RESEARCH DOSSIER
Simulation and Numerical Risk Evaluation
Abstract. Simulation and numerical risk evaluation transform assumptions about random loss-generating mechanisms into estimates of distributions, tail probabilities, capital measures, and decision-relevant sensitivities. This monograph develops the mathematical and computational foundations of Monte Carlo risk analysis, including compound insurance losses, dependent exposures, pseudo-random sampling, statistical error, importance sampling, conditioning, stratification, rare-event computation, quantile and expected-shortfall estimation, and reproducibility. It distinguishes conditional numerical precision from uncertainty about model specification, parameters, exposure data, and future regimes. A verified illustrative compound-Poisson experiment compares exact moments with 10,000, 100,000, and 1,000,000 simulated years. Further exact counterexamples show how nominally accurate Monte Carlo output can conceal rare losses and how quantile estimation becomes unstable when few observations reach the relevant tail. The original analytical synthesis provides a source-traceable research workflow for choosing estimators, validating assumptions, reporting uncertainty, and connecting simulated results to risk governance. The inquiry concludes that numerical evaluation becomes scientifically credible when a defined estimand, defensible probability mechanism, independently checked simulation implementation, and explicit numerical and model-uncertainty statements jointly support the decision.
Keywords: Monte Carlo simulation; numerical risk evaluation; aggregate insurance loss; collective risk model; rare events; importance sampling; conditional Monte Carlo; variance reduction; heavy tails; compound Poisson; value-at-risk; expected shortfall; tail estimation; dependence; reproducibility; numerical convergence; solvency; model uncertainty.
Introduction
Simulation supplies a practical bridge between probability theory and the evaluation of insurance losses whose distributions cannot be conveniently integrated in closed form. A model specifies possible outcomes and the mechanisms by which they occur; an algorithm samples those outcomes or reweights alternative samples; an estimator summarizes a chosen property; and an uncertainty statement characterizes the reliability of that numerical result. The bridge has several load-bearing assumptions. Correct coding cannot rescue an unsuitable loss distribution, while a well-founded distribution still requires careful numerical computation, diagnostic testing, and an intelligible relationship to the eventual decision.
The foundations are extensively documented in Simulation and the Monte Carlo Method by Rubinstein and Kroese (2016) and Glasserman's Monte Carlo Methods in Financial Engineering (2003). Their methods address integration, simulation design, variance reduction, and complex stochastic systems. Insurance adds specific complications: low-frequency high-severity claims, contractual limits, reinsurance attachment points, aggregation across dependent risks, claims emergence over time, and capital decisions driven by remote quantiles. Extreme-value theory, treated systematically in Embrechts, Klüppelberg, and Mikosch's Modelling Extremal Events (1997), supplies a complementary mathematical language for severe outcomes.
The central research question is how a computation can be judged reliable when the event of interest may occur once in thousands of years under the assumed model. That question requires separating the true but unknown physical or economic process, the probabilistic model of that process, the target functional under the model, and the random uncertainty arising from an algorithm's finite computational budget. Throughout this study, worked financial amounts are hypothetical accounting units. None is an empirical estimate for a named insurer or market.
The research method combines mathematical derivation, comparison of established algorithms, direct examination of original publisher and institutional records, a fully specified computational case, explicit counterexamples, and analytical interpretation. It does not introduce a new named theory, probabilistic doctrine, or universal capital rule. The original contribution lies in a rigorous synthesis and in independently checked numerical illustrations that make assumptions and failure conditions transparent.
Monte Carlo and Random Sampling
The Scientific Object of Numerical Risk Evaluation
A numerical risk evaluation answers a defined question about a loss variable or stochastic path. Let L be a real-valued loss over a specified period, with positive values denoting loss. Relevant quantities include E[L], Var(L), P(L>u), a quantile q_alpha, expected shortfall ES_alpha, the probability of an insurer's assets falling below obligations, and the financial consequence of policy interventions. Every quantity needs a monetary convention, exposure population, observation horizon, contract boundary, and a probability measure.
The formal distinction between an estimand and an estimator is foundational. The estimand is a property of the specified mathematical model, such as p=P(L>u). An estimator is a random statistic computed from simulated outcomes, such as p_hat=(1/n)sum 1{L_i>u}. Estimation error is p_hat-p; its magnitude is itself random. Separately, model error arises when the chosen probability law does not adequately represent the relevant world. Confidence intervals for p under the selected model do not measure this latter discrepancy.
