RESEARCH DOSSIER
Stress Testing and Scenario-Based Risk Evaluation
Abstract. Stress testing evaluates the consequences of adverse conditions by changing specified risk drivers, their dependence, the time sequence of events, or the operating environment in which losses occur. Scenario-based risk evaluation adds coherent narratives, explicit causal links, exposure assumptions, and decision-relevant consequences. This monograph develops the mathematical and empirical foundations of both approaches for insurance and broader risk systems. It distinguishes deterministic sensitivity shocks, internally consistent multivariate scenarios, stochastic scenario ensembles, and reverse stress tests. It explains why severity, plausibility, probability, and decision usefulness are separate properties; why insurance losses depend on liability terms, asset valuation, counterparty performance, and liquidity as well as insured events; and why a common shock can radically change tail losses while leaving expected aggregate loss unchanged. Original illustrative derivations demonstrate dependence-sensitive Value at Risk and Expected Shortfall, balance-sheet solvency erosion, competing liquidity constraints, and reverse-stress breakpoints. Comparative institutional evidence from the IAIS, European insurance stress tests, the UK Prudential Regulation Authority, U.S. ORSA practice, and climate-risk assessments is evaluated with strict distinctions between supervisory exercises, company-specific modeling, and general probability theory. The central conclusion is that a credible stress-test result is a conditional statement about a defined state of the world and a specified system response. Its evidential strength depends on scenario consistency, the chosen exposure boundary, complete cash-flow mapping, model validation, and transparent treatment of uncertainty. The work provides an original critical synthesis of established research and regulatory material without proposing a new named theory, doctrine, or proprietary model.
Keywords: Stress testing; scenario analysis; insurance risk; tail dependence; extreme losses; compound risk; enterprise risk management; ORSA; solvency capital; insurer liquidity; reverse stress testing; climate scenarios; systemic risk; model uncertainty; risk evaluation.
Introduction
The central scientific question of stress testing is how a specified configuration of adverse conditions changes the losses, financial capacity, and decisions of a risk-bearing system. The answer requires a precisely identified object of evaluation. For an insurer, that object can be annual underwriting result, ultimate claims cost, eligible own funds, capital requirement, available cash, collateral, reinsurance recoveries, or the ability to honor policyholder obligations through time. A stress scenario that materially changes one measure may have a modest effect on another. Consequently, a statement that an insurer “survives the stress” has meaning only when the measurement basis, horizon, legal entity, management-response assumptions, and threshold are declared.
The international supervisory basis is particularly clear. The International Association of Insurance Supervisors describes stress testing as measuring the financial impact of stressed factors and scenario analysis as studying combinations of circumstances. Its Insurance Core Principles, ICP 16 also recognizes the value of reverse stress testing and of using scenarios to expose model limits (IAIS, 2024). The IAIS principles concern insurer enterprise risk management for solvency purposes. A separate, broader methodological tradition includes risk measurement, Monte Carlo simulation, dependence modeling, decision theory, and sensitivity analysis. The two traditions meet when results are used to test financial resilience and inform choices.
This monograph treats scenario-based risk evaluation as a conditional inquiry: given a scenario and specified structural assumptions, what outcomes follow, what assumptions dominate, and what observations could change the conclusion? That formulation respects the important difference between a modeled conditional consequence and an unconditional prediction. An adverse scenario may be designed for severity and coherence without carrying a statistically defensible annual probability. Reverse stress testing begins instead with a failure condition and identifies combinations of events and responses that reach it. Both directions are needed to understand the boundary of resilience.
The distinction also determines how evidence should be interpreted. The 2024 EIOPA insurance exercise and the 2025 UK life insurance exercise are carefully defined supervisory experiments on participating institutions, reporting dates, regulatory balance sheets, and prescribed actions. They reveal vulnerabilities and strengths inside those boundaries. They do not license a universal probability of insurance-industry failure, and their outcomes are not directly comparable without adjusting for jurisdiction, population, regulation, valuation, scenario, and date. This study follows the EIOPA 2024 report and the PRA LIST 2025 results on those exact terms (EIOPA, 2024; Bank of England, 2025).
The method is a structured synthesis of primary standards, official supervisory exercises, publisher records of scientific literature, and mathematical demonstrations. Numerical examples are original, fully hypothetical constructions whose purpose is logical verification, not estimation of actual insurers. Sources for legal duties are distinguished from guidance and from analytical analogies. The endmatter specifies accessible source records; the associated research evidence ledger records which factual claims each document supports. The study is international in scope and provides research rather than institution-specific financial advice.
Stress Event Definitions
The object of a stress experiment
A stress event is a specified departure from a baseline condition, expressed through observable or modeled variables and interpreted within a defined system. Let the baseline state be x₀, the vector of stressed factors be x(s), and the mapping from factors to consequences be F. A deterministic scenario result is Y(s) = F(x(s), θ, e, a), where θ denotes model parameters, e represents exposures and contract terms, and a represents management actions. This compact notation identifies four logically separate sources of variation: the shock, the loss mechanism, the portfolio to which it applies, and the response after the shock. It does not assume that F is linear, that its parameters are estimated without error, or that every component of the scenario has a known probability.
An elementary sensitivity shock varies one parameter, such as claims inflation or interest rates, while other modeled inputs are held fixed. It is valuable for revealing local direction and magnitude. A multivariate stress specifies jointly evolving drivers, including possible causal ordering and feedback. A historical reconstruction replays selected observations or relationships from a past episode on current exposures. A hypothetical scenario selects adverse conditions for relevance or structural concern, even when no past event is a close analogue. A stochastic stress generates a distribution of results under a specified probability model; an illustrative deterministic scenario produces a defined conditional result. A reverse stress identifies states that cross a specified financial or operational boundary.
