financial terms and balance-sheet effects, without presenting a proprietary calibration or claiming a new scientific theory.
Keywords: Cyber insurance; catastrophe risk; probable maximum loss; PML; aggregate exceedance probability; occurrence exceedance probability; Value at Risk; Expected Shortfall; systemic risk; common shock; digital dependency; business interruption; model uncertainty; reinsurance; regulatory capital; solvency; liquidity; stress testing.
Introduction
Cyber catastrophe modeling is the estimation of the financial consequences of severe, correlated digital events for a specified insurance portfolio over a specified horizon. It is a distinct object of insurance science because the same technical incident can generate a multitude of contractually different claims and a single event can also impair the insurer's own operations. The magnitude of an incident, the probability of an insured loss and the amount of capital needed to withstand that loss are three different quantities. A defensible analysis connects them through an explicit chain of evidence, assumptions and financial transformations.
The problem is particularly consequential when cyber exposures are simultaneously heterogeneous and dependent. Two insured businesses with different security budgets, revenue models and locations may use the same identity platform, endpoint management product, managed service provider or payment infrastructure. Conversely, businesses registered in the same geographic area may have little digital infrastructure in common. Conventional location-based accumulation measures therefore capture only part of the relevant concentration. Digital dependencies add a second dimension to the established actuarial problem of loss aggregation.
This study addresses three research questions. First, what does probable maximum loss (PML) mean when the underlying distribution is not directly observable and the portfolio itself is changing? Second, through which technical and contractual pathways do severe scenarios become insured claims, and how can uncertainty in these pathways be propagated? Third, how should modeled gross and net losses enter insurer capital assessment under different regulatory and institutional arrangements?
The research is an international analytical synthesis of original supervisory documents, actual incident records, a contract specimen, industry scenario research and peer-reviewed actuarial literature. The quantitative examples are constructed by the author for transparent demonstration; they are neither fitted estimates nor predictions for an identified insurer. The legal discussion is jurisdiction-specific. The article distinguishes current contractual coverage from hypothetical coverage and distinguishes historical incidents from forward-looking scenario assumptions. The parent Science is Cyber Insurance. The dossier's scientific scope is tail loss and capital modeling, including the financial transformation necessary to obtain net loss; detailed exhaustion mechanics and treaty design belong to separate studies.
The selection of sources follows relevance and verification rather than a claim of exhaustive systematic review. It includes work that explicitly differentiates common factors from transmission, public evidence describing actual operational disruptions, official insurance capital rules and a publicly accessible policy form. Public data do not provide the confidential portfolio, wordings, exposure attributes and claim maturities needed to produce a calibrated commercial PML. The absence of those data is a boundary of inference, not a numerical assumption that can be silently filled in. All examples use declared units and distributions, and the resulting conclusions concern the relationships among model components rather than a new theory of cyber risk.
Conceptual and Historical Foundations
From incident severity to insurance loss
A cyber incident is an occurrence that compromises, impairs or threatens the operation or security of digital systems. A catastrophe in insurance analysis is a severe accumulation of covered consequences across an event or time interval relative to a portfolio's financial capacity. Both definitions require specification of the unit under observation. For a security engineer the relevant unit may be a host; for a business it may be an interrupted service; for an insurer it may be one insured legal entity and its coverages; for a reinsurance contract it may be an aggregation of claims satisfying an event definition. An incident may qualify as a catastrophe under one scale and not another.
Let E denote a technical event and X_i(E) the ultimate covered loss, after primary contract terms, of policy i. For a portfolio with n policies, the event gross insured loss is G(E) = sum over i of X_i(E). The word gross here refers to gross of reinsurance, not gross economic damage or ground-up loss. Annual aggregate loss G is the sum of event-related claims and other covered annual losses on the modeled basis. A model that silently excludes attritional losses cannot subsequently interpret its event-only result as total annual underwriting loss. Similarly, expected ultimate losses cannot be compared directly with cash payments during one quarter without a development and settlement bridge.
The actuarial literature offers complementary treatments of dependency. Awiszus et al. (2023) examine individual, common-factor and systemic risks through a broad modeling and pricing analysis. The distinction matters: co-movement generated by a common external cause is not the same mechanism as a failure transmitted through connected organizations. These mechanisms can coexist. A common supplier introduces exposure to a shared shock; a downstream business can transmit the consequences to a second business through the services it provides. The model must specify whether its dependency graph represents common suppliers, propagation paths or both.
Zeller and Scherer (2022) develop a marked-point-process treatment that connects heterogeneous insurance exposures with cyber attack and failure events. Their approach illustrates the importance of event and exposure covariates in a setting where the observed count and severity of claims can depend on the characteristics of the organizations insured. A completely homogeneous claim-count process is useful as a diagnostic baseline but is inadequate as an explanation of all cyber catastrophe loss.
The institutional development of accumulation analysis
Cloud and software supply chains made accumulation increasingly salient to the insurance market. The 2018 Cloud Down study by Lloyd’s and AIR Worldwide investigated provider dependencies and their modeled impact across US businesses. It is an example of industry scenario construction rather than evidence that its projected losses actually occurred. Its role in this research is methodological: identification of insureds sharing a provider is a necessary exposure-management task before a scenario can be assigned a financial magnitude.
The 2023 IAIS cyber report addressed underwriting and operational cyber exposures in the context of global insurance market monitoring. The available monitoring data and reporting of non-affirmative exposures constrain the conclusions. A survey finding about a group of reporting insurers represents those respondents and reporting conventions. It does not establish the complete exposure of the world insurance sector.
Market growth reinforces the need to maintain dated exposure snapshots. The NAIC 2026 market report records US cybersecurity insurance premium totals, including alien surplus lines, of approximately $9.464 billion for 2025 and $9.144 billion for 2024. Reported domestic-company premiums were approximately $7.459 billion and $7.083 billion respectively. On the rounded amounts these imply roughly 3.5% and 5.3% year-on-year growth. Reporting populations and classifications matter. Premiums reflect prices, insured limits, exposure composition and market participation; they are not counts of attacks or estimates of return periods. These US data do not define the worldwide insurance market.