The framework of coherent measures of risk proposed by Artzner, Delbaen, Eber, and Heath (1999) illustrates why the selected functional is substantive. A model can generate the same loss scenarios for several decision criteria, while its estimated variance, quantile, expected shortfall, and ruin probability convey materially different implications. A simulation designed to estimate an expected loss well may perform poorly for a 99.9th-percentile capital measure. Numerical design must follow the question, rather than the ease of computing an attractive statistic.
Numerical risk evaluation is an inferential activity, not merely software operation. A scientifically complete claim should state: the mathematical object; the probability assumptions; the estimation procedure; the evidence used for calibration; numerical error; structural and parameter uncertainty; diagnostics; the result's domain of validity; and the decision to which it is relevant. A missing element can change interpretation without visibly changing the number printed by the program.
Probability Models, Estimands, and Decision Horizons
A one-year property portfolio model might describe N claims in a year and associated claim severities X_1,...,X_N. A different model could treat reported claims, ultimate claims, or discounted settlement cash flows. These are different random variables, even when all are called annual losses. In a capital model, contractual coverage, recoveries, expenses, taxes, and asset dependence may also matter. The sequence of transformations must be explicit before a distribution can be simulated.
Let (Omega,F,P) be a probability space and L=g(Z_1,...,Z_d;theta) the loss determined by drivers Z, model parameters theta, and a measurable transformation g. The computational target might be m(theta)=E_theta[g(Z;theta)], p_theta(u)=E_theta[1{g(Z;theta)>u}], or a distributional functional. The notation reveals three questions: which drivers have been included, which relationships define g and P_theta, and how theta was established from data or judgment.
Time is part of the estimand. Annual aggregate claims and 20-year discounted adverse cash flow are not numerically interchangeable. A one-year high quantile may understate long-duration reserve uncertainty; a multi-year tail estimate may combine exposure growth, investment returns, inflation, and path-dependent management actions. Conditional and unconditional assessments also differ. A catastrophe risk conditioned on a particular climate state need not equal the risk averaged over an assumed state distribution.
Risk factors frequently contain mixtures. An exposure can experience an ordinary attritional claim process and an exceptional severe-event process. Dependence may result from common weather, legal inflation, disease outbreaks, supply chains, or correlated asset losses. Choosing normal approximations simply because first and second moments exist can be misleading when the decision is concentrated in a remote tail. The book Quantitative Risk Management by McNeil, Frey, and Embrechts (2015) places extreme outcomes and the dependence of risk factors at the center of quantitative practice.
Plain Monte Carlo: Sampling and Algorithms
For independent simulated values L_1,...,L_n distributed as L, the ordinary Monte Carlo estimate of E[L] is m_hat=n^{-1}sum_i L_i. If E[L^2]<infinity, Var(m_hat)=Var(L)/n and standard error declines at the familiar 1/sqrt(n) rate. Under suitable finite-variance conditions, the central limit theorem makes a normal approximation useful for a sampling interval. That interval describes repeated independent executions of the algorithm for a fixed underlying model. Its coverage cannot be silently interpreted as a claim that the model is correct.
Uniform pseudo-random numbers are transformed into samples from the required distributions. Algorithmic sequences are deterministic conditional on a seed and generator specification; their suitability depends on statistical quality and the simulation design. Inverse-transform sampling draws U uniformly on (0,1) and sets X=F^{-1}(U), with the generalized inverse appropriate for jumps. Other models use rejection, composition, Poisson generators, multivariate transforms, and simulation of stochastic paths. Glasserman's random-variable generation chapters (2003) provide established coverage of these foundations.
An independent stream is particularly important when nominally separate replications are used to estimate sampling variability. Reusing one seed and exactly the same random sequence produces identical runs, not independent confirmations. For parallel simulations, streams or counter-based designs must be assigned without unexamined overlaps. The independence assumptions required by a confidence interval need to be assessed independently of whether the code completed successfully.
Monte Carlo computation is frequently described by its scenario count, yet the information in a run depends on the incidence and magnitude of the observations relevant to its target. One million simulated ordinary claims cannot reveal the size of an unobserved catastrophic component whose probability distribution is misspecified. Nor does a fixed seed ensure future reproducibility if the software version, generator, transformations, or data inputs are changed without record. A seed is one component of an audit trail.
