RESEARCH DOSSIER
Cyber Insurance Reverse Stress Testing and Solvency Buffers
Abstract. Cyber insurance reverse stress testing identifies the combinations of insured loss, asset impairment, operational disruption, reinsurance shortfall, and payment timing that exhaust the resources necessary for an insurer to remain viable. The relevant boundaries include regulatory capital requirements, internal capital targets, continuity of payment, and continued access to effective risk transfer. This research develops a comparative scientific treatment of these boundaries, integrating insurance economics, prudential regulation, actuarial loss accumulation, contractual reinsurance mechanics, and liquidity management. It examines supervisory texts from the IAIS, European Union, United Kingdom, Canada, and United States, research on dependence and alternative risk transfer, and documented evidence of concentrated technology dependencies. Explicitly synthetic numerical examples isolate the effects of treaty attachment, exhaustion, recoverability, and settlement delay. The principal analytical finding is that a portfolio can retain positive regulatory capital headroom while losing the capacity to discharge claims as they fall due. Conversely, short-term liquidity financing can preserve payment continuity while leaving a permanent capital impairment unresolved. A scientifically defensible reverse stress exercise therefore reconstructs a sequence of events and contractual cash flows across legal entities and time horizons, tests the feasibility of management responses, and treats uncertainty about coverage, dependence, and collection as part of the exercise. The investigation establishes a practical basis for interpreting solvency buffers without proposing a new named theory or claiming an empirically calibrated probability of insurer failure.
Keywords: Cyber insurance; reverse stress testing; solvency; liquidity risk; capital adequacy; reinsurance recoverability; systemic cyber accumulation; operational resilience; insurance regulation; risk tolerance; ORSA; contingency funding; model uncertainty.
Introduction
An insurance undertaking makes promises concerning uncertain future payments. Its capacity to honor those promises depends on the balance between economic resources, contractual obligations, and the time at which resources can be mobilized. Cyber insurance intensifies the importance of this temporal balance because a shared technological dependency can simultaneously generate losses for numerous insured enterprises, interfere with the insurer's own systems, and produce a demand for liquidity across several entities in the same insurance group. The resulting problem is one of financial resilience under correlated disturbances.
A conventional prospective stress test asks what becomes of an insurer under a specified deterioration of its environment. Reverse stress testing starts from an unacceptable outcome and investigates the circumstances under which it becomes possible. This inversion gives the exercise a specific scientific object: identifying the boundaries of a business's resilience and the paths that reach them. A sufficiently detailed answer connects a technological event to insured liabilities, reinsurance contract response, cash settlement, capital valuation, and feasible governance decisions.
The inquiry concentrates on a risk-bearing insurer that writes affirmative cyber insurance or has other coverage potentially exposed to the same event. It examines primary insurance, ceded reinsurance, and group-level resources whenever those features materially affect the viability of the undertaking. Prospective probable maximum loss estimation and the erosion of individual insurance policy limits are complementary subjects; here their significance arises from their contribution to a failure boundary. The analysis does not estimate actual company financial health or recommend capital levels to specific firms.
The research asks which failure outcomes should be distinguished, which transmission mechanisms connect cyber losses with financial distress, and how permanent and temporary buffers change the set of paths that can produce those outcomes. The investigation uses three types of evidence. Prudential instruments establish jurisdiction-specific obligations, supervisory expectations, and definitions. Academic studies and documented incidents support the analysis of dependence, concentration, and transfer capacity. Transparent numerical experiments explain financial mechanisms with explicitly assumed inputs. The evidence classes are kept separate throughout.
The applicable legal texts and supervisory materials are examined as identifiable versions rather than treated as a unified global capital code. The assessment date is November 18, 2026. In particular, the version of SS19/16 applicable on that date is the November 2024 update according to the PRA's publication and version history; the May 2026 update has a later effective date. Legal comparisons below describe the specified instruments and their scientific significance. Their application to a particular undertaking depends on the relevant legal perimeter, domestic implementation, and operative rules.
The analytical method consists of tracing a causal chain, expressing accounting and cash constraints, locating roots at which a specified threshold is reached, and varying uncertain parameters. An illustrative calculation can establish the mechanism through which a loss, delay, or adjustment changes a boundary; it cannot establish the real-world frequency of the modeled shock. All monetary values in the examples are fictional, denominated in a generic unit of millions. The calculations have no empirical claim to represent the portfolio of a real insurer.
Conceptual and Historical Foundations
From individual losses to portfolio distress
Cyber insurance contains several economically different promises. First-party indemnification can include incident response, restoration, and interruption. Third-party liability can involve defense expenditure, settlement, and subsequent legal obligations. Other policy forms may respond to computer fraud, physical damage, or service interruption under particular wording. The consequences of one technological event therefore pass through multiple coverage tests before becoming an insurer's contractual obligation. Economic damage is not identical to insured damage, and insured damage is not identical to the cash disbursed immediately following an event.
The insured population of interest is the portfolio in force at the time of exposure, defined at the appropriate policy, legal entity, and temporal levels. Policies are associated with organizations, infrastructure, software, service providers, and contingent dependencies. Identifying these links provides a basis for investigating common causes. A portfolio of otherwise distinct businesses can be exposed to the same authentication service, cloud region, software update mechanism, or managed security provider. A single event may create highly uneven outcomes across businesses because the duration of interruption, redundancy, contractual definitions, and recovery arrangements differ.
The distinction between an incident and an insured event matters particularly when an insurer includes a wide range of risks in a catastrophe contract. Event aggregation depends partly on contract language, including definitions of occurrence, causal unity, temporal aggregation, and applicable classes of business. A technological common cause can be represented differently in separate primary policies and in a treaty. Financial analysis must preserve these distinctions rather than use event counts as an automatic proxy for reinsurance protection.
The Cloud Down study by Lloyd's and AIR Worldwide (2018) examined cloud dependencies in a modeled population of 12.4 million US businesses. Its analytical importance lies in the explicit consideration of widespread reliance on a provider. Its modeled population is neither a census of cyber insurance policies nor a record of realized paid claims. It illustrates why a risk engineer needs a map of technological dependencies alongside traditional attributes such as industry and insured turnover.
The July 2024 CrowdStrike disruption provides separate documentary evidence of a technological common cause. CrowdStrike's root cause analysis announcement (2024) associates the outage with a content configuration update affecting Windows sensors. Microsoft's contemporaneous outage statement (Weston, 2024) estimated that 8.5 million Windows devices were affected, representing less than one percent of Windows machines. Device counts express technological scale; the relevant insurance denominator would instead be covered entities and covered losses under their respective contracts. This event also establishes the importance of distinguishing malicious operations from accidental technology failures in scenario construction.
