Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Risk Model Sensitivity and Robustness

Abstract. Risk model sensitivity concerns the dependence of an insurance risk calculation on its parameters, assumptions, distributional representation, data and computational implementation. Robustness concerns the stability and decision relevance of the resulting calculation over explicitly described perturbations. A complete assessment must distinguish uncertainty about the insured event from uncertainty about the method used to quantify that event. This monograph develops that distinction mathematically and operationally for insurance loss severity, claim frequency, compound loss distributions, tail risk measures, dependence and capital decisions. It compares local derivatives, finite differences, scenario changes, variance-based and distribution-oriented global sensitivity, contamination arguments, extreme-value methods and simulation diagnostics. Original analytic examples show that distributions having identical means and variances can yield substantially different high-quantile and expected-shortfall estimates; that an apparently unchanged quantile can conceal a large change in conditional tail loss; that dependence can reverse quantile-based diversification conclusions; and that modest modifications of generalized Pareto tail shape can produce disproportionate differences in tail capital. An explicit source-to-claim research approach integrates the foundational risk-measure literature, current actuarial work and selected European insurance model-validation requirements without confusing statutory compliance with scientific robustness. The central finding is that sensitivity is a conditional property of a specified output, perturbation and decision context. Robustness therefore requires a reproducible comparison of economically relevant alternatives, numerical uncertainty and model-implied consequences, with clear disclosure of what remains unidentified by observed data.

Keywords: Risk model sensitivity; model robustness; insurance loss distributions; parameter uncertainty; model uncertainty; Value at Risk; Expected Shortfall; extreme value theory; generalized Pareto distribution; global sensitivity analysis; dependence uncertainty; simulation error; actuarial validation; Solvency II.

Introduction

An insurer's modeled loss is the result of a sequence of choices: defining the risk event and economic exposure; constructing records of observed losses; specifying a frequency and severity mechanism; describing dependence across events or portfolios; estimating parameters; applying reinsurance and accounting rules; simulating or analytically deriving a loss distribution; and selecting an output such as expected loss, a quantile, Expected Shortfall or a solvency capital requirement. A calculation can be internally exact and nevertheless highly contingent on one of those choices. Sensitivity analysis makes that contingency explicit.

The foundational literature makes the problem precise. Artzner, Delbaen, Eber and Heath (1999) articulated the coherent risk-measure axioms, including subadditivity. Rockafellar and Uryasev (2000) developed an influential Conditional Value-at-Risk optimization treatment. Embrechts, Klüppelberg and Mikosch (1997) demonstrated the depth of extreme-event modeling needed for insurance and finance, while McNeil, Frey and Embrechts (2015) placed loss distributions and aggregation within an integrated quantitative-risk treatment. These works motivate the choice of measurable output; they do not establish that any particular fitted distribution is empirically correct.

For this dossier, a risk model is represented as a mapping

[ Y = G(\boldsymbol{\theta},\mathcal{M},\mathbf{X},D,A), ]

where (Y) is a random loss or a derived decision quantity, (\boldsymbol{\theta}) denotes numerical parameters, (\mathcal{M}) the structural and distributional specification, (\mathbf{X}) the underlying random drivers, (D) the observed data and (A) the chosen algorithm. A reported risk statistic is (R=\rho(Y)), possibly after capital, contract and accounting transformations. Each component has distinct admissible perturbations. Changes in (\theta) within a fixed distribution family measure parameter sensitivity; replacing the family measures specification sensitivity; altering the joint law measures dependence sensitivity; replacing a numerical procedure measures computational sensitivity; and varying exposure, deductibles or limits measures insurance-contract sensitivity. The findings depend on which elements are held fixed.

The present analysis treats robustness as an empirically auditable relation between an output and a declared set of plausible variations, with attention to the downstream decision. It does not attach a new named axiom or proprietary model to this interpretation. Statistical robustness, in the more technical sense, is separately connected to an estimator's reaction to data contamination. Cont, Deguest and Scandolo (2010) argue in their risk-measurement robustness analysis that evaluating the mathematical functional apart from its estimated procedure overlooks a crucial source of instability. That insight provides a guide to the separation of formulas, data and estimation throughout the chapters.

All numeric illustrations here are synthetic, dimensionless and independently recomputable from displayed formulas. They are not observations of a real insurance portfolio, licensed capital calculations, or forecasts. Article 124 of the European Solvency II framework requires a model-validation cycle that tests the sensitivity to key assumptions. This is a specific European legal setting. The general mathematical arguments apply across jurisdictions, products and supervisory arrangements, while regulatory conclusions require the applicable local legal instrument.

Parameter Sensitivity

Defined outputs and perturbation scales

For a deterministic output (r(\theta)), the local derivative (\partial r/\partial\theta_i) measures a first-order response at a specified point, conditional on the remaining parameters. It is meaningful only where differentiability exists and where the chosen change is plausible. A derivative measured in currency units per percentage-point change cannot be compared directly with a derivative measured in currency units per unit change in a dimensionless shape parameter. Where (r,\theta_i>0), the elasticity (\mathcal E_i=(\theta_i/r)(\partial r/\partial\theta_i)) provides a dimensionless local comparison. For signed or near-zero variables, a declared reference scale and absolute finite differences are often more interpretable.