These methods answer different questions. “What if claims severity rises 20%?” is a sensitivity question. “What happens if inflation increases settlement costs while a market drawdown reduces asset liquidity and a reinsurer delays payment?” is a combined scenario. “What states bring the available capital below the relevant requirement?” is a reverse-stress question. They can share underlying models and data yet generate different evidence. Formal differentiation matters because false equivalence between them often leads to unearned statements about frequency or likelihood.
The ISO 31000:2018 risk-management guidelines establish a general management context for identifying, analyzing, evaluating, treating, monitoring, and communicating risks (ISO, 2018). They provide principles and a framework, rather than a numerical prescription for insurer shock calibration. The IAIS supplies insurance-specific expectations, including interaction between risk appetite, capital adequacy, enterprise risk management, and forward-looking assessment. EIOPA’s ORSA Guidelines similarly locate prospective assessment within insurers’ own risk management, with an EU regulatory setting (EIOPA, 2015). The operational question for this dossier is how to produce and interpret the quantitative evidence that such processes require.
Severity, plausibility, and likelihood
Severity refers to how adverse the consequences or driving conditions are against a specified benchmark. Plausibility is a judgment concerning internal consistency, physical possibility, institutional constraints, or credible causal pathways. Likelihood is a probability claim under a specified model, reference population, and time horizon. A scenario can be severe and plausible with no agreed probability. A high-probability scenario can be relatively benign. An impossible combination may generate a large modeled loss while teaching little about a credible mechanism, unless deliberately used to probe implementation limits.
For a loss variable L, a conventional quantitative baseline includes E[L], a high quantile VaRα(L), and a tail expectation such as Expected Shortfall. Stress testing may instead report L(s) conditional on a specified scenario s. A reported stressed loss of 100 units is not inherently a “one-in-100-year” outcome. The latter label requires a defined exceedance event, stationarity or other time-dependent probability assumptions, and a calibrated probability model. The difference grows in the presence of climate transition, cyber infrastructure changes, developing legal liability, and feedback from firms’ own responses.
Statistical calibration can provide disciplined reference points. Yet a historical sample represents the risks, exposure mix, recording methods, controls, and regime that generated it. Extreme observations may be sparse, and physical or economic processes may change. The value of a historical replay lies in an intelligible empirical anchor; it does not remove uncertainty about the current portfolio. The IPCC 2023 synthesis assessment emphasizes interactions among climatic and nonclimatic drivers, shifting exposures, and compound or cascading impacts (IPCC, 2023). Those findings support the need for scenario construction that allows more than isolated historical event reproduction.
Conditional evidence and decision criteria
A rigorous test declares the boundary and the output before calculating results. In an insurer exercise these include the legal-entity perimeter, the valuation date, currency, accounting and prudential basis, relevant reinsurance treaties, asset admissibility, policyholder behavior, and assumed settlement time. Outcomes can include losses, capital headroom, claims-paying cash flow, liquidity depletion, policyholder-service interruption, or recovery time. A scenario may have separate event, valuation, and multi-year projection horizons. A storm occurs on one date; claims settle over years; market liquidity may disappear immediately; regulatory own funds may be revalued at another time. Conflating these horizons produces apparent precision without comparability.
The test also requires a decision rule: a threshold, a ranking, a capital allocation choice, a contingency decision, or a question about model credibility. A regulatory threshold does not automatically define every insurer’s internal risk tolerance. Conversely, an internal tolerance indicator does not alter a statutory solvency requirement. The decision rule should be capable of changing under the result. Stress testing has practical value when it reveals a vulnerability, a robust strategy, or an assumption that requires investigation.
The Basel Committee’s 2018 stress-testing principles emphasize governance, methodology, documentation, and use of the results. These are banking-sector principles, so their direct regulatory scope concerns banks; their explanation of methodological discipline is useful as a cross-sector comparison. The insurer-specific regulatory authority remains the relevant insurance regime. This jurisdictional separation prevents the methodological analogy from being misrepresented as insurance law.
Historical and Hypothetical Scenarios
Historical replay and the counterfactual insurer
Historical scenarios ordinarily translate a selected event or sequence into today’s risk factors and financial exposures. This is a counterfactual calculation: how would the currently defined insurer have performed if selected past shocks had recurred? It requires selecting what is preserved from history. Replaying asset-price returns alone may ignore changed reserve sensitivity, investment duration, reinsurance terms, and hedging programs. Replaying catastrophe insured losses without adjusting property values, coverage, deductibles, geographic concentration, inflation, and claims handling may understate or overstate current loss.
The historical record is therefore evidence for constructing a scenario, while the scenario itself is a modeled transformation of current exposures. Historical return periods inferred for hazards are especially fragile when changes in climate, exposure location, vulnerability, and policy terms are ignored. The scientific record should preserve the historical observations, the adjustment formulas, and the counterfactual exposure basis separately. The distinction allows a reviewer to identify whether a surprising result comes from the historical event, the assumed current balance sheet, or the mapping between them.
An insurance scenario can also be generated around a historically unusual causal conjunction. A windstorm can affect insured property; simultaneously a regional infrastructure outage can delay adjustment and claims payments; demand surge can increase repair costs; reinsurer collateral requirements may tighten; securities may be liquidated into a stressed market. The insurer’s ultimate insured loss is only one part of the financial consequence. The scenario should specify which of these mechanisms are modeled, which are held fixed, and which are investigated qualitatively.
Designing hypothetical adverse states
A hypothetical scenario is built by choosing an adverse but structurally coherent state, evaluating how its components interact, and testing the insurer’s conditional response. One might combine elevated claims inflation, lower bond market values, adverse lapse behavior, and reduced collateral liquidity. The magnitude of each component requires justification: historical extremes, physical models, expert elicitation, regulatory design, exposure analysis, or a transparent sensitivity grid. No single technique makes the scenario universally “correct”; the evidence depends on the reason it was designed.