A reverse test can use these observations to support plausible shared-dependency mechanisms without treating the historical events as measurements of industry-wide insolvency risk. The exercise should identify which portfolios and contracts would respond under the modeled conditions, whether the technology incident is within the insurance perimeter, and how interruption duration translates into economic damage. The transition from historical incident evidence to portfolio stress is a conditional construction.
The scale and limits of market data
Public market data describe the commercial scale of cyber insurance. The NAIC's 2026 cybersecurity market report records approximately $9.464 billion in direct written premiums for 2025 including alien surplus lines, compared with approximately $9.144 billion for 2024, an increase of about 3.5%. On the report's basis excluding alien surplus lines, the figures are approximately $7.459 billion and $7.083 billion, respectively, an increase of around 5.3%. These amounts use the report's own reporting perimeter. Direct written premium is a flow associated with underwriting activity, not the outstanding stock of available capital, retained catastrophe exposure, or an estimate of aggregate cyber catastrophe loss.
This distinction has an important methodological consequence. A large premium volume may coexist with correlated exposure, low net retention, or high retention depending on the contracts written and the protection purchased. A premium growth rate cannot establish that the industry's loss-absorption capacity grew by the same percentage. Portfolio-level data are necessary to infer the effects of a particular stress. The analyst must account for differences in policy periods, coverage scope, premium reporting, premium allocation, and the reporting populations used in market summaries.
Similarly, aggregate incurred loss ratios describe realized experience within an accounting period. They combine the effects of claims recognition, reserving, exposure growth, and historical conditions. A reverse stress test investigates a different question: how actual financial obligations and resources would behave along an adverse path that may not have occurred in the data. Public market statistics establish context and may inform calibration, but they cannot determine a viability boundary without the insurer's contractual and financial characteristics.
Dependence and the meaning of extremes
The actuarial scientific issue is dependence across insured outcomes. Let X_i represent insured loss under policy i. Portfolio gross loss equals the sum of these losses together with any associated adjustment expenses included in the defined loss measure. The variance of this sum includes both individual variances and pairwise covariances, so the variance under independence can materially misstate aggregate uncertainty when shared factors affect many policies. Dependence in extreme outcomes requires further investigation because a model can reproduce ordinary co-movement yet underestimate simultaneous large losses.
Eling and Jung (2018) examine cross-sectional breach loss dependence using 3,327 data breach events from 2005 to 2016. Their abstract describes modeling monthly dependence with separate treatment of zero and positive losses. This study establishes a substantive precedent for treating cross-sectional cyber dependence as an actuarial problem. Its historical breach data and particular loss measure cannot establish the current insured-event distribution for cloud outages or an insurer's present portfolio. The abstract is the verified scope of citation.
Cyber accumulation has both probabilistic and structural dimensions. Probabilistic models describe uncertain event frequencies and severities under stated assumptions. Structural descriptions map the technical channels through which an event can propagate, identify alternative service pathways, and connect those pathways to policy terms. For reverse stress testing, both dimensions are useful. The structural representation constrains candidate scenarios; probabilistic information helps distinguish relatively credible paths from distant or unsupported possibilities. When empirical evidence is sparse, credibility judgments must remain traceable to documented assumptions rather than concealed inside a precise probability.
A scenario that exhausts a contractual limit may still be financially tolerable, and one that does not exhaust a technical-loss limit may be financially serious if other resources deteriorate at the same time. Severity is therefore defined relative to the undertaking's failure outcome. The financial consequence depends on loss retention, timing, own operations, and associated balance-sheet changes.
Reverse Stress Testing and Failure Boundaries
Supervisory and scientific meaning
The IAIS's Insurance Core Principles (2024), particularly ICP 16 and its associated ComFrame provisions, place prospective risk assessment and stress testing within enterprise risk management. Reverse stress testing starts from a specified adverse outcome and investigates circumstances capable of producing it. The exercise informs risk management and decisions about the undertaking's risk profile. Group assessment must consider the relationships between risks and the entities in which capital and obligations reside.
The PRA's SS19/16 ORSA guidance (2024), sections 7 and 8, addresses reverse testing in relation to business model failure, management actions, and the capital position. OSFI's Stress Testing Guideline (2009) similarly describes starting from an adverse outcome that threatens viability and investigating the contributing events. These sources provide convergent methodological motivation while retaining distinct supervisory and legal contexts.
In the European Union, Solvency II Article 45 establishes the own risk and solvency assessment, including assessment of overall solvency needs, continuing compliance with requirements, and deviations from assumptions underlying the SCR. Article 45(7) distinguishes the ORSA from the calculation of a capital requirement. This is a significant distinction for the research: the loss identified through a reverse assessment does not automatically become a statutory capital charge.
In the United States, the NAIC's ORSA Model Act 505 (2012), sections 6–7, provides an example of a model-law framework rather than a uniformly enacted federal mandate. Its premium exemption thresholds are $500 million for an individual insurer and $1 billion for an insurance group, using the defined direct written and unaffiliated assumed premium measure, subject to the applicable exceptions and supervisory provisions. Actual legal obligations depend on state adoption, the insurer's circumstances, and the relevant guidance. Neither those thresholds nor the existence of an ORSA process determines a specific cyber catastrophe capital amount.
The plurality of failure outcomes
A research exercise begins with precisely specified outcomes. The first is deterioration below a chosen capital target, including a board risk appetite threshold or a legally applicable boundary. The second is insufficient cash or readily mobilizable resources to meet obligations when due. The third is an inability to operate critical functions necessary to adjust claims, receive premiums, administer contracts, or execute recovery actions. The fourth is an unsustainable business configuration arising, for example, when risk transfer capacity cannot be renewed at a viable cost or when an undertaking cannot continue the business assumed in its recovery plan.
These outcomes may be connected in sequence. Liquidity pressure can lead to distressed asset sales, which impair capital. Capital deterioration can reduce access to funding or collateral. Operational disruption can delay both claim processing and reinsurance recovery. An insurer that remains solvent on a valuation basis may nevertheless require liquidity support, while an insurer with ample cash may experience a permanent loss of eligible capital.