The historical development of catastrophe modeling therefore has two intertwined trajectories: improvement of the technical description of interdependence and improvement of the financial description of portfolios. A more accurate map of digital dependencies without accurate insured values and contract terms can increase confidence in the wrong quantity. The institutional response must make both information sets usable in the same loss model.
Distinguishing three financial objects
Ground-up economic damage is the reduction in economic resources, income or wealth before the insurance contract determines indemnity. Gross insured loss measures the financial obligations within primary coverage before reinsurance recoveries. Net insured loss is the amount retained by the insurer on the specified reinsurance basis. All three may differ in timing, measurement and uncertainty.
Insurer capital is a further object. It relates to the ability to absorb adverse changes in the financial position over a specified horizon and within the applicable balance-sheet, risk and supervisory framework. An insured-event severity quantile can be a major input; it is not equivalent to the distribution of changes in available capital. A single model should maintain these distinctions throughout its exposure data, scenario engine, reporting and conclusions.
PML Uncertainty
A quantile, not a physical upper bound
The term PML is used in insurance practice for different measures of severe loss. A scientific report therefore defines the estimand before reporting a number. In this study, unless specifically qualified as a scenario PML, the PML at confidence level q is the lower quantile VaR_q(L) = inf{x : F_L(x) >= q}, where L is the stated loss variable and F_L is its cumulative distribution function. Losses are expressed as positive amounts. The chosen tail probability 1 − q is an exceedance probability under the stated model.
An annual aggregate exceedance probability (AEP) curve describes the distribution of the sum of annual losses. An occurrence exceedance probability (OEP) curve describes the distribution of the largest qualifying occurrence within the year. An OEP-based 200-year loss and an AEP-based 200-year loss therefore answer different questions. If the annual damage variable and event scope are fixed, a largest nonnegative event loss is bounded above by the corresponding annual sum, but annual sums may also contain losses outside the event catalog. This relationship should be checked in implementation rather than presumed for differently defined outputs.
The label 200-year loss conventionally denotes an annual exceedance probability of 0.005. It does not predict that an event occurs once every 200 calendar years. Under independent years with constant event probability 0.005, the probability of at least one exceedance in 200 years is 1 − 0.995^200, approximately 63.3%. Nonstationary hazards, changing coverage, dependence across years and discontinuities in the loss distribution limit the literal interpretation of this convention.
A scenario PML is a loss estimate conditional on a specified severe event narrative and exposure assumptions. It can be valuable even when no defensible annual probability is assigned. To describe it as a quantile, however, requires a calibrated probability distribution or a justified mapping between scenarios and probabilities. A list of extreme scenarios ranked by loss cannot determine a return period merely by the order of their magnitudes.
Expected Shortfall and the discrete-tail problem
Value at Risk identifies a threshold but is insensitive to the distribution of losses strictly beyond that threshold. Expected Shortfall (ES) at level q is the average of quantiles from q to 1, divided by 1 − q. It is defined for an integrable loss L:
ES_q(L) = (1 / (1 − q)) integral from q to 1 of VaR_u(L) du.
For a finite discrete distribution, write v = VaR_q(L). Then ES_q(L) = v + E[(L − v)+] / (1 − q), where a+ = max(a, 0). To see this, the tail interval consists of the probability masses strictly above v and, when required, a fraction of the mass at v. Summing these masses gives the stated expression. It differs from an unconditional average over all observations with losses greater than or equal to v when substantial probability is concentrated at the quantile itself. The applicability of conditional value-at-risk to general distributions is the subject of Rockafellar and Uryasev (2002); the finite-distribution calculation here is derived independently.
For a finite model with equally weighted annual simulation outcomes, ES can be calculated by ranking losses and averaging the upper tail with fractional weight at the boundary. The empirical quantile should use the same generalized inverse convention as the ES calculation. Interpolation routines that produce an intermediate loss not realized by the model may be appropriate for some statistical purposes, but their convention must be disclosed, particularly when the distribution includes discrete scenarios or contractual limits.
The sources of epistemic uncertainty
Process uncertainty concerns future events and outcomes even under known parameters. Parameter uncertainty arises from incomplete estimation of the selected model's parameters. Structural uncertainty concerns the selection and validity of mechanisms, probability distributions, event descriptions, and dependencies. Exposure uncertainty concerns missing policy data, unknown suppliers, ambiguous identifiers and changing sums insured. Contract uncertainty concerns policy interpretation, causation, occurrence aggregation, and future adjudication or settlement. These are different categories with different possible remedies.
A stochastic model with estimated event intensities can simulate process variation precisely while leaving structural uncertainty essentially unchanged. Increasing the number of Monte Carlo trials reduces simulation noise conditional on the assumed model. It does not make omitted attack pathways appear. Increasing the number of insured observations can improve estimation of individual claim frequencies and simultaneously leave the frequency of a global shared-provider outage uncertain. Zeller and Scherer (2024) specifically investigate methodological underestimation of accumulation in cyber. Their analysis supports treating the availability and structure of exposure and claims information as part of the accumulation problem.
A useful decomposition of future loss uncertainty starts with a model index M and parameters theta. Conditional distributions F(L | theta, M) can be averaged using explicitly documented parameter uncertainty or model weights. A predictive mixture is obtained by integrating conditional loss distributions, not by averaging the quantiles of those distributions. In general VaR_q of a mixture does not equal the mixture-weighted average of VaR_q. The corresponding model set can instead be reported as a range of plausible conditional PML estimates, with no assertion that the endpoints represent statistical confidence limits unless a valid confidence procedure establishes that interpretation.
The most important sources of error in an extreme digital catastrophe may not admit a well-supported probability law. A deterministic stress reflecting a plausible but poorly quantified mechanism should remain a stress. The distinction prevents a subjective model weight from being mistaken for a historically observed frequency.