The finite-difference experiment reports (\Delta_i r=r(\theta_i+\delta_i)-r(\theta_i)) together with (\delta_i), baseline assumptions and treatment of co-varying parameters. One-sided differences are appropriate at a contractual boundary, at zero frequencies, at shape-parameter restrictions and at discontinuities. A centered difference can conceal a sharp corner; an extremely small perturbation can amplify numerical rounding; an excessively large perturbation measures a different regime. The size and direction of the change are part of the result.

A stochastic output demands greater care. Differentiating the expectation (E[Y_\theta]) is generally different from differentiating its quantile or its tail expectation. An input can have little influence on expected loss yet dominate capital at a high percentile. The distinction is explicit in the actuarial risk-measure sensitivity approach of Tsanakas and Millossovich (2016), which links changes in uncertain inputs to changes in output risk measures.

Pareto severity as an exact local example

Let (X) follow a Pareto Type I severity law with minimum (x_m>0) and shape (\alpha>1):

[ P(X>x)=(x_m/x)^\alpha,\quad x\ge x_m. ]

For (0<p<1), the left quantile is (q_p=x_m(1-p)^{-1/\alpha}). Expected Shortfall for a continuous distribution is (ES_p=\alpha q_p/(\alpha-1)). Mean severity is (\alpha x_m/(\alpha-1)). The local semielasticity of the high quantile with respect to shape is [ \frac{\partial\log q_p}{\partial\alpha}=\frac{\log(1-p)}{\alpha^2}, ] and the corresponding Expected Shortfall log derivative is [ \frac{\partial\log ES_p}{\partial\alpha} =\frac{\log(1-p)}{\alpha^2}-\frac{1}{\alpha(\alpha-1)}. ] Both are negative. As the confidence level rises, the magnitude of the first term increases. As (\alpha) approaches one from above, the second term diverges in magnitude and the mean tail loss becomes particularly sensitive.

With (x_m=1), (p=.99), and (\alpha=2.5), the exact quantile is (6.309573) and Expected Shortfall is (10.515956). Changing only the shape to (2.3) gives (q_{.99}=7.405685) and (ES_{.99}=13.102365); changing it to (2.7) gives (5.504790) and (8.742901). The (-0.2) shape variation raises ES by approximately 24.6%, whereas the (+0.2) variation lowers it by approximately 16.9%. This asymmetry is a property of the nonlinear formula, not evidence of asymmetric data error. At this baseline, the ES semielasticity is approximately (-1.0035) per unit change in (\alpha), while its dimensionless shape elasticity is approximately (-2.5087). The local linearization becomes less accurate over the displayed finite changes.

The formula assumes (\alpha>1); an analyst may not extrapolate it into (\alpha\le1) as a finite expected value. It also assumes a Pareto law over the whole displayed support. The empirical question whether a fitted power law is justified requires separate statistical and tail diagnostics. The treatment of extreme insurance severities in Modelling Extremal Events is relevant precisely because extreme inference introduces threshold and extrapolation uncertainty.

Frequency, severity and coverage sensitivities

In a compound model (S=\sum_{j=1}^{N}X_j), assume independent and identically distributed severities, a count (N) independent of severities, finite severity second moment, and (N\sim\mathrm{Poisson}(\lambda)). Then

[ E[S]=\lambda E[X],\qquad \operatorname{Var}(S)=\lambda E[X^2]. ]

The derivative of expected aggregate loss with respect to (\lambda) is (E[X]). Relative sensitivity to frequency is one when severity is held fixed. Yet the influence of (\lambda) on a high quantile is indirect and may vary with the tail regime. For Pareto severity with (\alpha=2.5,x_m=1), (E[X]=5/3) and (E[X^2]=5). At (\lambda=100), aggregate mean is (166.666667), variance is (500), and standard deviation is (22.360680). The existence of these moments does not imply that the aggregate's far tail is approximately normal at the relevant solvency quantile.

For a per-occurrence deductible (d\ge x_m), the retained stop-loss amount ((X-d)+) has expectation [ E[(X-d)+]=\int_d^\infty P(X>x),dx =\frac{x_m^\alpha d^{1-\alpha}}{\alpha-1}. ] The derivative is (-P(X>d)). At (\alpha=2.5,x_m=1,d=5), the expected stop-loss amount is (0.0596285). Raising (d) to (5.5) reduces it to (0.0516850), or approximately 13.3%. These calculations describe one loss layer; they do not determine a policy premium without frequency, expenses, dependence, capital and contract terms. They also show that exposure and contract sensitivities can be more directly relevant to an underwriting decision than a generic mathematical derivative.

Global sensitivity and interactions

A local gradient is inadequate when inputs vary across a range, interact or trigger regime changes. Morris (1991) proposed elementary-effects screening designs for computational models. Sobol (2001) developed global sensitivity indices under a variance-decomposition approach; Saltelli and colleagues (2008) discuss the broader design principles. These methods examine an input domain rather than a single baseline.

For a transparent illustration let (U_1,U_2) be independent Uniform((-1,1)) drivers and (Y=U_1+U_2+U_1U_2). Since both inputs have mean zero and variance (1/3), orthogonality gives (\operatorname{Var}(Y)=1/3+1/3+1/9=7/9). The first-order Sobol indices are (S_1=S_2=(1/3)/(7/9)=3/7), while the second-order interaction contributes (1/7). Each total-effect index is (4/7). A one-factor-at-a-time account omitting the interaction therefore understates each factor's full contribution to output variance.