The exercise should specify the measurement horizon, legal entity, accounting or regulatory valuation basis, types of eligible resources, and timing of each threshold. For capital adequacy it must differentiate the amount and quality of funds from the capital requirement. For liquidity it must state which assets and facilities can actually be used within the horizon. For operational continuity it must state which critical processes become unavailable and for how long. A business model failure outcome is broader than a single ratio and can involve legally or operationally constrained responses.
A useful outcome definition has the form of an inequality. Let C be eligible own funds and K the relevant capital requirement, both on a consistent valuation basis. An internal target t defines the boundary C - tK = 0. Let B(s) be realizable cash resources at time s less obligations due by that time. The cash boundary is B(s) = 0. The reverse problem asks for admissible scenario paths having C - tK at or below zero, or B(s) at or below zero at some point during the defined horizon. Equality locates the boundary; strictly negative headroom denotes breach. Operational and legal constraints may impose additional boundary conditions.
This formulation allows the analyst to separate initial solvency, stressed solvency, and the feasibility of actions available after the shock. A risk committee may be concerned with a capital target well before a statutory boundary is reached. Both boundaries should be explicitly reported. A regulatory capital shortfall and legal insolvency are distinct legal and economic events; the former does not itself establish the latter.
Scenario construction and search
Let a scenario vector contain gross incurred claim loss, time distribution of payments, treaty recovery, counterparty impairment, expense and asset losses, and relevant operational constraints. Each component has a defined unit and place in the balance sheet or cash projection. The scenario search may identify the smallest weighted disturbance capable of reaching a specified outcome, subject to feasibility constraints. Weights define the meaning of smallness; they may arise from a documented statistical model, a structured expert assessment, or an explicitly stated engineering distance.
A statistically most likely scenario is meaningful only when the model provides a coherent probability measure over the candidate paths. A model lacking reliable joint probabilities can instead compare scenarios by transparent severity bounds and dominance relationships. Breuer and Csiszár (2013) describe entropic plausibility constraints for systematic stress tests in their published abstract. Distributional approaches of this kind constrain deviations from a reference distribution; their relevance here is methodological, and no specific implementation or mathematical result beyond the verified abstract is asserted. A point scenario under a fixed set of inputs is a different object from a worst-case distribution.
Dependence, contract terms, and timing create nonlinearities and discrete transitions. For example, treaty attachment changes marginal retention, and treaty exhaustion restores a higher marginal retained loss. An exclusion or event-definition dispute can alter the recoverability estimate discontinuously. A facility may cease to be available when a covenant is breached. Therefore a search based solely on one smooth local approximation can miss an important path. A robust exercise evaluates both the continuous regions and the transition points, and preserves alternative paths that reach the same outcome through different mechanisms.
A practical search starts by specifying a finite set of technical event families and mapping them into plausible insurance response. It then varies loss amounts, the distribution of payment times, contract aggregation, recovery delay and collectibility, and correlated deterioration of other resources. The search must retain the event definitions and assumptions that generated each crossing. The board and supervisors can then interpret whether a reported boundary is a plausible stress, an extreme structural sensitivity, or an artifact of the model.
A synthetic capital boundary
Consider an insurer with initial eligible capital C0 = 300 and a capital requirement K0 = 180. Its starting ratio is 166.67%. Assume an incremental non-claim economic loss D = 15 and, for illustration, a stressed capital requirement K(g) = 180 + 0.10g, where g is the gross ultimate covered claim loss. This linear function is an assumed sensitivity, not the Solvency II standard formula or any other statutory calculation.
Suppose a single excess-of-loss treaty covers 120 in excess of 40. Define nominal recovery R(g) = min(max(g - 40, 0), 120). Assume all modeled loss qualifies for that occurrence and that the treaty has no further restrictions, reinstatement cost, or prior exhaustion. Let q between zero and one be the fraction of nominal recovery ultimately valued as collectible in the scenario. Here q is a scenario valuation assumption, not a default probability. Discounting is omitted in this simplified exercise. Stressed capital is C(g,q) = 285 - g + qR(g). The requirement increases as stated independently of this recovery mechanism. This simplification isolates the effects of recoverability and retention; a regulatory engine would recalculate all relevant risk components and eligibility.
The headroom at the 100% boundary is 105 - 1.10g + qR(g). At the 125% internal target it is 60 - 1.125g + qR(g). The roots can be obtained in each region of the treaty by substituting R(g) = 0 below 40, R(g) = g - 40 from 40 to 160, and R(g) = 120 above 160. The following values are computed from these equations.
| Collectible fraction q | Gross loss at 125% target | Gross loss at 100% boundary |
|---|---|---|
| 1.00 | 160.00 | 204.55 |
| 0.75 | 80.00 | 177.27 |
| 0.50 | 64.00 | 141.67 |
| 0.00 | 53.33 | 95.45 |
The difference between the two columns gives the range of gross stress over which the undertaking falls below the internal target while remaining at or above the modeled 100% threshold. The range is sensitive to q because improved collectibility reduces the loss retained in capital and interacts with the location of the treaty's attachment and exhaustion points. A comparison between the rows identifies the value of recovery protection conditional on a specified gross loss path.
At g = 160, full collection reaches the 125% internal target exactly, since C = 245 and K = 196. Beyond expiration of that layer, incremental losses receive no additional treaty recovery. The additional gross loss capacity before the 100% boundary is approximately 44.55. With q = 0.75 the 100% boundary is approximately 177.27, illustrating how the same contractual limit supports a different financial resilience outcome once recoverability is impaired. The result is sensitive to the assumed increase in K and the non-claim loss D.
The calculation supplies a transparent diagnostic. It provides neither a market premium recommendation nor a prudential minimum. Its scientific contribution is the isolation of the conditional response of a capital boundary to the contract's piecewise structure and recovery valuation.
Liquidity Channels and Reinsurance Collection
From an ultimate loss estimate to payments
The gross ultimate loss recognized in a financial projection and the gross amount paid by a particular date are different quantities. Incident response expenditure may occur rapidly; interruption loss may accrue over time; liability litigation can take considerably longer. An undertaking's payments depend on reporting, proof, claims handling, policy conditions, settlement, and legal obligations. Recoveries from ceded reinsurance may follow different notice, documentation, payment, and dispute-resolution timetables.
EIOPA's cyber stress testing methodology (2023) explicitly examines both cyber risk arising from underwriting and the operational cyber resilience of the insurer itself. Its treatment of stress testing includes the financial consequences of policy response and reinsurance, and consideration of associated cash flows, including claim payments and reinsurance receivables. The distinction is methodologically important: an insurer can face surging demands on its operational systems at precisely the moment at which policyholders initiate numerous claims.