Extreme-value modeling and sparse data
Extreme-value methods can describe a tail above a suitably selected threshold. A generalized Pareto model, for example, uses distribution parameters scale beta > 0 and shape xi with positive support condition 1 + xi y / beta > 0. When xi is zero, the limiting distribution is exponential; positive xi produces a heavy tail. For a positive shape parameter, the theoretical model has an unbounded right tail and its mean is finite only for xi < 1. If fitted to ground-up costs, the resulting severity model must subsequently be transformed using the applicable insured limits.
Single-claim severity fitting is not equivalent to fitting an annual catastrophe loss distribution. Large single claims can arise from individual incidents; moderate claims that occur together can produce a much larger portfolio loss. Event clustering and uncertain event attribution affect the definition of the sample used for tail inference. For event loss extrapolation, the investigator must specify which events were recognized, how simultaneous claims were linked, how censored observations and deductibles were treated, and whether the portfolio at risk changed.
A goodness-of-fit measure on the center of a loss distribution provides little direct evidence for a 0.5% tail. Threshold sensitivity, stability of estimated parameters, censoring analysis, event-level validation and comparison against plausible scenario mechanisms are necessary to interpret extrapolated results. A severe empirical event also cannot, by itself, support a stable estimate of the frequency of all similar or more severe events.
Reportable PML specifications
Every PML result should carry its loss basis, occurrence or aggregate definition, confidence level or scenario identifier, model version, exposure valuation date, currency, valuation basis, inclusion of non-affirmative exposure, contract scope, recovery assumption and uncertainty statement. A loss number without that metadata is not a statistically reproducible result. A model comparison is meaningful only after these features are aligned or their differences are explicitly incorporated into the analysis.
Tail Scenarios
Exposure architecture and common-cause shocks
The first stage of a cyber catastrophe model is a dated exposure inventory. Policy-level records should identify the insured entity, coverage period, limit structure, attachment conditions, industry, operating countries, revenue exposure and relevant digital dependencies insofar as these are available and lawfully collected. Common dependencies may reside at several architectural layers: infrastructure, applications, identity, endpoint software, operational technology, managed services, payment services and network connectivity.
The same organization may use several suppliers, each of which is nested within an underlying infrastructure chain. Identifiers must be normalized without conflating a provider's corporate identity with a specific product or service. A shared corporate vendor name is not enough to establish common technical failure. Provider region, redundancy, tenancy, product version, rollout configuration, backup procedures and ability to substitute a service can make exposure conditionally different.
A simple event model assigns a set of shock types j, event intensities lambda_j and event marks V_j, describing attributes such as duration, effective reach, product version and restoration behavior. Given an event, a vulnerability function maps the event attributes and each insured's exposure to the probability and magnitude of disruption. For propagation, a network model may instead represent directed dependencies and conditional failure transmission. Calibration of transmission probabilities requires evidence; arbitrary propagation along every edge is a stress assumption, not an estimated fact.
Potential severe scenarios include faulty software distribution, widespread malicious exploitation, cloud-service interruption, identity-service failure, destructive malware and supply chain compromise. A scenario library should include accidental failures and malicious operations when policy scope makes them relevant. The 2018 cloud dependency study is a public example of investigating a shared service exposure. Its scenario numbers should not be transferred to a different current portfolio without remapping exposure and vulnerability.
Evidence from a widespread operational disruption
CrowdStrike's July 19 technical account describes a faulty Falcon content update distributed on July 19, 2024, from 04:09 to 05:27 UTC, affecting qualifying online Windows hosts using the relevant sensor version (7.11 and later). This provides a bounded technical initiating mechanism. The distribution interval is not the duration of business interruption. Restoration at different organizations required different interventions and proceeded on their own timelines.
Microsoft's contemporaneous estimate, reported by Weston (2024), was 8.5 million affected Windows devices, amounting to less than one percent of Windows devices. The denominator is devices, not businesses, insured policies or covered claims. The congressional hearing record documents vendor responses about rollout and testing. These accounts support analysis of a shared software mechanism and potential engineering controls but do not yield an ultimate insurance claims distribution.
An event model informed by this experience can specify online eligible endpoints, distribution timing, update propagation, organizational restoration resources, and the relationship between unavailable endpoints and critical services. A business that loses a small fraction of devices on an essential operational path may experience substantial interruption, while a different business may continue despite widespread workstation disruption. The mapping from endpoint reach to service downtime is consequently nonlinear and organization-specific.
The episode also illustrates the difference between a common initiating event and multiple contractually defined occurrences. A common defective update may be treated differently in different policy or treaty wordings. The causal event map should retain the technical origin, affected entities, time sequence and restoration information separately from the contractual aggregation decision. This permits alternative aggregation interpretations to be evaluated without rewriting incident facts.
Malicious scenarios and strategic dependence
Malicious cyber events pose additional questions of purposeful targeting and adaptation. An attacker can exploit a common weakness when the expected return is high, choose an important upstream service because it increases leverage, or shift technique when defensive controls become widespread. A model organized entirely around fixed independent hazards loses the strategic character of these mechanisms. An event frequency calibrated to a historical technique may change as technology, defensive operations and incentives evolve.
A defensible stress explores attacker reach and technical feasibility while keeping assumptions distinct from evidence. A catastrophic attack that interrupts a widely used service may require several simultaneous conditions: compromise of a particular access path, sufficient privileges, ineffective segmentation, widespread dependence and prolonged absence of substitutes. A scenario does not gain plausibility simply because its loss is large. Its mechanism should be technically coherent, and its loss drivers should be challengeable by reviewers with relevant security and operations expertise.
The 2023 Lloyd’s systemic risk scenario presented a modeled five-year global GDP loss measure of $3.5 trillion for major cyber disruption, expressed through a weighted severity construction. It is a macroeconomic scenario, not a forecast of insurance indemnities or an annual insurer PML. To translate macroeconomic disruption into insurance exposure would require identifying covered firms, perils, insurance market participation, limits, indemnity periods, causation and reinsurance. Applying a single insurance penetration percentage to cumulative GDP loss would bypass this chain.