Sampling Error, Confidence Intervals, and Effective Precision
For a rare-event probability p, the direct estimator p_hat=n^{-1}sum_i I_i has E[p_hat]=p and Var(p_hat)=p(1-p)/n when indicators are independent. Its relative standard error is sqrt((1-p)/(np)). This identity supplies an immediate feasibility test before running a large simulation. If p=0.0001 and n=100,000, the expected number of event observations is 10 and the approximate relative standard error is 31.6 percent. Seeking 5 percent relative standard error under the same assumptions needs approximately n=3,999,600 scenarios. This is an asymptotic precision calculation, not a guarantee of tail-model fidelity.
When the simulated event count is small, a normal interval for p may have poor coverage. Reporting a zero observed count as proof of zero probability is an elementary inferential error. If n independent Bernoulli trials have no successes, an exact one-sided 95 percent upper confidence bound for p is 1-0.05^{1/n}; for large n this is approximately 2.996/n. Even this bound conditions on independent, correctly generated Bernoulli trials with a fixed p. A misspecified generator or probability model invalidates the inferential premise.
For mean losses, use an independently computed sample variance and uncertainty interval where finite-variance and large-sample conditions are credible. Where severities are very heavy-tailed, finite observed variance in a finite run does not establish that the theoretical second moment exists. Standard central-limit approximations may be unreliable when moments are infinite or near-extremes dominate the sample. An alternative estimator, analytic tail treatment, or carefully qualified nonparametric method may then be required.
The JCGM Monte Carlo supplement (JCGM 101:2008) sets out a formal distribution-propagation approach in measurement science and makes explicit its mathematical conditions for outputs and numerical validity. Insurance models have different substantive inputs and outcomes, but the methodological lesson transfers: simulation results must be read together with assumptions about the inputs and the sufficiency of the computational sample. NIST's consistent Monte Carlo uncertainty analysis by Jamroz and Williams (2020) further emphasizes consistent, reproducible replicates in distributed computations.
Variance Reduction and Experimental Design
The purpose of variance reduction is to obtain a more accurate estimate of a particular model functional with a fixed computational budget, or the same accuracy with fewer simulation evaluations. Techniques do so by using mathematical information about the problem, introducing deliberate dependence between draws, conditioning on known variables, or changing the simulation distribution with a compensating weight. The relevant comparison is estimator error per unit of computational cost. A procedure that lowers raw sample variance but requires much more computation may not improve practical efficiency.
In antithetic sampling, draws may be paired using U and 1-U so that the average of the paired outputs has lower variance when the response has a suitable monotone structure. A control variate uses a variable C with known mean mu_C and correlation with the target output: estimate E[Y] through mean(Y-beta(C-mu_C)). The variance-minimizing coefficient in the ideal finite-variance setting is beta*=Cov(Y,C)/Var(C). If beta is estimated from the same sample, uncertainty analysis must account for its estimation and any resulting finite-sample properties. The presence of a strong control relationship should be demonstrated rather than assumed.
Stratification allocates sample effort to disjoint regions of the input space. If P(Z in A_j)=w_j and independent estimates within strata have variances sigma_j^2/n_j, an estimator sum_j w_j mean(Y|A_j) has variance sum_j w_j^2 sigma_j^2/n_j. Allocating effort to high-variance strata often improves precision. Latin hypercube sampling spreads sampled inputs across marginal probability intervals; the original comparison of sampling plans by McKay, Beckman, and Conover (1979) established improvements for specified estimator classes. Such improvements do not imply a distribution-free guarantee for arbitrary discontinuous tail events or dependent inputs.
Quasi-Monte Carlo designs use low-discrepancy point sets instead of ordinary independent pseudo-random samples. They can be highly effective for sufficiently regular integrands, especially when the effective dimension is controlled, but naive independent-sampling standard errors are inapplicable to a fixed deterministic point set. Randomized designs and replication can permit empirical error assessment. For high-dimensional insurance catastrophe losses with indicator discontinuities, the smoothing and effective-dimension assumptions require special attention. Glasserman's variance-reduction and quasi-Monte Carlo treatment (2003) distinguishes these techniques within a broader computational framework.
A practical experimental design should compare several estimators on the same mathematical target, using independent replication or an intentionally paired design. Report wall-clock cost, number of effective model evaluations, bias, empirical variance, and confidence coverage where a trusted benchmark exists. Prefer analytic identities as controls: frequency-severity formulas, zero-loss probabilities, conservation relationships, and limiting behavior under simplified assumptions often reveal implementation defects that higher scenario counts will conceal.