The PRA maintains a separate liquidity risk management statement (2019). The existence of a distinct supervisory instrument reflects the substantive separation of liquidity from capital assessment. A projected asset may contribute to a valuation-based capital position while remaining unavailable for payment at a specific date. The scientific task is to model the actual availability of resources, not merely their classification as assets.
Let h denote a cash horizon. Define B(h) as immediately available unencumbered cash, contractually and operationally realizable inflows, and monetizable assets, less cumulative required outflows and associated costs through h. Cash projected for h must be consistent with the same legal entity, currency, and operational constraints as the obligations. Where a financing facility is included, its draw conditions, collateral requirements, currency, and availability during the stress must be documented.
Liquidity across a path is determined by the minimum of B(h), not merely by the balance at the final date. A cash deficit can be temporary if recoveries subsequently arrive, but payment failure may cause costs or legal consequences before those inflows become available. Conversely, the achievement of a positive balance at one horizon does not establish adequate resources at subsequent horizons.
A joint capital and liquidity experiment
Retain the earlier synthetic portfolio. Assume that cash and promptly mobilizable assets at the modeled 30-day horizon amount to 85 before claim payments, and unrelated net outflows are 10. Suppose 65% of gross ultimate covered losses must be paid by day 30. If no reinsurance recovery is then available, liquidity is B30 = 75 - 0.65g. The cash boundary occurs at g = 75 / 0.65, approximately 115.38.
Assume next that one quarter of the collectible recovery value is deposited by day 30. The equation becomes B30 = 75 - 0.65g + 0.25qR(g). When q = 1, the headroom remains positive at g = 160, where it equals 1. Above treaty exhaustion the expression is 105 - 0.65g. The first cash boundary is then approximately 161.54. In both experiments the ultimate collection assumption is unchanged; only the timing of the recovery is varied.
At g = 120 and q = 1, nominal treaty recovery is 80 and modeled capital is 245, with K = 192. The capital ratio is approximately 127.60%, above the assumed internal target. Day-30 liquidity is negative 3 without receipts and positive 17 with the quarter-recovery assumption. This single example establishes the possibility of an internal capital target being satisfied at the same time as a cash deficit. The deficit does not result from an inconsistency between the models: capital recognizes the value of eventual recoveries while cash recognizes only actual availability by the prescribed time.
At g = 150 and q = 0.75, modeled nominal recovery is 110 and collectible recovery is 82.5. Capital is 217.5 and the requirement is 195, giving a ratio of approximately 111.54%. The day-30 cash balance without receipt is negative 22.5; the receipt of one quarter of collectible recovery improves it to negative 1.875. This case shows why both the amount and the timing of recoveries matter. Accelerating a receipt helps but may be insufficient to close the cash deficit.
These equations omit discounting, financing costs, and feedback between cash actions and capital. When payment discharges a liability already included in capital valuation, the payment itself does not create a second ultimate claim loss. Similarly, collecting a previously recognized receivable converts an asset into cash rather than creating additional economic capital. A full projection includes only incremental economic changes, such as new information about ultimate loss, impairment, discount unwinding, or asset sale losses, in the capital waterfall. Its cash waterfall separately records actual inflows and outflows.
Recovery valuation, contract response, and counterparties
Under Solvency II Article 81, calculation of amounts recoverable from reinsurance contracts and special purpose vehicles takes account of the time difference between recoveries and direct payments, together with expected losses from counterparty default. This provision highlights a distinction that is essential to reverse testing: nominal contract response, economic value of recoverables, and cash received on a specific date are distinct magnitudes. A loss scenario may deteriorate the value of a recoverable without entirely eliminating the nominal contractual agreement.
Four separate sources of uncertainty should be maintained in a contract-specific projection: the gross insured amount; the amount meeting treaty definitions; the amount economically recoverable after considering counterparty and dispute risks; and the timing of receipts. A single aggregate haircut can be useful in a simplified sensitivity test, as q is used above. It cannot explain the governance response unless the exercise specifies which channel the haircut represents.
For instance, a disagreement over event aggregation can change the attachment and number of available limits. A payment delay can result from a requirement for additional claims evidence. A counterparty loss may reduce recovery even if coverage is accepted. In a stress scenario, these mechanisms can coexist, but their effects should not be multiplied blindly when they measure overlapping reductions. Separate cash and valuation schedules allow the analyst to identify whether a proposed action improves the amount ultimately received or merely the time at which it becomes available.
The distinction can be demonstrated with a second synthetic example. Assume two insured loss clusters of 70 each, fully covered by the same 120 excess of 40 layer, with an annual recovery ceiling of 120 and no other restrictions. If the treaty treats both clusters as one occurrence, the 140 gross loss produces a nominal recovery of 100. If they are separate occurrences, each triggers 30, totaling 60. At q = 1, using the earlier assumptions for a total gross loss of 140, capital is 245 in the aggregated treatment and 205 in the separate treatment. With K = 194, the ratios are approximately 126.29% and 105.67%, respectively. Both legal constructions are assumptions here; only the wording and circumstances of a real treaty could determine which applies.
The result is not a universal rule that aggregation favors the cedent. If combined losses exceed one available limit, separate occurrences with reinstated protection may increase total recovery. Annual ceilings, reinstatement provisions, hours clauses, and the distribution of each event across the covered loss components determine the direction of the effect. The scientifically relevant finding is that a classification decision can move an undertaking across a capital target even when the same gross insured amount is assumed.
Public cyber-related coverage disputes provide evidence that wording and attribution can matter considerably. The New Jersey appellate decision in Merck v. ACE American (2023) interpreted a particular hostile or warlike action exclusion in the context of the litigated policies, focusing on its military-action meaning. This decision is used as a historical illustration of contract-specific interpretation, not a statement that subsequent policies or all jurisdictions apply the same result. Lloyd's Market Bulletin Y5381 (2022) sets out historical market requirements concerning state-backed cyber-attack exclusions, including attribution and significant impairment. The bulletin is a dated instruction for its stated market scope. These sources establish why a reverse test should examine actual wording and credible alternative response where a coverage question remains unresolved.