Variance decomposition is an analysis of variance. It does not by itself rank the drivers of a chosen .995 quantile or Expected Shortfall, and the conventional orthogonal interpretation assumes independent uncertain inputs with a stated sampling law. Dependence, epistemic ambiguity in input ranges and multimodality may defeat a naive application. Borgonovo's (2007) moment-independent importance measure addresses sensitivity of the output distribution beyond one moment. A useful comparison therefore reports a response-specific ranking rather than treating any sensitivity index as universal.

Structural Specification

What model structure actually determines

The structural specification determines which events can occur, which losses combine, which contracts respond and which dependencies matter. A model may be sensitive because a parameter is uncertain or because an entire mechanism is omitted. Frequency-severity independence, stationarity, exposure homogeneity, claims closure patterns, catastrophe clustering and copula selection are separate assumptions. Even perfect calibration of a model's finite parameter vector cannot compensate for a materially omitted mechanism.

A useful statistical decomposition distinguishes process variability from parameter estimation and specification. Conditional on model (\mathcal M) and parameter (\theta), (L) has a process distribution (P(L\mid\theta,\mathcal M)). The data induce uncertainty over (\theta), while competing defensible models induce uncertainty over (\mathcal M). For a Bayesian analysis, integrating over a parameter posterior yields a posterior predictive distribution, conditional on priors and likelihood choice. For a frequentist analysis, confidence intervals and bootstrap distributions describe repeated-sampling performance under specified assumptions. For scenario analysis, a collection of explicit alternative structures may be explored without assigning it probabilistic weights. Combining all three into a single apparently objective capital figure would hide methodological choices.

Compound distribution and dependence architecture

In the independent compound Poisson case the probability generating and Laplace transforms permit exact characterization where the moment-generating objects exist. For nonnegative severities, [ E[e^{-tS}]=\exp{\lambda(E[e^{-tX}]-1)},\quad t\ge0. ] This identity depends on independent Poisson counts and independent identically distributed marks. Common shocks, contagion, catastrophe clustering and reporting lags call for different specifications. Uncertain catastrophe occurrence may be correlated with claim severity and inflation, so independent count/severity sensitivity tests may miss material joint effects.

Within a portfolio, the margins do not identify the joint distribution. With fixed individual loss laws (F_1,\ldots,F_k), different couplings alter the sum (S). A scalar correlation coefficient does not determine joint extremes. McNeil, Frey and Embrechts (2015) analyze risk dependence and aggregation as a substantive modeling problem. An appropriately defined sensitivity experiment holds the marginal laws and exposure totals fixed while changing a defensible family of joint laws. A separate experiment changes the margins and dependence together to assess plausible economic stresses; the two should not be mislabeled as the same test.

A discrete dependence counterexample

Consider two identical insurance exposures, each paying (100) with probability (0.01) and zero otherwise. For each individual loss (X), (P(X=0)=.99), hence the left .99 quantile (VaR_{.99}(X)=0). Under independence, their sum (S) has (P(S=0)=.9801), (P(S=100)=.0198), and (P(S=200)=.0001); therefore (VaR_{.99}(S)=100). The summed quantile exceeds the sum of individual quantiles, illustrating the failure of subadditivity for this VaR definition.

Under perfect positive dependence, losses occur together: (P(S=0)=.99), (P(S=200)=.01), and (VaR_{.99}(S)=0). The VaR at exactly .99 has declined despite more concentrated tail events. Expected Shortfall defined by the quantile integral [ ES_p(S)=\frac{1}{1-p}\int_p^1VaR_u(S),du ] equals (101) for the independent sum and (200) for the perfectly dependent sum. Under the same convention, each standalone Expected Shortfall equals (100). The Expected Shortfall calculation distinguishes the damage beyond the quantile threshold even where the VaR boundary behaves counterintuitively.

This construction is exact and makes an important qualification visible: a capital comparison based on a single high quantile can change discontinuously when probability mass crosses a threshold. The choice of quantile convention for discrete distributions, the treatment of atoms and the tail probability must be stated. Artzner and colleagues' risk-measure analysis and Rockafellar and Uryasev's tail-risk optimization framework provide the relevant foundational context. Neither is a substitute for validating the joint loss model itself.

Insurance-specific model structures

Claims reserves are sensitive to the developmental shape of payment and reporting. Mack (1993) derived chain-ladder reserve standard errors under an explicit distribution-free moment framework; that result is conditional on the chain-ladder assumptions and should not be interpreted as protection against every structural change in claims handling, inflation or claims emergence. Reserve uncertainty, including claims development and parameter estimation, differs from the sensitivity of a forward-looking catastrophe or cyber portfolio.

Peters and colleagues (2023) examine cyber loss model risk, identifying parameter and model uncertainty as relevant to premium mispricing. The external data, heavy-tailed severities, technological change and loss aggregation characteristic of cyber portfolios make the structure itself a material object of sensitivity. This literature supports considering pricing consequences, while any asserted premium distortion for a specific insurer would require that insurer's own data, coverage and claims definitions.

Structural checks must also follow legal contract architecture: occurrence versus claims-made triggers, attachment and exhaustion of reinsurance, aggregate limits, reinstatements, event definitions, waiting periods, indemnity or parametric payout formulas, and inflation indexing. A change in exposure definition can alter an insurer's modeled risk even when all stochastic parameters remain numerically fixed. These contract transformations belong inside the model to be stressed, with a documented source for the contract terms when analyzing a real product.