From disruption to contract-defined claims
For each policy, ground-up business interruption can be represented as a time integral of the insured loss rate during an interruption, plus eligible extra expense, subject to the chosen contractual basis. Revenue loss is not automatically equal to insured business income loss, since saved expenses, alternative production and other adjustments depend on the policy. A duration model should distinguish technical recovery, recovery of service capacity and recovery of economic output.
Illustration: assume a policy indemnifies a loss rate of $20,000 per hour after a waiting period of 24 hours, treated for this example as a simple excluded initial interval. An interruption lasting 72 hours gives 48 indemnifiable hours and a business interruption amount of $960,000. If $100,000 of extra expense is separately eligible, the amount before a monetary retention is $1.06 million. With a $100,000 retention applying once to that combined loss and a $1 million limit, the payment is $960,000. Different waiting-period wording or coverage sequencing can materially change the result. These terms are hypothetical and intentionally explicit.
The AmTrustCyber specimen policy illustrates the importance of inspecting actual wordings. It contains separate own-business and dependent-business interruption coverage language, together with declarations and waiting-period provisions. It is a specimen, not the policy of an actual insured in this study. An insurer must map coverage-specific attachment, limits, exclusions and other conditions using the executed forms and applicable law. One policy can grant cover for unplanned system failure under specified conditions while another may require a different triggering event.
The model must also consider legal attribution, war-related exclusions and possible gaps in alignment between underlying and treaty coverage. The LMA 2023 clause bulletin identifies distinctions among model state-backed cyber attack clauses. The LMA 2024 treaty bulletin illustrates corresponding model wording for cyber treaty reinsurance. Neither bulletin establishes that all executed contracts incorporate those terms, or that all disputes have a uniform outcome. Scenario results should retain conditional coverage alternatives where conflicting interpretations or missing wording data could materially affect loss.
Non-affirmative cyber exposure may arise in insurance classes not explicitly labeled cyber where the operative policy terms potentially respond to a cyber-related event. EIOPA's 2022 supervisory statements address management of such exposures and clarity around systemic exclusions. Aggregation across affirmative and non-affirmative products must be performed at the actual contract-response level: counting every potentially exposed policy at full limit is a stress bound, whereas assuming all non-affirmative policies are excluded is a different and potentially equally misleading assumption.
Dependency as a tail multiplier
Consider 100 policies, each paying $1 million with annual marginal probability 0.01. In Model A, payments are independent. The count of payments is binomial with n = 100 and p = 0.01, with mean 1 and variance 0.99. In Model B, all policies pay when a shared shock occurs with probability 0.01. The count is 100 with probability 0.01 and zero otherwise. It also has mean 1, but variance 99.
Using the binomial cumulative distribution, the 99.5% loss quantile in Model A is $4 million. In Model B, it is $100 million. Expected loss equals $1 million in both. The calculation isolates the effect of dependence: identical marginal probabilities and severities give radically different portfolio tail outcomes. It is a teaching example, not an estimate of actual shared cyber dependency.
In a more realistic portfolio, let C be the common event state or common factor. The law of total variance gives Var(G) = E[Var(G | C)] + Var(E[G | C]). Even with conditional independence, the second term can dominate portfolio variation. Conditional dependence beyond the common factor, including cascading transmission, requires additional modeling. Pairwise correlations estimated from an ordinary period can fail to describe the joint distribution conditional on an extreme disruption.
Scenario design should therefore vary both the probability that a shared event occurs and the proportion of each connected business's loss that is covered. A robust book of individually well-controlled risks can remain highly concentrated on a provider; high dependency need not imply every connected business will fail or claim. Estimation requires both layers.
Regulatory Capital
The financial transformation from gross to net loss
Reinsurance reduces or reshapes the loss retained by the insurer when its contractual recovery conditions are satisfied. For an illustrative annual aggregate excess layer with attachment a and limit l, recovery is R(G) = min(l, max(G − a, 0)), and net loss N = G − R(G). A real occurrence treaty may require defining separate events; attachment, exhaustion, reinstatements, annual limits, exclusions, timing and other terms must then be applied in their actual sequence. The illustrative aggregate transformation avoids conflating those more complex structures.
Consider a synthetic annual aggregate portfolio loss with five states. Monetary figures are millions of US dollars and outcomes and probabilities are assumed, not empirical. The 99.5% VaR convention is the generalized inverse defined above. Reinsurance comprises a hypothetical single annual aggregate layer of 150 excess of 100, fully collectible and triggered in every eligible state. There is no premium, discounting or other cash flow in this illustration.
| Gross annual loss | Probability | Cumulative probability | Recovery | Net annual loss |
|---|---|---|---|---|
| 10 | 0.940 | 0.940 | 0 | 10 |
| 40 | 0.040 | 0.980 | 0 | 40 |
| 120 | 0.014 | 0.994 | 20 | 100 |
| 250 | 0.004 | 0.998 | 150 | 100 |
| 500 | 0.002 | 1.000 | 150 | 350 |
Expected gross loss is 14.68 and expected retained loss is 13.50; the difference 1.18 is expected recovery under the assumed layer. Gross VaR at 99.5% is 250 because 0.994 of probability lies below 250 and 0.998 at or below it. The most severe 0.5% contains all 0.002 probability of the 500 state and 0.003 probability at 250. Gross ES is (0.002 × 500 + 0.003 × 250) / 0.005 = 350. Net VaR is 100 and net ES is (0.002 × 350 + 0.003 × 100) / 0.005 = 200.
If expected loss has already been provided for, a narrowly defined loss-only unexpected-risk measure VaR minus mean gives 235.32 gross and 86.50 net. An ES-minus-mean measure gives 335.32 gross and 186.50 net. These figures do not represent an actual solvency capital requirement. Their purpose is to show how expected loss provisioning and a stated risk measure can change the interpretation of the same distribution.