Own operations and payment continuity
The own-operational channel can affect the financial position even when the triggering incident is unrelated to the underwriting portfolio. Claims intake may be interrupted, financial controls may become temporarily inaccessible, and the insurer may be unable to transmit instructions or supply information to a reinsurer. These are possible mechanisms to test rather than empirical claims about any given insurer. Their joint occurrence with underwriting loss can lengthen the time to settlement and produce additional expenses.
Payment continuity should therefore be evaluated within the actual configuration of an undertaking: access to payment systems, claims administration, authority to authorize exceptional expenditure, records available offline, and dependency on the same providers as the insured population. An asset held in a custody account is of limited immediate use if those empowered to issue instructions cannot access the required systems. A scenario may involve liquidity in economic terms but insufficient operational capacity to use it.
The reverse test can evaluate these assumptions through operational recovery exercises whose results feed into financial scenarios. For example, if alternate payment procedures have been tested and can support a defined volume within a specified time, the cash projection can reflect that demonstrated capability. A claim of immediate recovery without evidence of the underlying process would create a favorable but unsupported assumption.
Solvency Buffers and Prudential Architecture
Capital requirements and internal targets
The European Union's Solvency II Article 101 defines the SCR in relation to a one-year 99.5% value-at-risk measure of basic own funds and includes recognition of risk mitigation subject to the treatment of risks associated with that mitigation. The prescribed calibration concerns the applicable valuation and regulatory framework. It is not a direct statement that a particular cyber event has a 0.5% annual probability, and it cannot be inverted into such a probability without a coherent model of all relevant risks.
Solvency II Article 129 sets the MCR's corridor between 25% and 45% of the SCR subject to the absolute floor and other statutory calculation provisions. The MCR is a separate legal and financial boundary, including requirements concerning eligible resources. An SCR reverse test must therefore distinguish an internal management trigger, the SCR threshold, any further regulatory thresholds, and the separately assessed MCR position. The synthetic 100% threshold above is an ordinary ratio boundary within the example; it is not an MCR calculation.
The ratio of eligible own funds to a capital requirement is informative only when numerator and denominator use the appropriate basis. A loss can impair the numerator while simultaneously changing the risk profile underlying the denominator. Risk transfer can affect the underwriting risk component and introduce credit or other risks. Changes in business mix, retained limits, asset exposure, and group perimeter can affect the denominator. Using a fixed K may be justified for a narrowly defined instantaneous sensitivity but should be identified as such.
A capital buffer measured as C - K is the excess above a specified minimum. A management buffer measured as C - tK is the excess above an internal target. Neither quantity alone measures the cash available to support claim payments, and both are sensitive to the eligibility and location of funds. A group may appear adequately capitalized while a particular legal entity has insufficient eligible resources. An undertaking may also possess subordinated funding or other resources that contribute to an eligible capital measure but cannot provide immediate cash without an additional transaction.
The PRA's SS19/16 discussion of capital and reverse stress reinforces the importance of assessing the quality and availability of resources and the actions that would be taken under adverse conditions. Its supervisory treatment applies within its specified perimeter. In Canada, the OSFI stress-testing guideline supports examining severe combinations of risks relevant to the institution. In the United States, the NAIC model-law approach locates the insurer's own assessment within a state-based regulatory environment. These materials do not yield one universal capital buffer percentage. They support a comparative method that states the jurisdiction, target, resources, and failure definition before reporting a stress boundary.
The economic function of buffers
A solvency buffer absorbs deterioration of assets or an increase in liabilities while preserving the financial resources needed to support continuing obligations. It can also create decision time. If an insurer has sufficient headroom before a mandatory threshold, it may be able to adjust future underwriting, commission funding arrangements, or execute other feasible changes without crossing that boundary during the modeled period.
The economic cost of maintaining a buffer varies with its composition. Cash offers strong immediate liquidity but may have an opportunity cost relative to investments. Other eligible capital resources may provide substantial loss absorption but limited near-term liquidity. Reinsurance transfers specified uncertainty but entails a premium expense and counterparty exposure, while collateralized arrangements can alter the timing and certainty of funding. A retained investment return that contributes to capital may deteriorate under adverse asset conditions.
These effects should be assessed jointly. Consider an undertaking with more initial capital but unchanged cash and treaty structure. In the earlier example, increasing C0 by 30 shifts the capital headroom equation upward by 30. It leaves the cash equation unchanged unless the increase in capital is supplied in immediately usable resources. The 100% boundary with q = 0 and all other assumptions unchanged would move from 95.45 to approximately 122.73. The day-30 cash boundary remains 115.38 without recovery receipts. Increasing capital by issuing equity whose proceeds are maintained as cash instead would also affect the cash equation. The economic form of a buffer matters as much as its accounting amount for the particular failure outcome.
Similarly, a committed facility adds capacity to the cash projection if it is available at the required time. A draw usually creates a corresponding liability and thus does not provide the same permanent loss absorption as an injection of equity. Interest, collateral calls, or changes in market valuation can generate secondary effects. A contingency facility is valuable for liquidity, but its ability to sustain solvency depends on its terms and eligibility under the relevant regime.
Reinsurance and alternative transfer can reduce gross catastrophe retention within the response of the instrument. Cremer and colleagues (2024), in their study of cyber risk transfer, report findings from 20 semi-structured interviews concerning reinsurance, alternative risk transfer, and available capacity. Their qualitative findings help identify institutional issues to consider, including the diversification of capacity. They do not provide calibrated failure frequencies or demonstrate that a specific alternative instrument is invariably effective for cyber risk.
A collateralized structure can be analyzed by identifying the contractual trigger, collateral holding arrangements, permitted assets, timing of release, and the extent of any basis risk. Protection based on an external index may pay a different amount from the undertaking's own insured losses. Protection triggered by the indemnified loss may take longer to establish and verify. These are contract design trade-offs, and their consequences can be examined by replacing the relevant recovery schedules within the stress experiment.
Management actions under a common shock
An effective reverse stress assessment distinguishes pre-event risk mitigation from post-event intervention. Pre-event measures may include changes in underwriting limits, purchasing reinsurance, diversifying exposures, holding liquid assets, and establishing tested contingency arrangements. Post-event measures may include drawing committed funding, accessing group resources, selling assets, or altering future underwriting. The latter are constrained by the time required for execution and by the situation after the shock.
The analysis should calculate a boundary before management actions and a further boundary conditional on actions whose feasibility has been demonstrated. The difference between these boundaries measures the conditional value of the actions within the model. No benefit should be attributed twice when two actions draw on the same asset or facility. If an action improves liquidity by selling assets, the consequences of the sale for the remaining capital position must also be calculated.