Loss Process Simulation
Frequency–Severity Aggregation
The collective risk model writes aggregate loss as S=sum_{i=1}^N X_i, with S=0 when N=0. If N is independent of identically distributed severities and has E[N]=lambda, the mean is E[S]=E[N]E[X]. Under finite second moments, Var(S)=E[N]Var(X)+Var(N)(E[X])^2. In the important compound-Poisson case N~Poisson(lambda), Var(N)=E[N]=lambda and hence Var(S)=lambda E[X^2]. These identities are useful independent benchmarks for a simulation engine.
The distribution of N reflects the assumed claim count mechanism. A Poisson model supposes equidispersed independent increments under its basic process interpretation. Overdispersion, policyholder heterogeneity, contagion, seasonality, or claims clustering may require a mixed Poisson or other count model. Severity distributions can be discrete, continuous, mixed, truncated, censored, or contractually modified. A rare catastrophe component often dominates E[X^2] even when its contribution to the claim count is negligible.
A severity model should distinguish gross economic loss from insured payment. Deductibles, limits, coinsurance, waiting periods, and exclusions change the simulated payable amount. For a per-occurrence deductible d and payment limit l, an illustrative payment function is Y=min(max(X-d,0),l), assuming the contractual wording supports that representation. Reinsurance may introduce retention, aggregate attachment, reinstatements, hours clauses, or other path-dependent terms. These transformations should be applied to the same simulated underlying events in the appropriate order; adding expected recoveries to an unrelated gross-loss simulation can erase dependence and contractual nonlinearity.
A simulation can be forward, generating events then payments; or conditional, generating selected components given other information. Conditional simulation is often more efficient when the underlying problem has known structure. For portfolios with reporting delays, the simulation may require occurrence date, reporting date, paid development, inflation, and a cut-off valuation date. The estimation object must state whether it measures ultimate cost or observed cash flow and whether discounting is applied.
Dependence, Exposure, and Aggregate Loss
Independent sampling of each coverage line creates a specific joint distribution: the product of the fitted marginals. Where underlying risks share hazards or macroeconomic drivers, the independence assumption can materially alter the simulated tail. Correlation captures some linear dependence but does not determine joint extreme-event frequency. Two joint models with identical marginal distributions and Pearson correlations may exhibit different tail configurations and aggregate quantiles.
A useful architecture separates exposure generation, event generation, physical or financial intensity, individual vulnerability, contract transformation, and aggregation. Common-shock scenarios should enter at the event level where the causal channel can be represented. If flood losses in two regions depend on the same storm path, independent per-region random loss draws can misrepresent catastrophe accumulation. An insurer-wide asset-liability simulation should preserve any economically material dependence between claim shocks and the assets intended to fund them.
Embrechts, Puccetti, and Rüschendorf (2013) demonstrate in model uncertainty and VaR aggregation that dependence uncertainty can produce substantial uncertainty in aggregate quantile assessments. Their findings concern model-based dependence bounds and examples, not universal empirical loss levels for insurance portfolios. A meaningful report therefore provides sensitivity to joint-distribution choices, especially in a high-quantile decision.
Dependence is also time-dependent. Inflation correlations, asset-liability feedback, climate exposures, risk mitigation, and recovery rates can change under stress. Where the support for an assumed joint distribution rests on short historical records, scenario analysis should make the limitations visible. Simulation can evaluate the consequences of alternative dependence assumptions; it cannot determine which assumption is true from computational volume alone.
Exact and Simulated Compound-Poisson Case Study
Design, distribution, and analytic checks
Consider an invented annual portfolio with total frequency NPoisson(10), independent severities X taking the value 1,000 with probability 0.995 and 100,000 with probability 0.005. This is a deliberately transparent two-point loss model; it is not calibrated to claims data and makes no claim about any actual line of insurance. The severe layer is rare at the individual-claim level but material to annual variance and tail risk. Under Poisson splitting, the numbers of ordinary and severe claims are independent N_oPoisson(9.95) and N_s~Poisson(0.05), and annual loss is S=1,000 N_o+100,000 N_s.