For insurance applications, an effective design dossier should answer a sequence of questions. Which initiating event or structural change is assumed? What transmission channels convert the event into claims, liabilities, assets, and cash flows? Are the drivers jointly plausible, including their timing and correlation? What mitigations operate in the scenario, and which could fail at the same time? Are management actions feasible at the stated speed and volume? Which thresholds and outputs change? Which conclusions persist if the assumptions are perturbed? The answers should appear in the technical specification, not only in an explanatory narrative written after results have been obtained.
The NIST risk-assessment guide is one useful source for structured threat-event, vulnerability, and impact thinking in information-system settings (NIST, 2012). Its scope is federal information-system and organizational risk assessment; it is used here to illuminate scenario decomposition for operational and cyber shocks, rather than as an insurance-capital rule. In insurance, common technology vendors, cloud dependencies, software defects, and service interruption can be modeled as initiating events with downstream policy, claims, investment, and operational effects. A scenario that assumes independent cyber incidents while the actual system shares a provider has omitted a potentially decisive common cause.
Scenario matrices and the meaning of a scenario ensemble
It is often more informative to evaluate a structured collection of scenarios than to enlarge a single worst case indefinitely. Consider a two-dimensional grid in which claim inflation is 0%, 10%, or 25% and the fraction of a catastrophe portfolio affected by a common event is 5%, 15%, or 30%. The nine outcomes reveal interaction, changes in rankings, and nonlinear thresholds. A grid is a finite experimental design, not an estimated joint probability distribution. Weighting each cell equally would introduce a probability assumption that the scenario construction itself has not established.
Where probabilities are justified, a stochastic ensemble can simulate losses under an explicit probability law Pθ. The sample may represent expected distributional behavior under Pθ, including tails, conditional events, and uncertainty quantified by the model. A coherent scenario library can also be used without probability weights to study robustness over a credible set of adverse possibilities. An equal-weighted arithmetic average of intentionally selected severe scenarios should never be presented as an unbiased expected loss estimate. The analyst should instead report scenario-specific outcomes, a range, and the conditions under which the comparison holds.
The mathematical basis for simulation is developed in Glasserman’s Monte Carlo methods, while McNeil, Frey, and Embrechts provide a broad account of quantitative risk management and multivariate loss models (Glasserman, 2003; McNeil et al., 2015). The works support careful estimation and aggregation under specified models; neither a simulation engine nor a large number of generated draws validates an incorrectly specified loss law.
Empirical supervisory experiment: European insurers in 2024
EIOPA’s 2024 exercise provides a concrete case in which combined shocks and management actions change the interpretation of resilience. Its published 2024 Insurance Stress Test report examined adverse consequences of intensified or prolonged geopolitical tensions, including macroeconomic and financial-market shocks (EIOPA, 2024). The aggregate solvency ratio declined from 221.8% at baseline to 123.3% under the fixed-balance-sheet stress; specified reactive management actions raised the post-stress ratio to 139.9%. The 98.5-percentage-point initial decline and 16.6-point subsequent improvement are arithmetic differences between reported aggregate ratios, not estimates of annual failure probability. The report records eight participants below the 100% requirement before reactive management actions, all of whom reported actions that restored their positions above that threshold in the exercise.
The same report reveals why capital and liquidity require separate treatment. EIOPA reported a significant stressed aggregate cash-and-equivalents liquidity shortfall before reactive actions, while other liquid assets remained available to meet the specified liquidity demands. A headline solvency ratio alone would omit the need to sell assets, meet surrenders, and settle cash outflows. The exercise is evidence about the selected firms, scenario, balance-sheet methodology, and assumptions; its results are not a universal statement about every insurer or every geopolitical event.
The EIOPA stress-test overview describes the exercise’s policy objective as assessing resilience and improving supervisory understanding, not as predicting the exact future. This distinction is substantive. A realized event may combine different shocks, financial responses, market conditions, and timelines. Test design determines what the results can support.
Dependence and Common Shocks
Aggregate loss mathematics
Let individual insured loss components be L₁, …, Lₙ, and total gross loss be S = Σᵢ Lᵢ. When means and second moments exist, E[S] = Σᵢ E[Lᵢ] and Var(S) = Σᵢ Var(Lᵢ) + 2Σᵢ<ⱼ Cov(Lᵢ,Lⱼ). The expectation identity holds regardless of dependence. The variance identity shows that covariance can materially change the dispersion of aggregate losses even when every marginal loss distribution and the expected aggregate loss are unchanged. For heavy-tailed distributions with infinite variance, the variance formula is unavailable, but the need to specify dependence remains.
Insurance concentration makes common shocks particularly important. Two properties may have similar insured values but vastly different joint exposure depending on their geographic proximity, construction features, flood defenses, and claims inflation. Two cyber policies may appear distributed across policyholders while depending on the same infrastructure provider. Several reinsurance treaties may reference counterparties whose financial condition deteriorates together with claims. The appropriate dependence object can be a latent common hazard, shared market factor, correlated vulnerability, network path, conditional intensity process, or a complete joint loss distribution.
Pearson correlation describes linear dependence under the moments required for its definition. It cannot, by itself, determine simultaneous extreme probabilities or the shape of nonlinear dependence. Embrechts, McNeil, and Straumann explain these limitations in Correlation and Dependence in Risk Management (2002). In scenario evaluation, the consequence is practical: a calibrated correlation matrix may be consistent with more than one joint tail behavior. Separate scenarios can probe alternative common-cause structures without pretending that a particular tail dependence coefficient is fully identified from sparse data.