A stress in which the most severe outcome receives no recovery because of a hypothesized exclusion or noncollectibility produces net outcomes 10, 40, 100, 100 and 500. Expected net loss becomes 13.80, net VaR remains 100, and net ES becomes 260. This is a clear illustration of why a single net VaR can fail to capture the financial effect of a recovery weakness beyond the quantile. The hypothesized exclusion and counterparty failure have different legal and financial mechanisms; their equal cash-loss treatment in this example does not imply that they are interchangeable in practice.
Economic and prudential capital
Economic capital in an internal model ordinarily refers to an amount of financial resources associated with a risk appetite and valuation basis. Capital required by a supervisory framework is a legal or regulatory measure determined according to that framework's applicable rules. The balance-sheet definition of available capital, the confidence level or risk measure, the time horizon and the treatment of mitigation differ across frameworks.
In the European Union, Solvency II Article 101 specifies that the Solvency Capital Requirement corresponds to the Value-at-Risk of basic own funds subject to a confidence level of 99.5% over a one-year period. It includes specified underwriting, market, credit and operational risk categories and recognizes risk mitigation subject to the associated risk treatment. A cyber underwriting catastrophe quantile is only one potential component of the modeled financial position. There is no general identity between the 99.5% quantile of gross annual cyber losses and the undertaking's Solvency Capital Requirement.
The distinction is especially important for events that simultaneously create policy claims, impair the insurer's own systems, affect investment values, alter the collectibility of reinsurance recoverables or generate adverse effects in other business lines. If F_0 is the initial own-funds position and F_1 its modeled end-of-horizon value, the modeled adverse change can be represented by F_0 − F_1 under a declared convention. Estimating F_1 requires the full financial and valuation assumptions relevant to that model. The sign and scope of each term should be reconciled with balance-sheet accounts, including expected losses already reflected in reserves and premiums.
The EIOPA 2023 cyber stress methodology explicitly addresses underwriting and insurer operational cyber risks. This supports designing scenarios that cover both channels while retaining their distinct accounting and causal treatment. An insurance company may have a material exposure through the customers it insures even if its own technical environment is resilient, or may face severe operational disruption without a comparable underwriting catastrophe.
The international supervisory reference is IAIS ICP 17, including its capital adequacy treatment at the relevant solo and group levels. The IAIS Insurance Capital Standard adopted in 2024 is directed at internationally active insurance groups. It should not be presented as a uniform national capital rule for all insurance undertakings or as a required standalone cyber PML method.
In the United States, the NAIC's risk-based capital overview explains the regulatory intervention role of RBC. Appropriate statutory treatment depends on the insurer type, applicable state requirements and reporting basis. RBC and Solvency II SCR ratios are not directly equivalent. Comparing nominal percentages across these systems without specifying numerator, denominator and intervention meaning creates a misleading measure of capital adequacy.
In the United Kingdom, the PRA 2022/23 annual report describes a historical insurance stress-testing exercise that included cyber and natural catastrophe scenarios for general insurers and Lloyd’s syndicates. Conclusions from a specified stress exercise concern the specified scenarios and participating institutions; they do not guarantee resilience under every possible cyber catastrophe. This cross-jurisdictional analysis identifies distinct instruments and purposes rather than assigning a uniform global confidence level.
Capital adequacy and liquidity after a catastrophe
Suppose an insurer has hypothetical eligible own funds of $400 million and required capital of $250 million, giving a ratio of 160%. An additional $100 million modeled net deterioration in own funds reduces the numerator to $300 million, while a re-estimation of required capital gives $275 million. The post-stress ratio is approximately 109.1%. A further $40 million operational deterioration, assumed not to overlap with the first loss and holding the revised denominator fixed, produces approximately 94.5%. These calculations are simple ratios, with no tax, valuation, regulatory eligibility or management-action adjustments.
The example makes clear why a stress report has to model changes in both numerator and denominator, rather than subtracting a catastrophe loss from a previously calculated regulatory capital number. It is also possible for a loss to have been partially reflected in provisions before occurrence recognition. Modeling should track incremental changes rather than deducting previously recognized expected amounts twice.
Cash-flow resilience introduces a second set of questions. Claims advances, defense expenses and emergency assistance may be due before reinsurance recoveries are due or received. A solvent insurer can face constrained liquid resources during that interval. Liquidity modeling should specify payment schedules, recovery timing, collateral, available liquid assets under stress and contingent funding constraints. A recovery counted in ultimate net loss is not automatically available cash at the needed date.
A severe shared event can affect multiple business lines and financial counterparties. If a common shock affects primary coverage, reinsurer credit and an insurer's own claims-processing capacity, independent simulations of these risk categories followed by unconditional aggregation may obscure adverse dependence. The model should use a common scenario state and then reconcile the valuation channels so that actual losses are neither omitted nor duplicated.
Risk appetite and decision use
A board-level PML framework should distinguish current capital resources, risk appetite thresholds, underwriting limits on dependencies, treaty protections and escalation points. Changes in PML may result from portfolio changes, revised contract interpretations, model updates or actual changes in the hazard. Reporting a single time series without a change attribution prevents decision makers from understanding whether underwriting action or model remediation is required.
Capital allocation to businesses or segments can use marginal risk contributions under a declared risk measure, but reinsurance, deductibles and policy limits introduce nonlinearity. A risk contribution computed before a treaty transformation cannot be assumed to allocate the retained portfolio risk consistently. The allocation question must state whether it concerns pricing, risk appetite, capital charges, or the distribution of group capital; these uses can call for distinct methods.
Evidence Validation and Sensitivity Analysis
Calibration and data lineage
The credibility of the model rests on a documented chain from source evidence to exposure attributes, event occurrence, vulnerability, contractual claims, recoveries and capital effects. Each variable requires a definition, unit, observation period, population, version and missing-data treatment. For an observed incident, the record should separate what was directly documented from reconstructed causal pathways, reporting estimates and assumptions made for modeling. The claims database should identify valuation date, paid and incurred amounts, reported limits, event grouping and the distinction between open and closed cases.