Under a common cyber shock, assumptions about external capacity require particular scrutiny. Potential counterparties may be affected by the same event, market pricing may change, and operational disruption may affect transaction execution. A group-level resource is available to a subsidiary only if its transfer is legally authorized, operationally executable, and possible within the cash horizon. The stress assessment should inspect aggregate group resources and the entity-specific constraints governing their use.
Management actions can therefore be specified as conditional transactions with owner, earliest execution date, maximum amount, sequence, eligibility, required approvals, and adverse feedback. A choice to restrict new underwriting may reduce future exposure while leaving current claim obligations unchanged. Renewal of a treaty may affect the next underwriting period but provide no retroactive protection for a preceding occurrence. The time at which an action starts affecting each balance must be identified.
Evidence Validation and Sensitivity Analysis
Evidence by type and inference
A research dossier on extreme financial outcomes requires a disciplined distinction between observation, legal authority, academic inference, and model assumption. The IAIS framework and jurisdiction-specific instruments establish the scope and rationale of assessments. They do not calibrate the illustrative insurer used here. The NAIC market report establishes premium amounts within its reporting perimeter. It does not measure the failure boundary of a particular insurer. Corporate incident reports establish aspects of technical events, while neither the Microsoft device estimate nor the cloud model establishes the number of affected policies.
The regulatory evidence addresses different horizons and scopes. For example, Article 101's one-year calibration does not prescribe a 30-day cash constraint. An ORSA involves the undertaking's own assessment rather than simply a repetition of a regulatory ratio. Market bulletins and case law address stated contractual or market circumstances, and their significance elsewhere depends on the governing contract and applicable law.
Academic evidence provides methods and observed empirical associations within its own datasets. The Eling and Jung study concerns data breach losses, not the loss distribution of every cyber coverage component. The Breuer and Csiszár reference is used for the distributional stress-testing idea stated in its abstract, with no inference about inaccessible technical details. The Cremer and colleagues article supplies qualitative institutional evidence, not numerical calibration. The differing evidentiary scopes explain the different use of these works in the argument.
The historical PRA Insurance Stress Test 2022 feedback (Gerken, 2023) identifies 54 participants: 17 general insurers, 21 Lloyd's syndicates, and 16 life insurers, with cyber among the general insurance stress areas. This is documentary evidence of supervisory attention to underwriting cyber stress. Participation in that exercise does not provide evidence that any specific firm would resist a scenario of the form used in this monograph. The scientific significance is the use of stress analysis at the level of regulated insurance business.
Evidence quality in a firm-specific implementation depends on the provenance and reconciliation of policy, claims, reinsurance, and finance data. An exposure register should identify the in-force period and common technology dependencies. A claim history should distinguish gross and net amounts, expense definitions, paid amounts and estimates, and maturity of development. A contract inventory should contain the governing language and actual limitations. A capital and cash opening balance should reconcile with the appropriate reporting basis. Data quality limitations must propagate into the range of reverse scenarios rather than be suppressed.
Sensitivity of the capital boundary
The piecewise example admits analytical sensitivity calculations. In the interior of the reinsurance layer, R(g) = g - 40. For target t, the headroom equation is 285 - 180t - 40q - (1 + 0.10t - q)g = 0. The interior root is therefore
g = (285 - 180t - 40q) / (1 + 0.10t - q),
provided it lies between 40 and 160. Below attachment the root is obtained with qR = 0; above exhaustion it is obtained with qR = 120q. For the values used here, 1 + 0.10t - q is positive, so headroom decreases throughout the modeled layer. Each root must be checked against its region. This qualification matters because extending the interior equation beyond the exhaustion point would exaggerate the benefit of the treaty.
At q = 0.5 and t = 1, the interior root is (105 - 20) / 0.6 = 141.67, within the layer. At q = 0.75 and t = 1, the interior expression would yield a root outside the layer; the correct exhaustion-region equation gives (105 + 90) / 1.1 = 177.27. At q = 1 and t = 1.25, the interior root is exactly 160, at the boundary between regions. These checks establish consistency between the table and the piecewise contract response.
The capital requirement sensitivity itself is an important input. Replace 0.10 by a nonnegative slope a. At q = 1 and above exhaustion, the 100% capital root becomes 225 / (1 + a). For a = 0, 0.10 and 0.20, the respective roots are 225, approximately 204.55 and 187.50; all remain above 160. A higher assumed capital requirement response reduces the tolerated gross loss even with full collection. A real regulatory calculation can produce changes that differ materially from these linear examples. The analysis indicates a direction conditional on the assumed function, not a universal law about regulatory capital.
The incremental non-claim loss D also matters. Above exhaustion, holding the other assumptions fixed, the 100% boundary for full collection is (300 - D - 180 + 120) / 1.10. If D increases from 15 to 35, the root falls from approximately 204.55 to 186.36. The 20-unit additional permanent loss reduces gross stress capacity by approximately 18.18, reflecting the concurrent increase in the denominator. If D represents additional claim-related expense already included in g, it should be removed from D to avoid double counting.
A structural sensitivity can also replace the single occurrence assumption with a documented range of plausible aggregation responses, as in the 140-unit example. The point is to examine a legal and contractual discontinuity that a continuous q sensitivity alone might conceal. An adequate test records the nominal recovery under each interpretation, the financial boundary associated with it, and the basis for selecting or retaining the alternative.
Sensitivity of cash depletion
Let p denote the proportion of ultimate gross claims paid by day 30 and f the proportion of collectible recovery received by then. The simplified cash equation is B30 = 75 - pg + fqR(g). Assume no early recovery. With p values of 0.50, 0.65, and 0.80, the first cash boundary is respectively 150.00, approximately 115.38, and 93.75. Faster claim payment reduces the gross stress needed to exhaust immediately available resources; the same ultimate capital result can therefore be compatible with several cash outcomes.