An exact two-risk counterexample
Consider two identical Bernoulli catastrophe losses X and Y, each equal to 100 monetary units with probability p = 0.01 and zero otherwise. Each expected loss is one unit, so total expected loss E[X + Y] is two units under every dependence structure compatible with these marginals. The individual distributions and mean aggregate loss are fixed. Only joint occurrence changes.
Under independence, P(X = 100, Y = 100) = 0.0001. The aggregate loss S has probabilities P(S = 0) = 0.9801, P(S = 100) = 0.0198, and P(S = 200) = 0.0001. Under perfect positive common-shock dependence, X = Y almost surely, so P(S = 0) = 0.99 and P(S = 200) = 0.01. Both models satisfy E[S] = 2. Their tail distributions differ sharply.
Define the loss quantile VaRα(S) as inf{x : P(S ≤ x) ≥ α}. At α = 0.995, the independent model gives VaR = 100, while the common-shock model gives VaR = 200. Under the standard quantile-integral definition ESα(S) = (1/(1 − α)) ∫ from α to 1 of VaRᵤ(S) du, the independent model gives ES₀.₉₉₅ = (0.0049 × 100 + 0.0001 × 200)/0.005 = 102. The common-shock model gives ES₀.₉₉₅ = 200. The discrete probability masses are treated according to the integral definition; ES is not computed by conditioning only on losses strictly greater than VaR.
| Quantity | Independent events | Perfect common shock |
|---|---|---|
| Probability of both losses | 0.01% | 1.00% |
| Expected total loss | 2 | 2 |
| 99.5% Value at Risk | 100 | 200 |
| 99.5% Expected Shortfall | 102 | 200 |
The table proves a precise and limited point: equal marginal probabilities and equal total expected loss do not determine joint tail risk. The difference is created entirely by the dependence assumption. It is an illustrative mathematical construction, not an insurer exposure estimate or a calibrated annual catastrophe model. Its scientific value is that every number can be reproduced from the stated probabilities.
The distinction between quantiles and expected tail loss has an extensive literature. Rockafellar and Uryasev’s conditional Value-at-Risk optimization gives a tractable treatment of a tail-risk objective, while Artzner and colleagues’ coherent measures of risk establish axiomatic requirements for risk aggregation (Rockafellar and Uryasev, 2000; Artzner et al., 1999). Their work does not imply that any scenario chosen by an analyst has a known probability. The counterexample demonstrates why tail-risk results should be assessed under alternative plausible joint structures.
From common cause to cascade
Dependence can arise in multiple causal forms. A common-cause event directly affects many insured exposures, such as a regional flood or shared digital service failure. A cascade propagates through an intermediate system, such as electricity outages that affect supply chains, produce business interruption, and delay emergency response. Financial contagion may operate through reinsurance recoverables, funding constraints, asset sales, or reduced market liquidity. Each mechanism has different timing and intervention points. A stress specification should therefore identify edges and sequences rather than simply choosing a single high correlation coefficient.
An insurer’s retained loss is a nonlinear transformation of gross losses through policy limits, deductibles, attachment points, reinstatements, aggregate caps, exclusions, and reinsurance protections. If net loss is N = g(S; contract terms), then E[N], VaR(N), and liquidity needs may change discontinuously as an attachment point is crossed. A stress that increases gross claims by a fixed percentage cannot automatically increase net claims by the same percentage. A reinsurer’s nonpayment or a contractual coverage dispute can alter both net ultimate loss and the timing of cash receipts. These factors require explicit treaty and counterparty data in a real application.
The IPCC’s assessment of compound and cascading risks identifies interacting physical and socioeconomic processes and cross-sector effects (IPCC, 2022). The relevance to insurance is structural: hazard, exposure, vulnerability, supply chains, and financial response can move together. The scientific assessment does not provide a universal insurance-loss multiplier for these interactions. The multiplier must be constructed and validated for the specific portfolio and scenario.
Model distribution versus scenario consistency
Copula models, factor models, conditional event trees, and network simulations can represent different dependence structures. They cannot substitute for a clearly stated risk ontology: which event affects which exposure through which mechanism. In a statistically fitted model, the dependence parameters can be estimated from relevant observations, with uncertainty and stability diagnostics. In a hypothetical scenario, dependence may be deliberately imposed to test a vulnerability. Both approaches are legitimate if their evidentiary status is visible.
A robust evaluation often reports at least three dependency cases: a calibrated baseline, a common-factor stress, and a structural-break or contagion case. The cases are chosen to reveal the sensitivity of tail and balance-sheet outcomes to dependence, not to force a single worst-case number. The important comparison is how the ranking of risks and protective actions changes across these plausible mechanisms. Inference should remain bounded by the assumptions actually tested.
Balance-Sheet Consequences
A financial transmission system
An insurer transforms event losses into financial consequences through two linked processes. First, changes in claims, expenses, discount rates, and policyholder behavior alter liabilities. Second, changes in market prices, credit quality, cash flows, and counterparty performance alter assets and available funds. Capital adequacy is a relationship between eligible financial resources and the applicable capital requirement. Liquidity is the ability to meet cash obligations at the required times. These related quantities respond differently to adverse events and cannot be collapsed into one number.
At a simplified economic-balance-sheet level, own funds may be represented as OF = A − L, where A is the value of assets and L is the value of liabilities under a specified valuation basis. A regulatory coverage ratio may be written C = eligible own funds / solvency capital requirement, with the numerator and denominator measured according to the applicable jurisdictional rules. This is an explanatory identity, not an attempt to replace the full prudential valuation, eligibility, tiering, tax, matching-adjustment, or transitional-measure rules of any jurisdiction.