Distributional Alternatives

Why equal moments are insufficient

A finite sample generally cannot identify the far tail uniquely. Multiple fitted families can closely reproduce the bulk of observed losses, mean and variance yet disagree at a one-in-100 or one-in-1,000 event. A model comparison confined to likelihood, AIC or an aggregate goodness-of-fit statistic may privilege the numerous small claims, whereas a solvency decision is governed by a thin tail. Distributional analysis therefore needs explicit evaluation of the region the model is used to predict.

A precise synthetic comparison illustrates the point. Model P is Pareto Type I with shape (\alpha=2.5) and (x_m=1). Its mean is (5/3) and variance is (20/9). Model L is lognormal with [ \sigma^2=\log(1.8)=0.5877867,\qquad \mu=\log(5/3)-\sigma^2/2=0.2169323, ] so that its mean and variance equal those of Model P exactly. Both laws are positive and continuous. Their matched first two moments do not match their tail probabilities, quantiles or expected-shortfall values.

Tail probability level Pareto VaR Pareto ES Lognormal VaR Lognormal ES
.950 3.314454 5.524090 4.384152 6.330757
.990 6.309573 10.515956 7.392597 9.903050
.999 15.848932 26.414887 13.278263 16.790914

The lognormal is more severe at the .95 quantile and .95 Expected Shortfall, and at the .99 quantile, yet the Pareto has greater .99 and .999 Expected Shortfall. By .999 the Pareto is larger on both measures. The crossing is derived from two explicitly specified laws and provides no empirical claim about which law an insurer should select. It shows that an assessment of distributional alternatives must name the risk measure and percentile before using the word robust.

For the Pareto, (q_p=(1-p)^{-1/2.5}) and (ES_p=(5/3)q_p). For the lognormal, (q_p=\exp(\mu+\sigma\Phi^{-1}(p))) and [ ES_p=\exp(\mu+\sigma^2/2), \frac{\Phi(\sigma-\Phi^{-1}(p))}{1-p}. ] These formulas make the table reproducible without fitted datasets or random seeds. Reporting only mean and standard deviation would leave the materially different tail behavior concealed.

Alternative tails and threshold decisions

Extreme-value theory supplies a principled asymptotic basis for modeling sufficiently high exceedances under suitable regularity assumptions. Balkema and de Haan (1974) studied residual-life asymptotics; Pickands (1975) addressed extreme-order-statistic inference. The generalized Pareto approximation for exceedances over a high threshold is useful, but its asymptotic justification does not turn every finite threshold and dataset into a well-specified Pareto tail.

The practical threshold problem is a bias–variance tradeoff. A low threshold admits more observations but may violate the approximation; a high threshold improves asymptotic plausibility but leaves few exceedances and volatile estimates. Threshold stability plots, parameter uncertainty, mean residual life diagnostics and alternative body–tail splice choices provide different kinds of evidence. Instability across defensible thresholds is a substantive finding, even when one selected model happens to have a favorable fit statistic.

Distributional families should be selected with the economic mechanism in view. A lognormal or gamma model may capture moderately skewed claim severity; Pareto-type regular variation can represent heavier extremes; mixtures can accommodate subpopulations; finite policy limits truncate insured payouts even when ground-up claims are unbounded. A catastrophe model may be constructed from hazard, vulnerability and exposure modules rather than a single severity family. Sensitivity must preserve the actual link between peril, location, insured value and contractual response.

Data contamination and functional robustness

Let the baseline loss be exactly (1), and let a contaminated loss take value (200) with probability (\varepsilon=0.005) and (1) otherwise. The left .99 quantile remains (1): 99.5% of outcomes are at the baseline value. Expected Shortfall at .99, defined through the upper one-percent quantile integral, is (100.5), because half of the upper 1% consists of losses of (200). The baseline Expected Shortfall is (1). Thus an unchanged quantile coexists with a hundredfold increase in tail average. At (\varepsilon>.01) the left .99 quantile jumps to (200). This example separates quantile threshold behavior from sensitivity to tail magnitude.

The numerical severity (200) is illustrative. It may be replaced by any (M>1): for (0<\varepsilon<.01), [ ES_{.99}=1+\frac{\varepsilon}{.01}(M-1), ] while (VaR_{.99}=1). An arbitrarily large (M) makes ES arbitrarily large under fixed positive contamination probability. This is a tail-magnitude statement; it is not a claim that every ES estimator has the same statistical robustness property. Cont, Deguest and Scandolo (2010) explicitly study robustness of measurement procedures, which requires looking at estimation and the data perturbation metric rather than only at the idealized population functional.

A robust estimate can also be systematically biased for the quantity an insurer genuinely needs. Trimming extreme but valid losses may stabilize a premium estimate while understating catastrophe capital; mechanically retaining every suspicious point may import reporting errors into the tail. Data corrections, duplicate event removal, inflation restatement, censoring adjustment and event-definition review require transparent rationales. Neither outlier deletion nor blind retention is a universal remedy.

Insurance consequences of distributional uncertainty

The choice between distributions can affect risk loads, capital allocation, retained exposure and reinsurance attachment. A treaty priced on expected layer loss depends on a tail integral, while a solvency test may depend on a high aggregate quantile. A change in tail fit can therefore alter one decision more than another. For the same portfolio, multiple distributional diagnostics and multiple decision outputs are necessary to explain which assumption is genuinely material.