Public incident records frequently identify affected systems or organizations, but reporting windows vary and coverage information is often unavailable. Claims data may exclude non-insured firms and events below retentions. Industry statistics can be affected by changes in policy participation, reporting thresholds, company practices and development maturity. As illustrated by the NAIC premium figures, measured market growth cannot by itself identify catastrophe frequency. The exposure denominator needs its own evidence before event rates can be estimated.
Marked-process models such as the 2022 actuarial study provide a way to consider heterogeneous event types and covariates, but estimating an event intensity requires defining what is counted. In a shared shock, treating a single originating failure with 1,000 reported claims as 1,000 independent catastrophe events inflates event frequency and changes the dependence structure. Conversely, assigning every claim during a broad time interval to a single event risks merging unrelated losses. Event-reconstruction uncertainty should remain visible in the data.
Data acquisition may involve privacy, confidentiality and security constraints. Detailed dependency mappings can themselves be sensitive operational information. The model owner should define permissible data uses, access controls, retention, aggregation and validation procedures. In a scientific publication, proprietary insured identifiers and details that enable targeting should be avoided. Transparent assumptions and aggregate demonstrations can preserve reproducibility of reasoning without making an actual insurer's confidential exposure inventory public.
Numerical sensitivity and validation examples
The five-point gross distribution permits a transparent sensitivity test. Transfer probability 0.004 from the loss-10 state to the loss-500 state, holding all other states fixed. The probabilities become 0.936, 0.040, 0.014, 0.004 and 0.006. The cumulative probability at 250 becomes 0.994, and the new 99.5% VaR is 500. ES is also 500 because the entire upper 0.5% lies in the 500 state. Expected gross loss rises from 14.68 to 16.64, a change of 1.96.
The example demonstrates threshold instability when probability mass crosses the selected tail boundary. The expected loss increases about 13.35%, while the VaR doubles. A small absolute probability shift cannot always be treated as a small change in tail capital. This observation is a consequence of the constructed loss distribution, not a claim that cyber PML empirically doubles under a particular frequency revision.
The shared-shock example can be verified without Monte Carlo. For independent claims, P(K = 0) = 0.99^100 and P(K = k + 1) = P(K = k) × (100 − k)/(k + 1) × 0.01/0.99. Summation yields cumulative probabilities approximately 0.9816 at k = 3 and 0.9966 at k = 4. These bound the 0.995 quantile at four claims. For the fully dependent model, P(K = 0) = 0.99 and P(K = 100) = 0.01, so any quantile above 0.99 is 100. The expected count is one in each construction.
When the model is stochastic, validation should evaluate marginal frequencies, losses by coverage, dependency concentrations, event reach and conditional severity at each stage. End-to-end historical validation is valuable but limited for extremely rare events. With a probability of 0.005 per year, a 20-year sample of independent stationary annual observations has an expected 0.1 exceedance. The probability of observing none is 0.995^20, approximately 90.5%. Consequently, absence of an observed exceedance over that interval provides weak evidence about calibration of a 99.5% annual quantile. A one-year simulated return-period result cannot substitute for this missing historical evidence.
Simulation diagnostics should include repeated seeds, convergence as sample size increases, quantile uncertainty and sensitivity to extreme scenario sampling. If a simulation uses 100,000 independent annual trials, an expected 500 observations fall in the upper 0.5%, subject to sampling variation and ties. That count is an indication of numerical precision, not validation of the scenario probabilities. Importance sampling can increase effective information about severe paths, but probability weights must be carried through estimation of the original distribution. Selecting the most severe simulated outcomes after oversampling them without correction yields biased tail estimates.
A simple reproducibility record for the finite examples is the loss vector, the probability vector, the quantile convention, the recovery transformation and the sensitivity modifications. Independent calculations can sum probability, compute the mean, locate the quantile and evaluate ES using its excess-loss expression. This avoids reliance on software interpolation defaults. For a model used in production, validation also requires exact exposure snapshot hashes, dependency version records, model code or vendor version, run configuration, calendar convention, currency basis and reproducible financial terms.
Contract, model and operational challenges
Validation should challenge event modeling and financial modeling separately. For a technical scenario, test whether the affected product, version, service or dependency is consistent with the event mechanism. For business loss, test whether duration, partial recovery, substitution, revenue loss and extra expense can plausibly arise from the disruption. For insured loss, check declarations, coverage triggers, waiting periods, retentions, limits, exclusions, relevant law and possible disputes. For net loss, test treaty eligibility, ordering and recoverability. For capital, reconcile the resulting loss and other modeled changes to the balance sheet.
The United States actuarial standards ASOP No. 56 on modeling and ASOP No. 38 on catastrophe modeling are relevant professional references within their specified scope. They are not globally binding technical standards for every model user. Their existence and scope underscore that modeling is a professional activity with purposes, assumptions, reliance considerations and required judgment; the proposed validation record here is an analytical synthesis adapted to cyber catastrophe risk, not an assertion that every individual checklist item is textually mandated by those standards.
Independent challenge is particularly important for vendor models where proprietary event catalogs or vulnerability functions are unavailable to the insurer. The insurer can still examine input completeness, output definitions, model documentation, scenario relevance, sensitivity and reconciliation with claims experience. The inability to inspect all internal source code is a limitation of the verification performed. A vendor's model output should not acquire greater evidentiary status merely because an opaque calculation reports more decimal places.
A triangulation approach compares structural scenarios, statistical loss information and expert review. Statistical estimates provide information about observed distributions conditional on data and sample definition. Scenario analysis provides information about coherent but poorly observed mechanisms. Expert review can challenge feasibility and dependencies. Divergence across these methods is an investigation result rather than a defect to hide by averaging. The reasons for disagreement must be established before combining results.