A usable scenario maps each stressed driver into asset values, liability values, recoverable amounts, and required capital. A fall in bond prices does not always have the same net own-funds effect because discount-rate changes can also alter liabilities. An increase in claims inflation may affect reserves and future premiums, while a lapse shock can alter liquidity and liability duration. A credit downgrade can change assets, capital charges, and contractual collateral requirements simultaneously. The result depends on the regulatory framework, asset-liability matching, product mix, and the legal entity’s actual exposures.
The EU framework’s Solvency II Article 45 requires insurers and reinsurers within its jurisdictional scope to undertake an own risk and solvency assessment addressing their specific risks and solvency needs (European Union, 2009). The IAIS insurance ERM guidance explains the relationship among risk appetite, limits, capital adequacy, and prospective assessment (IAIS, 2024). U.S. ORSA guidance is organized under state law based on NAIC Model #505; the NAIC’s enterprise risk management overview documents the role of current and prospective solvency assessment and its recent updates (NAIC, 2026). These are jurisdictionally distinct frameworks, not identical quantitative tests.
An exact illustrative solvency calculation
Consider a hypothetical insurer with economic assets of 1,250 monetary units, liabilities of 1,000, and eligible own funds of 250. For demonstration alone, assume the applicable capital requirement is 130 and that the 250 units are fully eligible. Baseline coverage is 250/130 = 192.31%. A combined market and underwriting stress reduces assets by 120 and increases liabilities by 35; suppose the capital requirement simultaneously rises to 155. Post-stress assets are 1,130, liabilities 1,035, and own funds 95. Coverage is 95/155 = 61.29%. The example demonstrates an adverse change in both components of the ratio, which makes the decline substantially larger than a measure based on the change in assets alone.
| Balance-sheet indicator | Baseline | Combined stress |
|---|---|---|
| Assets | 1,250 | 1,130 |
| Liabilities | 1,000 | 1,035 |
| Eligible own funds (assumed) | 250 | 95 |
| Capital requirement (assumed) | 130 | 155 |
| Coverage ratio | 192.31% | 61.29% |
Suppose a shareholder then supplies 60 units of new eligible capital and a separately justified risk-reduction action reduces the capital requirement by 20 units, with no other costs or adjustments. Own funds become 155, the capital requirement becomes 135, and coverage reaches 114.81%. This third result is conditional on capital actually being available and the capital-reduction mechanism being allowable and effective. It is an explicit scenario extension, not an assumption that the action will be feasible in a systemic crisis. If raising capital is prohibited by a supervisory test specification, this response must be excluded from that exercise.
These numbers illustrate why management actions need individual evidence. Selling liquid assets to produce cash ordinarily changes asset composition rather than creating own funds, and transaction costs can diminish own funds. Transferring liabilities via reinsurance can reduce retained risk but introduces counterparty dependence, premiums, collateral conditions, and possible basis risk. A dividend suspension can preserve future resources but does not necessarily recover capital already lost. Claims-handling capacity and communication with policyholders may be operational constraints that a capital-ratio calculation misses.
The independent liquidity dimension
Let beginning cash be B, stressed cash inflows received within a fixed horizon be I, contractual and behavioral outflows be O, and readily monetizable assets after valid haircuts be H. A deliberately simplified horizon liquidity position is Q = B + I + H − O. A negative Q indicates a modeled shortfall under those assumptions. The formula is useful for bookkeeping, but the full dynamic assessment must respect the sequence of payments, settlement dates, margin calls, forced-sale losses, legal restrictions, and currency mismatches. An end-of-quarter positive total does not guarantee sufficient intraday or weekly cash.
Take a hypothetical beginning cash balance of 30, expected near-term inflows of 20, outflows of 90, and assets with a nominal sale value of 50 subject to a 20% haircut. The stress liquidity position is Q = 30 + 20 + 40 − 90 = 0. The insurer just meets the aggregate horizon requirement in this stylized case. If only 25 of the 40 monetizable units can settle before the largest payment date, the position at that date is negative by 15 unless other cash is available. The horizon and settlement assumption therefore reverse a reassuring conclusion without changing the total asset balance.
Liquidity and solvency interact through forced asset sales, collateral obligations, and management responses. A highly capitalized insurer may encounter near-term cash stress; a lower-capital firm may remain liquid for a time. The EIOPA 2024 liquidity findings show the importance of distinguishing cash holdings from other liquid assets and including behavioral outflows. Their reported sample-wide results concern the specific EIOPA exercise and should not be substituted for the liquidity position of an individual insurance entity.
UK life insurance evidence in 2025
The Bank of England’s Life Insurance Stress Test 2025 examined eleven major UK life insurers active in the bulk purchase annuity market under the Solvency UK framework (Bank of England, 2025). The aggregate Solvency Capital Requirement coverage ratio moved from 185% to 154% under the core financial-market scenario; the reported aggregate capital surplus above requirements fell by £8.6 billion. The exercise also examined concentration and funded reinsurance exposures. These are measurements of the participating firms at the specified valuation date under the specified scenario, including permitted standardized actions.
The scenario’s three-stage structure illustrates the significance of time and valuation mechanisms: initial repricing, subsequent credit downgrades and defaults, and stabilization with prescribed responses. The Bank of England explains that changes in credit quality affect matching-adjustment portfolios, eligible own funds, and capital requirements. The result provides insight into a business model in which long-duration liabilities are supported by investment portfolios and matching arrangements. It does not supply a general rule for property/casualty insurers, other jurisdictions, or alternative monetary environments.
The UK publication also provides a caution about exploratory exercises. The concentration analysis considered shocks to selected asset classes under its stated correlation assumptions, and the PRA itself discussed the limitations of excluding broader correlations among related assets. The funded-reinsurance exploratory result depends on recapture, collateral composition, and valuation assumptions. Proper research reproduces these constraints in interpretation. It is scientifically inappropriate to compare 154% in the UK exercise directly to 123.3% in EIOPA’s European exercise as though both were estimates on one common population and shock distribution.