The 2024 study by Borgonovo, Clemente and Rabitti examines sensitivity requirements for internal models and argues for complementary settings rather than reliance on a single metric. Its proposal is scholarly evidence about supervisory methodology; it is not itself a binding universal regulatory requirement. Applied to distributional selection, its practical implication is to compare model behavior under alternative, explicitly justified perturbations rather than treating one fitted output as self-validating.

Tail Sensitivity

Risk measures, confidence levels and decision scope

For a loss variable (L), define (VaR_p(L)=\inf{x:F_L(x)\ge p}). Define (ES_p(L)=(1-p)^{-1}\int_p^1 VaR_u(L),du) when the integral is finite. For continuous loss distributions, ES equals the conditional mean above the (p)-quantile; with atoms, the integral definition avoids ambiguity in selecting the boundary mass. The two statistics ask different questions. VaR identifies a percentile threshold; ES averages the losses in the upper tail. Neither alone establishes whether the selected tail law is credible, whether a risk can be hedged or whether a given jurisdiction prescribes that measure.

The tail confidence level must be chosen in relation to the intended use. A change from .99 to .995 is a change in the question. It is not an innocuous rounding of the same risk statistic. For a Pareto law, (q_p=x_m(1-p)^{-1/\alpha}); raising (p) makes both the quantile and its shape sensitivity larger in absolute terms. Comparing sensitivities at several confidence levels provides information about where model disagreement becomes consequential.

Generalized Pareto tail-shape experiment

Let exceedance probability above threshold (u) be (\zeta=P(L>u)=.05). Assume that conditional excesses have generalized Pareto survival ((1+\xi y/\beta)^{-1/\xi}), with (\beta=2) and (\xi>0), for (y\ge0). For (p>1-\zeta), [ q_p=u+\frac{\beta}{\xi}\left[\left(\frac{\zeta}{1-p}\right)^\xi-1\right]. ] For (\xi<1), the expected shortfall is [ ES_p=q_p+\frac{\beta+\xi(q_p-u)}{1-\xi}. ] The formula assumes the upper-tail approximation, a stable exceedance rate and an appropriate effective threshold. It is not an unconditional inference from five percent of observations alone.

At (p=.995) the ratio (\zeta/(1-p)) equals (10). With (\xi=.2), the quantile excess (q_p-u) is (5.848932), and ES excess (ES_p-u) is (9.811165). With (\xi=.4), holding the other stated values fixed, the quantile excess increases to (7.559432) and ES excess to (15.932387). The shape change doubles (\xi), increasing the modeled ES excess by roughly 62.4%. A fixed threshold location adds the same (u) to both levels; reporting the excess isolates tail-shape sensitivity from an arbitrary monetary threshold.

As (\xi) approaches one from below, theoretical ES diverges for an unbounded generalized Pareto tail. With (\xi\ge1), the mean excess is infinite. If insured losses are contractually capped, the payout distribution has a finite upper endpoint, and an unbounded severity tail cannot be carried into the insured-loss ES calculation without applying that cap. This distinction between gross severity and insured retained loss is fundamental to an insurance sensitivity analysis.

Tail estimation versus extrapolation

A simulation estimating (VaR_{.999}) from 1,000 independent scenarios has about one observation in the upper 0.1% on average. In fact, if each scenario has independent probability .001 of exceeding the true .999 quantile, the probability of observing zero such exceedances is (.999^{1000}\approx.3677). Simulation count alone does not measure all uncertainty; parameter error and structural assumptions remain. Importance sampling, conditional methods or analytical tail approximations can improve numerical efficiency, but they introduce proposal-distribution and weight diagnostics of their own.

A high-percentile estimate derived from peaks over threshold uses more information in its fitted tail model than a single empirical order statistic, at the price of assuming an asymptotic form and estimating tail parameters. The foundational extreme-value monograph and Pickands' tail inference argument explain why extreme-value methods are methodologically relevant, while their mathematical conditions must remain explicit. A high fitted quantile with a narrow Monte Carlo interval may still be structurally fragile.

Multivariate tail dependence and stress

Tail risk can be dominated by the dependence mechanism rather than by the individual severities. The discrete two-loss counterexample in the Structural Specification chapter already shows that changing a coupling while holding marginals fixed can move ES from (101) to (200). In continuous catastrophic loss systems, common weather drivers, inflation shocks, correlated business interruption and reinsurer default can alter the joint upper tail. Correlation calibrated to ordinary years may offer limited information about coincident extremes.

Reverse stress testing begins with a material adverse output and examines the assumptions or conditions capable of producing it. Makam, Millossovich and Tsanakas (2021) develop divergence-based sensitivity analysis supporting forward and reverse sensitivity with simulated scenarios. The technique is an existing published method, and its computational properties depend on the selected divergence, constraints and available scenarios. A reverse stress is useful as a diagnostic of required joint conditions; it is not a forecast probability unless a defensible probability model for the stress has been separately established.

For a regulatory illustration, European Solvency II Article 242 validation tools address output stability across repeated calculations with the same input data, sensitivity to key assumptions and reverse stress testing. Article 124 addresses the model-validation cycle more broadly. The rules attach to the relevant Solvency II internal-model regime; they are not an assertion that every insurer worldwide follows an identical legal rule.

Numerical and Sampling Effects

Three distinct sources of computational disagreement

The first computational issue is random simulation error: repeated estimations with identical model assumptions but different random draws can differ. The second is numerical approximation error: discretization, quadrature, truncation of loss severity, finite-difference step and convergence tolerances can distort a deterministic target. The third is software and implementation error: code can fail to implement the intended probability law, policy layer or aggregation. All three should be separated from statistical parameter uncertainty and from alternative model structure.