Minimum uncertainty disclosure
An internal capital report should document material uncertainty in the event catalog, eligibility assumptions, missing suppliers, coverage interpretation, loss development, treaty response, model structure, and balance-sheet dynamics. Parameter ranges should have named evidentiary or judgmental bases. The report should include baseline and stressed estimates with a narrative of the drivers, and distinguish stochastic percentiles from scenario magnitudes.
This study's original illustrations are intentionally independent of a proprietary dataset. Their precision is arithmetic, not empirical. Neither the selected events nor the cited market information is used to infer an actual probability for an insurer's catastrophic loss. Direct PDF page-image access was incomplete for some official reports; the cited numerical and supervisory statements were restricted to the examined extracted text. The university repository abstract for the 2002 risk-measure paper supports only its general-distribution scope. The finite-distribution formula is derived explicitly in the text. These limitations constrain the claims and are recorded in the evidence ledger.
Integrated Scientific Synthesis and Research Agenda
Reconciling event probability and financial severity
The scientific synthesis is that a cyber PML is jointly determined by event mechanism, exposure dependency, contract response and financial aggregation. Improving each component independently is valuable, but the complete estimate can remain wrong if the interfaces between components are inconsistent. For example, a correctly modeled event reach can be applied to a stale book of insurance; an accurate business interruption severity can be mapped to the wrong waiting-period structure; a correctly estimated gross loss can be transformed through an inapplicable treaty; a correctly estimated retained loss can be inserted into the wrong capital accounting basis.
The shared-shock illustration shows that marginal loss estimates are insufficient for portfolio catastrophe risk. The five-state illustration shows that net VaR may stay unchanged when an extreme recovery gap materially increases ES. The capital ratio illustration shows that the consequences of a given net catastrophe loss depend on both own-funds deterioration and the post-stress level of required capital. These results are mathematically consistent applications of established concepts. They do not establish a new theoretical system, nor do they estimate the size of a real-world portfolio.
PML is consequently a measure with a declared scope, not a synonym for catastrophe size. Restricting a model to a known peril and a selected insured-loss population is scientifically acceptable when the result is identified accordingly. Describing such a model as the total capital impact of cyber risk is a different claim requiring additional evidence.
An integrated workflow for accountable model use
A practical workflow begins by deciding which financial quantity, horizon and extreme-tail convention the institution needs. It proceeds by inventorying risks and contracts, identifying shared digital dependencies, defining event families, estimating or stressing event likelihood, specifying conditional losses, applying financial terms and examining alternative contractual responses. Net and gross exceedance curves are generated on aligned exposure assumptions. Then operational, asset, liability, recovery and liquidity risks enter the relevant financial assessment.
The workflow contains a feedback relationship. Discovery of a new common provider changes the exposure inventory and scenario relevance. Changes to exclusions or policy conditions alter the mapping into claims. A new event can inform technical vulnerability assumptions without re-estimating its annual occurrence probability from a single observation. A financial stress that reveals a tail beyond the risk appetite may lead to underwriting restrictions, improved data acquisition or changes in risk financing. Each response should be evaluated against the same underlying quantity so that changes in reported PML can be attributed.
Scenario selection is a particularly important governance task. A library limited to historically observed events can omit emerging mechanisms. A library of dramatic possibilities without plausibility constraints can overstate what can be inferred. Model governance requires a documented distinction among a modeled probabilistic event, a severe-but-plausible stress, and a hypothetical reverse stress designed to test where institutional resilience fails. None of these exercises can be properly understood without its decision purpose.
Reverse stress testing starts from a failure condition, such as an inability to maintain a defined financial or liquidity threshold, and identifies combinations of losses and other effects that could produce it. The resulting scenario identifies a vulnerability boundary. It is not necessarily a forecast or a probability estimate. In cyber underwriting, reverse stress can reveal sensitivity to shared dependencies, contract mismatches, unusually prolonged restoration, or a delayed recovery even when none of these individually breaches a loss threshold.
Portfolio diversification and external capacity
Risk diversification in cyber insurance depends on whether different exposures are exposed to the same severe mechanisms. A varied set of industries can still share cloud, endpoint and identity services. Adding more underwritten risks can reduce sampling variance at the individual level while increasing absolute exposure to a common event. Portfolio monitoring should accordingly report exposure concentration by mechanism, dependency and financial response instead of relying only on industry or country diversification.
Access to external risk-bearing capacity is influenced by the ability to define and quantify the risk transferred, capacity limits, recovery terms and the dependence of counterparties on the same systemic event. A nominal treaty limit measures available protection under contractual conditions, not assured economic recovery under every scenario. A material mismatch between primary and treaty scope or a delay in cash recovery should be evaluated in net tail analysis, as the example demonstrates. The detailed design of layers and contractual event aggregation remains a distinct research subject.
Comparative supervision provides complementary lenses. The EU one-year own-funds VaR target, US RBC intervention framework, IAIS international group standards and historical UK stress testing are different institutional arrangements. A scientific comparison uses the same hypothetical underlying portfolio and explicitly maps each distinct measure where information is sufficient. It does not assume that superficially similar capital percentages establish equivalent protection.
Research priorities
The empirical research agenda includes more reliable event identifiers in claims datasets, clearer reporting of loss bases and development, exposure-normalized data on shared suppliers and product versions, and methods for evaluating non-affirmative exposure. Studies should examine the relationship between technical reach and economic interruption conditional on redundancy and restoration resources. Data on coverage and recoveries, when lawfully available in aggregate form, would help distinguish an economic catastrophe from an insured one.
Statistical research should investigate how event rates, severity and dependency change when portfolios and technology evolve, how model uncertainty changes estimates of rare-event loss, and how discrete financial terms affect tail metrics. Comparative modeling studies should predefine the portfolio snapshot and loss basis before comparing outputs. Disagreement caused by different model assumptions is scientifically different from disagreement caused by different contract inputs. A structured decomposition could make comparative studies more interpretable without dictating a particular commercial model.