With q = 1, p = 0.65 and f = 0.25, the layer-region cash headroom is 65 - 0.40g. Its extrapolated root would be 162.5, beyond exhaustion; the correct root is 161.54 in the exhaustion region. Again the region check prevents a false result. With the same p but f = 0, the first cash root falls within the layer, where the treaty has no immediate cash effect because no receipts are assumed. This illustrates a recurring difference between economic risk transfer and short-horizon funding.
| Sensitivity assumption | First day-30 cash boundary |
|---|---|
| 50% claims paid; no collection | 150.00 |
| 65% claims paid; no collection | 115.38 |
| 80% claims paid; no collection | 93.75 |
| 65% claims paid; 25% of full collectible recovery received | 161.54 |
A realistic cash projection must use more than one horizon and must incorporate the minimum balance over the path. For example, a large payment at day five and reimbursement at day twenty may generate a deficit before the favorable day-30 balance. The example here cannot rule out such an interim failure. Researchers should construct cash-flow buckets whose granularity reflects the timing of the relevant obligations and the time needed to execute financing actions. Cash in one currency may also require conversion before it can settle an obligation in another; transfer restrictions and unfavorable market conditions can affect the amount available.
Validation of the reverse search
A transparent reverse search is reproducible if the exposure perimeter, scenario inputs, functions, and crossing criterion are known. The synthetic capital and cash examples use monotone headroom functions on their treaty regions, making bisection or direct algebra sufficient. A more general model may exhibit nonmonotonicity, multiple disjoint feasible regions, or discontinuities. A search algorithm should evaluate region endpoints, examine changes in contract response, and avoid treating the first root found from an arbitrary starting value as the globally nearest boundary.
Numerical validation should compare the result against a separate calculation or independently checkable formula. At each reported root, one should substitute the unrounded value into the constraint and confirm that residual headroom is near zero within numerical tolerance. Values immediately below and above the root should be tested for the expected sign. The reported table rounds to two decimals only after the root has been evaluated. Rounding is a presentation decision rather than a change to the crossing criterion.
An additional validation concerns the accounting identity linking assets, liabilities, incurred claims, and recovery values. A model that recognizes the same loss as an incurred expense and then deducts its full cash payment a second time from capital will overstate impairment. A model that recognizes an unpaid recoverable and then adds its entire receipt to economic capital will overstate available own funds. The separate balance and cash schedules should be reconciled so that transfer between asset classes does not appear as new wealth.
In a firm-specific setting, a stress test should be re-performed using alternative plausible parameterizations and source definitions. One important alternative is a weaker recovery assumption conditional on the same gross loss. Another is a faster claims payment schedule with delayed recovery. A third is an operationally correlated shock that restricts the use of cash. Results should be reviewed by personnel responsible for insurance contracts, actuarial projection, finance, treasury, and operational continuity, with their factual disagreements recorded.
The present work cannot validate an actual exposure register, real-time claim file, or proprietary reinsurance wording because none was provided. Accordingly, it distinguishes methods that have been examined through the synthetic experiment from institution-specific implementation tasks. It does not attach a probability, return period, or recommended capital amount to the calculated thresholds.
Integrated Scientific Synthesis and Research Agenda
The sequence of financially material events
The investigation brings together three channels that are often measured separately: gross insured accumulation, permanent net loss and capital adequacy, and temporary payment pressure. The causal sequence starts with a technological disturbance. The disturbance creates financial consequences for entities in a business network. Coverage and claims assessment convert some of those consequences into insured liabilities. Reinsurance terms determine nominal ceded response, and the availability of collection determines its economic value and timing. The insurer's own financial and operational resources determine its capacity to manage the resulting obligations.
A scenario can branch at each stage. A technical event may affect fewer insured entities than initially assumed, or a common service dependency may expose organizations omitted from an exposure map. A coverage issue can change the gross insured amount. Aggregation can change the ceded response. Counterparty impairment can reduce the recovery value while the contractual liability remains. A temporary payment delay can exhaust cash without permanently changing ultimate net loss. These branches are substantive sources of financial variation.
The purpose of reverse testing is to connect them to a defined adverse outcome. The analyst should retain the causal construction of a stress path, rather than report a gross loss number detached from the assumptions that generated it. Two paths with the same gross incurred loss can have different liquidity and capital consequences. Indeed, two paths with the same ultimate retained loss can differ sharply in timing and operational feasibility. The synthetic examples establish these relationships without relying on claims about the actual distribution of cyber catastrophe losses.
A consequential interpretation follows. A high nominal limit is insufficient evidence of financial resilience. Protection must be evaluated at the relevant occurrence and aggregate levels, on an appropriate basis of economic recovery, within the time horizon of obligations, with explicit regard to legal entity and operational availability. This conclusion is an analytical synthesis of established contractual, financial, and prudential mechanisms; no new named authorial theoretical system is asserted.
Institutional incentives and governance
An underwriting team may view a contract as accepted capacity, an actuarial team as a change in a loss distribution, and a treasury team as a prospective stream of receipts. These interpretations reflect different organizational responsibilities. A reverse stress assessment integrates them by requiring a common event and consistent contractual response across models, while preserving the separate outputs needed for capital, liquidity, and operational decisions.
Without such reconciliation, favorable assumptions can appear in several places. The underwriting model can assume full treaty recovery; the liquidity plan can assume early receipt; and the contingency plan can assume that the same counterparty supplies further funding. Each assumption may be individually plausible while the combined scenario is weakly supported. Reviewing the scenario as a joint construction reveals these dependencies.
The governance value of the exercise arises from the possibility of identifying actions before the boundary is reached. An undertaking can investigate concentration limits, contractual clarity, collateral, available liquid resources, and operational workarounds with reference to specific adverse paths. It can distinguish actions that preserve capital from those that accelerate cash availability. It can also identify which risks would remain beyond its chosen tolerance after the proposed measures.
An actionable result should contain a clearly stated failure outcome, scenario pathway, exposure and contract perimeter, time to crossing, sensitivity to uncertain assumptions, candidate actions, and assessment of the resources required to execute those actions. It should distinguish the amount of financial headroom from the confidence in its measurement. Information about the likely mechanism is as important for decision making as a headline loss figure.
Further scientific questions
Research would benefit from datasets connecting the dependency structure of insured entities with contract-level response, ultimate development, payment timing, and reinsurance collection. Such data would allow researchers to estimate the joint distribution of gross loss, net loss, and funding needs for comparable exposures. Access raises practical concerns about commercial sensitivity, privacy, coverage terms, and the security implications of publishing dependency maps. Aggregated or carefully controlled data arrangements may facilitate work while restricting disclosure of identifiable dependencies.
A second area concerns dependence that changes with technology and business practices. Historical data breach relationships may provide information about certain risks but are not necessarily representative of current cloud architecture, managed service arrangements, or shared software dependencies. Researchers should investigate the stability of estimated relationships under changes in technology, exposures, contractual language, and security controls.