A correct Monte Carlo interval conditional on a fitted model reflects simulation uncertainty only. It does not include unseen historical events, uncertainty in tail family, climate nonstationarity or a contract interpreted incorrectly. Reporting a confidence interval without naming its target and conditioning is therefore insufficient. A useful statement is: “95% interval for the Monte Carlo estimator of this fitted model's .99 Expected Shortfall, conditional on the specified parameters and algorithm.” A separate uncertainty statement is required for parameter refitting or model selection.

Event counts and finite-sample information

If a modeled extreme event occurs independently with probability (q) in each of (n) observations, the expected event count is (nq) and the probability of zero observed events is ((1-q)^n). With (q=.005,n=200), the expected count is one and (P(K=0)=.995^{200}\approx.3670). At (q=.01,n=250), the expected count is (2.5) and (P(K=0)=.99^{250}\approx.0811). Such calculations do not show that the true event rate equals the scenario rate; they show why even a correctly specified model may supply sparse evidence about rare outcomes.

In dependent catastrophe seasons, treating every individual loss as an independent observation exaggerates the effective information about event frequency. Event clusters, changing reporting limits, inflation, litigation, policy limits and portfolio mix further complicate comparison. Resampling units should correspond to the mechanism: observations may need to be resampled by event, calendar period, region or underwriting cohort, not mechanically by individual claim. A bootstrap is an estimator of a specified sampling process, not an independent verification of the process itself.

Efron's (1979) bootstrap method provides a general resampling foundation. Its standard nonparametric form has limited power to recover genuinely unobserved tail events; every resample draws from the observed empirical support. For heavy-tailed statistics, small effective samples and unstable fitted parameters, percentile intervals can have poor coverage. Parametric, semiparametric or threshold bootstrap variants require their own model assumptions and diagnostics.

Reproducible simulations and variance reduction

A simulation study should record the pseudo-random generator, seed, number of runs, sampling scheme, stopping rule, distributional versions, code revision, numerical precision and hardware-dependent behavior where material. It should compare results from distinct seeds and increasing scenario counts. Common random numbers can improve paired scenario comparisons because two model variants are evaluated against matched random innovations; they do not create independent evidence about parameter realism.

Convergence should be evaluated at the actual output of interest. An expected-loss estimate may stabilize quickly while the .999 quantile and ES remain noisy. Diagnostics include repeated-run standard errors, stabilization plots, comparison with analytical identities in tractable subcases, and checks for dominated weights under importance sampling. An algorithm that produces identical answers under identical inputs can be reproducible yet inaccurate because of discretization or logic errors.

The 2024 actuarial working-party study by Kelliher and colleagues provides operational-risk validation guidance, including consideration of distributions, key assumptions, adjusted data and simulation error. Its application to an insurer's operational risk differs from a legal requirement in a country outside its scope. It is valuable here as independently published practical validation scholarship.

Controlled uncertainty experiments

A defensible study can be organized as nested experiments. First fix the data, structure and parameters and vary only the random seed to quantify Monte Carlo error. Next fix the data and model family and re-estimate parameters under appropriate sampling or posterior procedures. Then change the family, threshold, dependence or exposure definition while recording why each alternative is admissible. Finally, test economically meaningful scenarios involving more than one driver, such as inflation interacting with catastrophe frequency or claims delay interacting with reserve development.

The nested separation matters because mixing all factors in one ensemble can make the final spread look informative while preventing attribution. A wide output range may be caused mostly by numerical noise, a genuine tail-shape uncertainty, or a change in insurance coverage. Those are different scientific conclusions with different remedies. Where an uncertainty source cannot be assigned a defensible probability law, an interval or scenario set may be more candid than a falsely precise posterior weight.

Implementation verification and independent checks

Analytical identities are useful software test oracles. A Pareto simulator should reproduce the displayed quantile and conditional tail-expectation formulas within sampling uncertainty; a compound Poisson simulation should approximate (E[S]=\lambda E[X]) and (\mathrm{Var}(S)=\lambda E[X^2]) under the stipulated moments; a dependence engine should reproduce the exact probabilities in the discrete two-loss example. Independent numerical implementations can expose coding mistakes that repetition of the same implementation cannot.

A software check still has a defined reach. Reproducing all synthetic identities validates those calculations, not an insurer's data, peril model, reserve policy or prospective capital estimate. The distinction between method verification and substantive validity is essential to scientific reporting.

Robustness Reporting

An interpretable record of assumptions and outcomes

A robust report begins by identifying the subject: covered peril, exposure population, time horizon, monetary basis, inflation convention, claims definition, contractual layers, accounting treatment and the modeled decision. It defines the baseline and every tested alternative. Each sensitivity result must specify (i) the output and its unit, (ii) the perturbation and its domain, (iii) what was held fixed, (iv) the analytic or computational method, (v) sampling and numerical uncertainty, and (vi) whether the resulting change crosses a decision threshold.

A useful reporting matrix has separate rows for parameter variation, distribution choice, dependence structure, exposure/contract definition, finite-sample refitting, and numerical approximation. For every row, it records a baseline result, stress result, absolute difference, proportionate difference where meaningful, and a short scientific justification. Where model alternatives are mutually inconsistent or not probability-weighted, they remain separately labeled rather than being averaged into a synthetic “confidence interval.”