Reverse Stress and Breakpoints
Beginning with an adverse outcome
Conventional forward stress analysis specifies shocks and calculates consequences. Reverse stress analysis specifies a loss of viability or threshold and identifies conditions that would generate it. The scientific object can be regulatory capital breach, economic insolvency, inability to pay a near-term claim, exhaustion of a reinsurance layer, failure of essential operations, or breach of a risk appetite indicator. A reverse-test result is meaningful only when the failure criterion, available protective actions, and search space are given.
Let g(s) denote a failure measure under scenario s, and let g(s) ≤ 0 represent a defined breach. A reverse-stress search can be written as finding s in an admissible set S such that g(s) ≤ 0. If the investigator minimizes a distance d(s,s₀) from a baseline to the breach, the meaning of the word “nearest” depends on that chosen distance. A distance in historical standard deviations is not equivalent to physical severity, joint annual frequency, or causal credibility. Optimizing without regard to physically or contractually possible combinations can identify mathematically nearby but implausible states. The solution must be interpreted as conditional on the feasible set and objective.
The IAIS reverse stress testing guidance explicitly connects reverse stresses to circumstances in which an insurer may fail or move below a predefined financial level and highlights management actions to reduce that risk (IAIS, 2024). The function of such analysis is diagnostic: it exposes points where recovery, contingent financing, risk transfer, operations, or governance may become inadequate. It does not convert every mathematically found break point into a likely future event.
A solvency breakpoint derivation
Return to an independent hypothetical baseline with eligible own funds of 250 and a capital requirement of 130. Define a continuous adverse loss x that reduces eligible own funds by x while the capital requirement rises by 0.1x. Then OF(x) = 250 − x, SCR(x) = 130 + 0.1x, and C(x) = (250 − x)/(130 + 0.1x). The capital coverage threshold C(x) = 1 is reached when 250 − x = 130 + 0.1x. Hence x = 120/1.1 = 109.09 monetary units to two decimal places.
Economic own funds become zero only at x = 250 in the same simplified construction. Therefore the capital requirement can be breached substantially before economic net assets reach zero. In a real jurisdiction, further regulatory thresholds, eligibility constraints, tax and reserve adjustments, and intervention levels may apply. This derivation is an explanatory algebraic experiment, not a representation of an actual insurer’s prescribed solvency formula.
The breakpoint changes if the model permits actions. Suppose a valid prospective action adds m units to own funds before breach, at a cost or probability that must be modeled. Then the boundary is 250 + m − x = 130 + 0.1x, or x* = (120 + m)/1.1, only if the action is reliably available, immediately eligible, and does not change the capital requirement or other terms. A delayed action that arrives after a liquidity failure cannot be credited in the earlier time step. The mathematical lesson is that recovery capacity must be specified in the state transition, not inserted as a retrospective adjustment to a stress result.
Multiple failure channels
The most informative reverse scenarios may involve several inequalities. A simplified insurer could require C_t ≥ 1 at designated dates, cash Q_t ≥ 0 at all payment times, and the availability of an essential operational service. The earliest breach may arise from a collateral call before the capital balance falls below its solvency threshold. A common reinsurer failure may produce both new retained claims and slower recoveries. A catastrophe may create exceptional claim volume, while an investment-market shock reduces the ability to realize cash without loss.
A reverse analysis can enumerate single-cause, combined-cause, and cascading failure paths. These should be reported as conditional pathways with an explanation of the constraints that make each path credible. Introducing probabilities is possible only after a defensible event-frequency and dependence model has been specified. A ranking by minimum severity distance is distinct from ranking by estimated frequency, expected loss, or effectiveness of possible intervention.
Breakpoints as a guide to resilience decisions
The analytical value of reverse testing emerges when it identifies an actionable constraint: insufficient rapid liquidity, overly concentrated counterparties, fragile reinsurance attachments, operational recovery delays, or a capital buffer sensitive to correlated risks. The result can motivate collateral planning, exposure reduction, renegotiation of treaty terms, diversification, improved liquidity reserves, or crisis rehearsal. Whether an intervention is efficient requires a separate decision analysis accounting for cost, policyholder interests, constraints, and the range of scenarios in which it is effective.
The Bank of England’s insurance stress-testing program illustrates how public supervisory exercises and their underlying specifications are revised across cycles. An insurer’s internal reverse tests should likewise be reviewed when exposures, products, financial markets, or dependencies change. A static map of prior failure paths becomes progressively less informative as the business and hazard environment evolve.
Interpretation and Limitations
What a result establishes
A scientifically sound stress result has the form: under the explicitly described scenario, valuation basis, model, exposure set, horizon, and behavioral assumptions, the defined outcome has a calculated value or range. Its strength rests on traceability of inputs to observed or justified facts, explicit modeling of structural links, mathematical correctness, and the sensitivity of conclusions to credible alternatives. A numerical result expressed to two decimal places may be reproducible yet still be epistemically uncertain because model structure or future conditions are uncertain. Reproducibility and predictive validity are different properties.
Three complementary evidentiary tests are especially important. First, internal consistency asks whether the specified shocks can coexist and whether timing, accounting, and mathematical operations agree. Second, external grounding asks how the exposures, contract structures, historical observations, and physical or institutional processes have been verified. Third, decision robustness asks whether the conclusion changes when plausible alternative models, dependencies, and management constraints are used. A stress analysis can pass one test and fail another. For example, a mathematically correct simulation may rely on a reinsurance recovery assumption contradicted by treaty terms or counterparty conditions.