A third area concerns the joint behavior of cedents and reinsurers. The financial effect of correlated underwriting losses is influenced by the distribution of risk across risk-bearing entities and their own protection arrangements. A research program could examine the concentration of recoverables by counterparty, geography, collateral structure, and contractual trigger, jointly with the time distribution of payments. The evidence required to establish the resulting systemic effects is substantially more detailed than gross market premium figures.
A fourth area concerns validation of management responses. A scenario's assumed recovery action can be compared with actual operational testing, transaction documents, legal permissions, and prior experience of execution. Capital resources and liquidity sources should be evaluated in relation to their verified availability. A research program could examine how the reliability of management actions changes under concurrent operational and market stress, rather than assuming that actions available in ordinary conditions remain feasible during a crisis.
These scientific questions identify areas for further empirical inquiry. The analytical conclusions of the present work are limited to the verified supervisory and documentary sources and to the conditional mathematical experiments. They provide a basis for asking empirically tractable questions about cyber insurance resilience rather than claiming that the underlying market distributions have already been measured.
Conclusion
Cyber insurance reverse stress testing identifies the financial and operational paths by which an undertaking can reach an unacceptable outcome. Its scientific value lies in specifying the outcome in advance and tracing the consequences of technology events through insured liabilities, reinsurance, valuation, payments, and management actions. Capital headroom, liquidity headroom, and operational continuity form distinct dimensions of the undertaking's resilience.
International and jurisdiction-specific supervisory materials support prospective assessment of adverse conditions, reverse stress testing, and own risk and solvency assessment within their respective regulatory contexts. Their application requires explicit attention to the applicable legal entity, valuation basis, effective date, and reporting scope. The Solvency II SCR is a defined regulatory capital measure; an internal solvency target and a minimum cash balance provide additional, separately defined boundaries.
The synthetic experiments establish the joint importance of gross loss severity, treaty attachment and exhaustion, recoverability, and the settlement horizon. Under the modeled assumptions, a gross loss of 120 with full ultimate collection produces a capital ratio of approximately 127.60% while producing a day-30 cash deficit of 3 if no recoveries have arrived. Earlier receipts can remove that deficit without increasing ultimate nominal protection. Alternative contractual aggregation of the same gross amount of 140 moves the modeled ratio from approximately 126.29% to 105.67%. The examples demonstrate distinct financial mechanisms; they do not estimate the frequency of a real insurer's distress.
Solvency buffers consequently require interpretation in terms of their amount, quality, location, and availability. Permanent capital, contingent liquidity, and reinsurance protection contribute through different transactions and time profiles. Stress analysis should recognize the resources an insurer can actually mobilize, the losses it must ultimately absorb, and the operational conditions determining its ability to execute a response.
The overall conclusion is that the financial resilience of cyber insurance is best investigated through a consistent reconstruction of the entire stressed pathway, with explicit uncertainty and independently checkable calculations. Such an investigation can clarify the margin separating a manageable disturbance from a failure outcome and support decisions about exposure, transfer, capital, liquidity, and continuity. Its scientific integrity depends on distinguishing documented evidence from assumptions, legal requirements from management choices, and reproducible conditional results from claims of empirical probability.
References
International Association of Insurance Supervisors (IAIS). (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups. December 2024. Source
European Insurance and Occupational Pensions Authority (EIOPA). (2023, July 11). Methodological principles of insurance stress testing: Cyber component. EIOPA-BoS-22/514. Source
Prudential Regulation Authority (PRA). (2024). Solvency II: Own Risk and Solvency Assessment. Supervisory Statement SS19/16, November 2024 update. Source
Prudential Regulation Authority (PRA). (2016–2026). Solvency II: Own Risk and Solvency Assessment. SS19/16 publication and version history. Source
Office of the Superintendent of Financial Institutions (OSFI). (2009). Stress Testing Guideline. Source
National Association of Insurance Commissioners (NAIC). (2012). Risk Management and Own Risk and Solvency Assessment Model Act (Model 505). Source
National Association of Insurance Commissioners (NAIC). (2026). Report on the Cybersecurity Insurance Market. Source
European Parliament and Council of the European Union. (2009). Directive 2009/138/EC (Solvency II), Article 45: Own risk and solvency assessment. EIOPA Single Rulebook. Source
European Parliament and Council of the European Union. (2009). Directive 2009/138/EC (Solvency II), Article 101: Calculation of the Solvency Capital Requirement. EIOPA Single Rulebook. Source
European Parliament and Council of the European Union. (2009). Directive 2009/138/EC (Solvency II), Article 129: Calculation of the Minimum Capital Requirement. EIOPA Single Rulebook. Source
European Parliament and Council of the European Union. (2009). Directive 2009/138/EC (Solvency II), Article 81: Recoverables from reinsurance contracts and special purpose vehicles. EIOPA Single Rulebook. Source
Gerken, C., Prudential Regulation Authority. (2023, January 23). Insurance Stress Test 2022: Feedback. Source
Prudential Regulation Authority (PRA). (2019). Liquidity risk management for insurers. Supervisory Statement SS5/19, publication page. Source
Lloyd's and AIR Worldwide. (2018, January 23). Cloud Down: Impacts on the US economy. Research report. Source
CrowdStrike. (2024, August 6). Channel File 291 Incident: Root Cause Analysis available. Source
Weston, D. (2024, July 20). Helping our customers through the CrowdStrike outage. Microsoft. Source
Lloyd's. (2022, August 16). Market Bulletin Y5381: State backed cyber-attack exclusions. Source
Superior Court of New Jersey, Appellate Division. (2023, May 1). Merck & Co., Inc. v. ACE American Insurance Co., consolidated Nos. A-1879-21 and A-1882-21. Source
Cremer, F., Sheehan, B., Mullins, M., Fortmann, M., Materne, S., & Murphy, F. (2024). Enhancing cyber insurance strategies: Exploring reinsurance and alternative risk transfer approaches. Journal of Cybersecurity, 10(1), tyae027. DOI: 10.1093/cybsec/tyae027. Source
Breuer, T., & Csiszár, I. (2013). Systematic stress tests with entropic plausibility constraints. Journal of Banking & Finance, 37(5), 1552–1559. DOI: 10.1016/j.jbankfin.2012.04.013. Source
Eling, M., & Jung, K. (2018). Copula approaches for modeling cross-sectional dependence of data breach losses. Insurance: Mathematics and Economics, 82, 167–180. DOI: 10.1016/j.insmatheco.2018.07.003. Source