The interpretation of robustness must be conditional on the tested domain. “Stable within a ±5% change of claim frequency” conveys no evidence about sensitivity to tail shape or correlation. Conversely, a large sensitivity under a highly speculative, economically implausible scenario does not automatically make a model unsuitable for its intended use. The scientific question is whether reasonable alternative assumptions can change a material insurance conclusion and whether the evidence supports a preferred specification.

Source validity and regulatory relevance

Regulatory statements must identify their jurisdiction and legal status. European Solvency II Article 121 statistical-quality requirements require appropriate actuarial and statistical techniques, credible information, realistic assumptions and sufficient data quality for relevant internal models. Article 124 validation standards expressly address the sensitivity of internal-model results to key assumptions. The implementing Article 242 validation tools contain more operational detail, including reverse stress. These statements describe the cited provisions and do not assert country-by-country implementation, model approval or compliance of a particular insurer.

A good scientific record also distinguishes research literature from law. Borgonovo, Clemente and Rabitti (2024) propose complementary model sensitivity settings; Cont, Deguest and Scandolo (2010) investigate measurement-procedure stability; Makam, Millossovich and Tsanakas (2021) supply simulation-compatible divergence stresses. They answer different questions and cannot be collapsed into a single mandatory test. The correct test portfolio is determined by intended output, model characteristics and material uncertainty.

A reproducible dossier of illustrations

The synthetic illustrations can be independently recomputed as follows. For the Pareto example evaluate (q_p=(1-p)^{-1/\alpha}) and (ES_p=\alpha q_p/(\alpha-1)). For the matched lognormal, set (\sigma^2=\log(1.8)) and (\mu=\log(5/3)-\sigma^2/2), then apply the lognormal quantile and Expected Shortfall formulas already provided. For the generalized Pareto example, use (\zeta=.05,\beta=2,p=.995,\xi\in{.2,.4}) and report excess over (u). For the two-event dependence example, enumerate the four joint Bernoulli outcomes under independence, then replace them with the two outcomes of perfectly coincident events. For the simulation-count example, evaluate exact Bernoulli probabilities ((1-q)^n). Each calculation is closed form; no pseudo-random seed is required.

The table below consolidates the question each example resolves. It serves as a check against overgeneralization.

Controlled illustration Fixed conditions Modified condition Identifiable conclusion
Pareto shape Scale and confidence level Severity shape (\alpha) Tail quantile and ES respond nonlinearly
Matched Pareto and lognormal Mean and variance Distribution family Moment matching does not identify the tail
Rare contamination Baseline loss and .99 risk level Small high-loss mass Quantile may stay fixed while ES increases sharply
Two Bernoulli losses Individual marginal distributions Dependence coupling Aggregated VaR and ES can respond very differently
Generalized Pareto excess Threshold, rate and scale Tail shape (\xi) Tail-capital extrapolation can amplify shape uncertainty
Extreme-event sample Event probability Number of draws Absence of events is compatible with a nonzero rare-event rate

A decision analysis should then connect the result to an action without pretending the example is a calibration. Reinsurance attachment might be sensitive to the distribution's layer-loss integral; diversification credit might be sensitive to tail coupling; reserve sufficiency might be sensitive to claims-development assumptions; and economic capital may be sensitive to percentile, model horizon and eligible exposure. Where the ranking of practical alternatives changes within a defensible perturbation region, the decision is assumption-dependent and that dependence must be part of the recommendation.

Materiality, transparency and limits of assurance

Materiality is a decision-relative concept. A 20% change in an estimated loss statistic may be important to a thinly capitalized portfolio and less important to an immaterial exposure. Likewise, a 1% change near a regulatory or contractual boundary can be operationally decisive. The report should therefore pair numerical changes with an explicitly documented threshold, capital buffer, premium margin or scenario tolerance. A stability claim without such a reference can be mathematically true but economically uninformative.

Every reported limitation should be placed with the affected conclusion. Tail distributions may remain observationally indistinguishable over the available sample. Credible dependence under simultaneous catastrophes may be unidentifiable. The simulation may not incorporate post-event claim inflation, litigation, new coverage wordings or emerging technologies. Expert elicitation can broaden scenarios while introducing judgment uncertainty. The record should describe which limitation is reducible by more data, stronger computation or alternative analysis, and which requires a substantive modeling or decision choice.

The scientific value of robustness analysis is diagnostic and comparative. It permits an insurer or researcher to see which sources of model uncertainty can alter a conclusion, what evidence restricts the alternatives, and how a decision performs across them. The dossier's original contribution is the integrated analysis and fully stated examples. It makes no claim of a new theoretical framework, novel statistical estimator, regulator-endorsed threshold or empirical discovery.

Conclusion

Risk sensitivity has no single magnitude independent of the output, perturbation and modeling assumptions. A derivative characterizes one local direction. A global variance index quantifies contribution to output variance under a defined input distribution. A tail stress explores high-consequence events or distributional modifications. A contamination analysis studies responses to data changes. A simulation convergence analysis measures numerical stability. Each is scientifically useful within its declared scope, and each leaves other uncertainties unresolved.

The exact examples establish five complementary results. A small change in a heavy-tail severity shape can cause a larger and asymmetric change in Expected Shortfall. Equal severity means and variances do not imply equal insurance tail risk or stable ranking across quantiles. A low-probability extreme loss can greatly increase ES while leaving a selected VaR unchanged. Dependence assumptions can modify aggregate risk even when every marginal distribution is held fixed. Simulation with few expected tail observations carries irreducible finite-sample information limits under the stipulated event model.