Model risk includes misspecified marginal loss distributions, unstable dependence, unobserved common factors, inadequate event catalogs, flawed valuation, treatment of policyholder behavior, and numerical error. Expert judgment is often indispensable, but its use should be separated from measured quantities and made reviewable. Statistical intervals derived under one model do not encompass all possible structural uncertainty. The quantitative risk-management literature emphasizes tail behavior, dependence, and aggregation, while the simulation literature shows how computational sampling error can be assessed under a chosen model (McNeil et al., 2015; Glasserman, 2003). Neither source removes the need for scientific judgment about model adequacy.
Tail estimation, back-testing, and scarce events
Rare events create a severe observational problem. An annual event probability near one in several hundred years is difficult to estimate reliably from a few decades of relevant and stable data. Selecting a distribution family or extrapolating extreme quantiles can be necessary but produces sensitivity to tail assumptions and thresholds. When processes are nonstationary, past observations may estimate a mixture of regimes rather than the distribution relevant to the next planning horizon.
Back-testing can compare earlier predictions with subsequent observations and reveal calibration defects, but a short record of extreme events provides limited power to discriminate among rare-tail models. Scenario testing complements this evidence by asking whether specified structural vulnerabilities have been missed. It cannot turn an unobserved scenario into an empirically verified likelihood. A transparent report preserves the distinction among data uncertainty, parameter uncertainty, model uncertainty, and uncertainty about future scenarios.
The IPCC climate-risk synthesis provides a clear example of why shifting hazards and exposures demand long-horizon, conditional assessment (IPCC, 2023). EIOPA’s climate-scenario ORSA guidance illustrates the application of climate materiality and scenario reasoning within an insurer’s own solvency assessment (EIOPA, 2022). The guidance is methodological and supervisory; company-level financial results require firm-specific exposures, assumptions, and calculations.
Causal identification and behavioral feedback
Some stress pathways are mechanical under given contracts and balances; others depend on uncertain behavior. Policyholder surrender, claims reporting, repair-price demand surge, counterparty collateral requests, and investment-market depth may change because the adverse event occurs. Multiple firms taking similar management actions can alter the very market prices and liquidity conditions used in the test. A scenario based on fixed prices during large-scale liquidation may fail to represent the costs generated by the response itself.
A rigorous analysis differentiates exogenous shock assumptions from endogenous reactions. If the test assumes insurers collectively sell assets, the possibility of market impact and price feedback requires explicit treatment. If the test assumes prompt reinsurance recoveries, counterparty liquidity and settlement conditions should support that timing. If a scenario assumes a new capital raise, the test should specify whether shareholders and capital markets remain able and willing to supply it. These are causal questions that cannot be answered solely by increasing the number of simulation trials.
The EIOPA 2024 exercise is particularly instructive because it evaluated fixed-balance-sheet outcomes and permitted reactive management actions. The official results and recommendations demonstrate that the exercise design changes the meaning of the headline ratio (EIOPA, 2024). The UK PRA exercise used its own defined action set, reporting perimeter, and scenario progression. Results should therefore be compared on matched assumptions, not merely because both reports use the word stress.
A hierarchy of claims and limitations
A scenario calculation may support a conditional balance-sheet claim, a model-comparison claim, or a practical vulnerability claim. It supports a probability claim only with adequate probabilistic calibration. It supports a causal claim only when the mechanism has been specified and justified. It supports a normative recommendation only when the objective, constraints, and consequences for relevant stakeholders are articulated. Treating all four levels as the same scientific statement creates overreach.
For supervisory sources, the hierarchy also includes legal status. The Solvency II Directive is a legal instrument within its applicable jurisdiction; EIOPA materials include guidelines, exercises, application guidance, and recommendations with distinct regulatory roles. IAIS principles are international insurance supervisory standards whose national effect depends on jurisdictional implementation. NAIC material describes a model-act-based U.S. state system. BCBS principles are banking-specific. These differences determine how each source can be cited and what it can legitimately establish.
This monograph’s empirical scope is limited to public supervisory reports and institutional descriptions, with no access to confidential insurer-level stress submissions or proprietary catastrophe models. The European and UK numerical results are reported as published, without independent re-performance of participating insurers’ models. Scientific books and journal articles were checked against publisher records and accessible descriptions, while full proprietary contents were not claimed when unavailable. The original numerical examples are transparent mathematical constructions, not empirical findings. External peer review, regulatory certification, and proof of model accuracy beyond the stated checks are not claimed.
Synthesis for evaluation practice
The research supports several practical analytical conclusions. Stress scenarios should be specified jointly with the output and failure criterion they are intended to test. Dependence should be represented through identifiable shared mechanisms and evaluated under alternatives. Asset, liability, capital, and liquidity effects should be connected in one coherent time sequence. Management actions should be modeled as constrained, potentially costly decisions rather than guaranteed offsets. Probability labels should be reserved for probabilistically calibrated models, while conditional scenario results should retain their own explanatory status. Finally, decision relevance is established by showing which vulnerability or action changes under the scenario and how stable that conclusion remains when assumptions vary.
These conclusions synthesize established risk science and observed supervisory practice. They are ordinary analytical results, not proposals for a new theoretical system. The principal scientific contribution of this dossier lies in integrating the logical status of scenario evidence, exact dependence counterexamples, time-dependent financial constraints, and cross-jurisdictional supervisory observations into one reproducible account of stress-based evaluation.
Conclusion
Stress testing and scenario-based risk evaluation are formal methods for investigating how adverse configurations of circumstances affect risk-bearing systems. They require a precise scenario, a defined exposure and valuation boundary, an explicit mapping of shocks into consequences, and a stated decision criterion. A stress scenario is informative because its conditional implications can be examined, challenged, and compared with alternatives. Its severity and narrative coherence do not by themselves establish an unconditional probability.