Regulatory research should investigate the interaction of underwriting and own-operation disruption, including the accounting, recovery and liquidity channels by which a shared event affects an insurance group. Publicly available solvency disclosures may be insufficient to reconstruct internal models; future studies need to distinguish what can be audited from public evidence and what requires cooperation with insurers or supervisors.
The present research establishes the analytical requirements for interpreting PML and using it in capital modeling. A specific calibrated PML would require a separately authorized empirical investigation with a legally accessible exposure set, actual contract data, observed event and loss information, and documented model validation. Numerical examples illustrate the consequences of assumptions and are reproducible from the values provided. The scientific contribution is the integrated interpretation of technically generated loss, quantified extreme dependence and financial resilience, grounded in established actuarial and prudential frameworks.
Conclusion
Cyber catastrophe probable maximum loss is a defined extreme-loss measure conditional on a portfolio, risk mechanism, time horizon, probability or scenario specification, loss basis, contract interpretation and model version. A scenario severity and an annual quantile are different objects. Gross insured loss, net retained loss and the change in an insurer's financial position are different financial quantities.
The research supports several connected conclusions. Shared dependency can radically alter portfolio tails with no change in marginal expected losses; therefore marginal frequencies and severities are inadequate for catastrophe modeling in isolation. An explicit event and exposure architecture is required to represent common-cause disruption and propagation. Historical incident records support the mechanism and observable reach of specific events, but ordinarily do not provide a complete insured-loss distribution. Economic scenario totals cannot be converted into insurer PML without contractual and exposure evidence.
Contractual response determines which interruptions become covered claims. Reinsurance transforms those claims into retained loss subject to contractual and financial conditions. Expected Shortfall complements quantile-based measures by revealing severe outcomes beyond a fixed threshold. The hypothetical recovery stress in this monograph demonstrates how a significant extreme-tail difference can leave net VaR unchanged. These examples establish methodological relationships, not empirical estimates or new named theories.
Capital modeling extends the loss distribution into the full relevant balance-sheet framework. Regulatory requirements differ across jurisdictions and institutional scope; individual cyber PML is not a universal statutory capital requirement. Scenario analysis should include potential interaction among underwriting losses, operational disruption, reinsurance, other financial effects and liquidity, while reconciling overlaps and valuation. The model's uncertainty must remain visible in the reported result.
A credible cyber catastrophe assessment is therefore a reproducible, auditable account of how an initiating event becomes a covered financial obligation and how that obligation affects the institution's resilience. The practical scientific priority is to improve the evidence linking shared digital dependency, observed disruption, actual contract response and capital consequences. This is a necessary foundation for meaningful estimation of rare insured losses in an evolving digital environment.
References
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Source
Actuarial Standards Board. (2021). Actuarial Standard of Practice No. 38: Catastrophe modeling (for all practice areas). Revised edition. Effective December 1, 2021. Source
Awiszus, K., Knispel, T., Penner, I., Svindland, G., Voß, A., & Weber, S. (2023). Modeling and pricing cyber insurance: Idiosyncratic, systematic, and systemic risks. European Actuarial Journal, 13, 1–53. Source
CrowdStrike. (2024, July 19). Technical details: Falcon update for Windows hosts. Source
European Insurance and Occupational Pensions Authority. (2022, September 22). EIOPA publishes supervisory statements on exclusions related to systemic events and the management of non-affirmative cyber exposures. Source
European Insurance and Occupational Pensions Authority. (2023, July 11). Methodological principles of insurance stress testing: Cyber component. Source
European Parliament & Council of the European Union. (2009). Directive 2009/138/EC (Solvency II), Article 101, Calculation of the Solvency Capital Requirement. EIOPA single rulebook. Source
International Association of Insurance Supervisors. (2023). Global Insurance Market Report: Special topic edition on cyber. Source
International Association of Insurance Supervisors. (2024). Insurance Capital Standard (ICS). Official standard-setting overview. Source
International Association of Insurance Supervisors. (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups, ICP 17: Capital adequacy. Online consolidated tool. Source
Lloyd’s & AIR Worldwide. (2018). Cloud down: Impacts on the US economy. Source
Lloyd’s Market Association. (2023). Bulletin LMA23-002-PD: State-backed cyber attack exclusion clauses. Source
Lloyd’s Market Association. (2024, March 1). Model state-backed cyber war exclusion clauses for cyber treaty reinsurance. Bulletin LMA24-011-CM. Source
Lloyd’s. (2023, October 18). Lloyd’s systemic risk scenario reveals global economy exposed to $3.5trn from major cyber attack. Source
National Association of Insurance Commissioners. (2026). Report on the cybersecurity insurance market. Source
National Association of Insurance Commissioners. (n.d.). Risk-based capital. Insurance Topics. Accessed November 18, 2026. Source
Prudential Regulation Authority. (2023). Prudential Regulation Authority annual report 2022/23. Bank of England. Source
Rockafellar, R. T., & Uryasev, S. (2002). Conditional value-at-risk for general loss distributions. Journal of Banking & Finance, 26(7), 1443–1471. DOI: 10.1016/S0378-4266(02)00271-6. Source
Technology Insurance Company, Inc. (2021). AmTrustCyber sample policy: Declarations CYDEC 0321 and policy form CY990001 0321. Distributed by Victor Insurance. Source
U.S. House of Representatives, Committee on Homeland Security. (2024). An outage strikes: Assessing the global impact of CrowdStrike’s faulty software update. Hearing, 118th Congress. Source
Weston, D. (2024, July 20). Helping our customers through the CrowdStrike outage. Microsoft. Source
Zeller, G., & Scherer, M. (2022). A comprehensive model for cyber risk based on marked point processes and its application to insurance. European Actuarial Journal, 12, 33–85. First published online 2021. Source
Zeller, G., & Scherer, M. (2024). Is accumulation risk in cyber methodically underestimated? European Actuarial Journal, 14, 711–748. Source