The central practical conclusion is that risk-model robustness is a documented relation between decision-relevant outputs and defensible alternatives. Scientific validation requires checking distributional assumptions, parameter estimation, structural dependence, contract transformation, numerical performance and material consequences in separate but connected analyses. The applicable supervisory framework supplies additional obligations within its own jurisdiction. A final report must preserve source-specific attribution, exact mathematical conditions, reproducible computations and the limits of external evidence. Such documentation turns uncertainty from a concealed feature of a point estimate into an explicit basis for insurance research and informed institutional decisions.

References

  • Artzner, Philippe, Freddy Delbaen, Jean-Marc Eber, and David Heath. 1999. “Coherent Measures of Risk.” Mathematical Finance 9 (3): 203–228. DOI and publisher record.

  • Balkema, A. A., and Laurens de Haan. 1974. “Residual Life Time at Great Age.” The Annals of Probability 2 (5): 792–804. DOI and journal record.

  • Borgonovo, Emanuele. 2007. “A New Uncertainty Importance Measure.” Reliability Engineering & System Safety 92 (6): 771–784. DOI and publisher record.

  • Borgonovo, Emanuele, Gian Paolo Clemente, and Giovanni Rabitti. 2024. “Why Insurance Regulators Need to Require Sensitivity Settings of Internal Models for Their Approval.” Finance Research Letters 60: 104859. DOI and publisher record.

  • Cont, Rama, Romain Deguest, and Giacomo Scandolo. 2010. “Robustness and Sensitivity Analysis of Risk Measurement Procedures.” Quantitative Finance 10 (6): 593–606. DOI and publisher record.

  • Efron, Bradley. 1979. “Bootstrap Methods: Another Look at the Jackknife.” The Annals of Statistics 7 (1): 1–26. DOI and journal record.

  • Embrechts, Paul, Claudia Klüppelberg, and Thomas Mikosch. 1997. Modelling Extremal Events for Insurance and Finance. Berlin: Springer. Publisher bibliographic record.

  • European Insurance and Occupational Pensions Authority. “Solvency II Directive, Article 124: Validation Standards.” EIOPA Single Rulebook, source provision. Article 124.

  • European Insurance and Occupational Pensions Authority. “Commission Delegated Regulation (EU) 2015/35, Article 242: Validation Tools.” EIOPA Single Rulebook, source provision. Article 242.

  • European Parliament and Council of the European Union. 2009. Directive 2009/138/EC on the Taking-Up and Pursuit of the Business of Insurance and Reinsurance (Solvency II), Article 121 and related provisions; consolidated version dated January 17, 2025. EUR-Lex consolidated text.

  • Kelliher, Patrick O. J., Madhu Acharyya, Andrew J. Couper, Edward N. V. Maguire, Choong A. Pang, Christopher M. Smerald, Jennifer K. Sullivan, and Paul M. Teggin. 2024. “Operational Risk Working Party – Validating Operational Risk Models.” British Actuarial Journal 29: e1. DOI and Cambridge record.

  • Mack, Thomas. 1993. “Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates.” ASTIN Bulletin 23 (2): 213–225. DOI and publisher record.

  • Makam, Vaishno Devi, Pietro Millossovich, and Andreas Tsanakas. 2021. “Sensitivity Analysis with χ²-Divergences.” Insurance: Mathematics and Economics 100: 372–383. DOI and publisher record.

  • McNeil, Alexander J., Rüdiger Frey, and Paul Embrechts. 2015. Quantitative Risk Management: Concepts, Techniques and Tools. Revised edition. Princeton: Princeton University Press. Publisher book record.

  • Morris, Max D. 1991. “Factorial Sampling Plans for Preliminary Computational Experiments.” Technometrics 33 (2): 161–174. DOI and publisher record.

  • Peters, Gareth W., Matteo Malavasi, Georgy Sofronov, Pavel V. Shevchenko, Stefan Trück, and Jiwook Jang. 2023. “Cyber Loss Model Risk Translates to Premium Mispricing and Risk Sensitivity.” The Geneva Papers on Risk and Insurance – Issues and Practice 48: 372–433. Springer article record.

  • Pickands, James, III. 1975. “Statistical Inference Using Extreme Order Statistics.” The Annals of Statistics 3 (1): 119–131. DOI and journal record.

  • Rockafellar, R. Tyrrell, and Stanislav Uryasev. 2000. “Optimization of Conditional Value-at-Risk.” Journal of Risk 2 (3): 21–41. DOI and publisher record.

  • Saltelli, Andrea, Marco Ratto, Terry Andres, Francesca Campolongo, Jessica Cariboni, Debora Gatelli, Michaela Saisana, and Stefano Tarantola. 2008. Global Sensitivity Analysis: The Primer. Chichester: Wiley. Publisher book record.

  • Sobol, I. M. 2001. “Global Sensitivity Indices for Nonlinear Mathematical Models and Their Monte Carlo Estimates.” Mathematics and Computers in Simulation 55 (1–3): 271–280. Publisher article record.

  • Tsanakas, Andreas, and Pietro Millossovich. 2016. “Sensitivity Analysis Using Risk Measures.” Risk Analysis 36 (1): 30–48. DOI and journal record.