The mathematical analysis demonstrates that dependence can transform joint tail consequences while preserving individual loss probabilities and aggregate expected loss. The two-risk construction yields an unchanged expected loss of two units but moves 99.5% Value at Risk from 100 to 200 and Expected Shortfall from 102 to 200 when independent rare events are replaced by a perfect common shock. The balance-sheet construction shows that simultaneous asset erosion, liability increases, and rising capital requirements can generate a much larger capital-coverage decline than any single balance-sheet component suggests. The reverse-stress calculation identifies a specified regulatory coverage breakpoint before the exhaustion of economic own funds. These are exact results under stated assumptions; they provide testable explanations of how structure and thresholds determine outcomes.
Public supervisory evidence confirms that the financial question is multidimensional. EIOPA’s 2024 exercise exposed substantial capital and liquidity consequences under a geopolitical scenario and demonstrated the difference between fixed-balance-sheet and permitted reactive-action results. The UK PRA’s 2025 life insurance test revealed the significance of asset downgrades, liability valuation, and the prescribed recovery sequence for a particular life-insurance population. IAIS standards, EU ORSA requirements, U.S. NAIC guidance, and other sources place such conditional analysis within wider risk governance while retaining distinct jurisdictions and methodological purposes.
The resulting conclusion is institutionally and mathematically grounded: the quality of a stress-test conclusion depends on consistency of the imagined event, completeness of the response system, validity of the source evidence, and robustness to plausible alternatives. Scientific interpretation should identify what has been proven by calculation, what has been observed in real exercises, what remains a model-dependent assumption, and which conclusions could guide decisions under uncertainty. That disciplined separation makes stress testing a durable method of risk evaluation across insurance products, financial systems, and emerging sources of interconnected loss.
References
Artzner, P., Delbaen, F., Eber, J.-M., & Heath, D. (1999). Coherent measures of risk. Mathematical Finance, 9(3), 203–228. https://doi.org/10.1111/1467-9965.00068
Bank of England, Prudential Regulation Authority. (2025, November 17). Life Insurance Stress Test: 2025 results. https://www.bankofengland.co.uk/prudential-regulation/publication/2025/november/list-2025-results-report
Bank of England. (2025). Insurance stress testing: Insurers (official program and methodology resources). https://www.bankofengland.co.uk/stress-testing/insurers
Basel Committee on Banking Supervision. (2018, October 17). Stress testing principles. Bank for International Settlements. https://www.bis.org/publications/201810-guidelines-stress-testing-principles
Embrechts, P., McNeil, A. J., & Straumann, D. (2002). Correlation and dependence in risk management: Properties and pitfalls. In M. A. H. Dempster (Ed.), Risk Management: Value at Risk and Beyond (pp. 176–223). Cambridge University Press. https://doi.org/10.1017/CBO9780511615337.008
European Insurance and Occupational Pensions Authority. (2015). Guidelines on Own Risk Solvency Assessment (ORSA). https://www.eiopa.europa.eu/publications/guidelines-own-risk-solvency-assessment-orsa_en
European Insurance and Occupational Pensions Authority. (2022). Application guidance on running climate change materiality assessment and using climate change scenarios in the ORSA. https://www.eiopa.europa.eu/system/files/2022-08/application_guidance_on_running_climate_change_materiality_assessment_and_using_climate_change_scenarios_in_the_orsa_0.pdf
European Insurance and Occupational Pensions Authority. (2024, December 17). 2024 Insurance Stress Test: Report (EIOPA-BoS-24/553). https://www.eiopa.europa.eu/document/download/f8a234b0-a84a-49ff-975e-c47f8849bfc0_en?filename=Report+-+Insurance+Stress+Test+2024.pdf
European Insurance and Occupational Pensions Authority. (2024). Insurance stress test 2024 (program, findings, and recommendations). https://www.eiopa.europa.eu/insurance-stress-test-2024_en
European Union. (2009). Directive 2009/138/EC on the taking-up and pursuit of the business of insurance and reinsurance (Solvency II), Article 45, as amended. EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/ALL/?uri=CELEX%3A32009L0138
Glasserman, P. (2003). Monte Carlo Methods in Financial Engineering. Springer. https://doi.org/10.1007/978-0-387-21617-1
Intergovernmental Panel on Climate Change. (2022). Climate Change 2022: Impacts, Adaptation and Vulnerability, Working Group II, Summary for Policymakers. https://www.ipcc.ch/report/ar6/wg2/chapter/summary-for-policymakers/
Intergovernmental Panel on Climate Change. (2023). Climate Change 2023: Synthesis Report, Summary for Policymakers. https://www.ipcc.ch/report/ar6/syr/summary-for-policymakers/
International Association of Insurance Supervisors. (2024, December). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups (ICP 16). https://www.iais.org/uploads/2025/01/IAIS-ICPs-and-ComFrame-adopted-in-December-2024.pdf
International Organization for Standardization. (2018). ISO 31000:2018: Risk management—Guidelines (edition 2; reviewed and confirmed 2023). https://www.iso.org/standard/31000
McNeil, A. J., Frey, R., & Embrechts, P. (2015). Quantitative Risk Management: Concepts, Techniques and Tools (revised edition). Princeton University Press. https://press.princeton.edu/books/hardcover/9780691166278/quantitative-risk-management
National Association of Insurance Commissioners. (2026). Enterprise Risk Management (NAIC insurance topics; ORSA guidance history and 2026 updates). https://content.naic.org/insurance-topics/enterprise-risk-management
National Institute of Standards and Technology. (2012, September). Guide for Conducting Risk Assessments (NIST SP 800-30 Rev. 1). https://csrc.nist.gov/pubs/sp/800/30/r1/final
Rockafellar, R. T., & Uryasev, S. (2000). Optimization of conditional value-at-risk. The Journal of Risk, 2(3), 21–41. https://doi.org/10.21314/jor.2000.038