The expectation of one severity is 0.995(1,000)+0.005(100,000)=1,495. Thus E[S]=10(1,495)=14,950. Its second raw severity moment is E[X^2]=0.995(1,000,000)+0.005(10,000,000,000)=50,995,000. Consequently, Var(S)=10(50,995,000)=509,950,000 and SD(S)=22,582.0725. The probability of at least one severe claim in a year is 1-exp(-0.05)=0.0487706. These are exact model implications, apart from displayed rounding. Any simulation using the stated process should converge toward them as sampling error declines.
The model's ordinary annual expected loss is 9,950, while the severe layer's expected loss is 5,000. Severe claims therefore represent one third of expected annual loss while accounting for one half of one percent of individual claims. The variance contribution of the severe layer is 500,000,000 out of 509,950,000, or about 98.05 percent. These ratios show how a small frequency weight can dominate dispersion. In a contractually layered portfolio, a very similar mechanism can dominate capital and reinsurance decisions, even when it contributes little to claim counts.
Reproducible simulation
The following calculation was independently executed using NumPy's default_rng(20261114); successive samples of size 10,000, 100,000, and 1,000,000 were drawn without resetting the generator between sample sizes. For each size, the experiment independently drew ordinary and severe counts with Poisson means 9.95 and 0.05, and computed annual S. Sample variance uses the n-1 denominator. Empirical quantiles use the smallest observed threshold whose empirical CDF reaches the selected level (method="inverted_cdf"). All values are hypothetical accounting units; they are outputs of a documented experiment rather than source claims taken from a book or external data.
| Simulated years | Mean S | Sample variance | Severe-event frequency | Empirical VaR 95% | Empirical VaR 99% |
|---|---|---|---|---|---|
| 10,000 | 14,911.90 | 501,676,106.00 | 0.049300 | 21,000 | 113,000 |
| 100,000 | 14,950.03 | 510,032,853.33 | 0.048770 | 21,000 | 113,000 |
| 1,000,000 | 14,986.87 | 514,674,160.17 | 0.049062 | 21,000 | 113,000 |
| Exact model benchmark | 14,950.00 | 509,950,000.00 | 0.048771 | Distribution-defined | Distribution-defined |
The table does not suggest monotone convergence. Each row is a different random sample, and a longer sample can be farther from an expectation than a shorter one. The large sample's mean exceeds the analytic mean by 36.87, while its sample variance exceeds the analytic variance by about 0.93 percent. That outcome is consistent with finite-sample randomness. Reproducibility means the specified computational environment can recreate the reported numbers, while inferential reliability requires replication, standard-error assessment, and model validation beyond one deterministic seed.
The simulation can be reproduced with the following complete, standalone Python calculation, assuming NumPy is installed. It uses no downloaded datasets and creates no network traffic.
import numpy as np
rng = np.random.default_rng(20261114)
for n in (10_000, 100_000, 1_000_000):
ordinary = rng.poisson(9.95, n)
severe = rng.poisson(0.05, n)
loss = 1_000 * ordinary + 100_000 * severe
print(n, loss.mean(), loss.var(ddof=1),
np.mean(severe >= 1),
np.quantile(loss, 0.95, method="inverted_cdf"),
np.quantile(loss, 0.99, method="inverted_cdf"))
Interpretation and hidden weaknesses
The 95th-percentile threshold is close to a structural transition in this two-component model: severe annual claims occur in approximately 4.88 percent of years, so the most adverse approximately five percent of years almost entirely consist of years with at least one severe claim, apart from high ordinary-count outcomes and distributional overlap. The empirical quantile happens to remain 21,000 in the displayed samples, but its apparent stability cannot establish robustness under a different severity probability or severity amount. A small increase in the severe-claim intensity could move the percentile into a different mixture region and produce a large quantile change.
An important diagnostic is the severe-claim count itself. Under N_s~Poisson(0.05), observing at least one severe claim in about 4.88 percent of simulated years is consistent with the model's event process. Yet this validates only code implementation against an assumed frequency distribution. An empirical finding that the actual severe frequency is 0.008 instead of 0.005 per claim would change the capital assessment regardless of whether the code perfectly reproduces Poisson(0.05).
A full operational insurance assessment would require exposure-level data, claims development, inflation, reinsurance terms, dependence, asset returns, and an explicit currency and valuation date. The simplified model intentionally excludes these features so that the mathematical checks are transparent. This is an advantage for testing numerical machinery and a limitation for extrapolating results to insurance practice.
Importance Sampling
Change of Measure and Likelihood Weights
Importance sampling estimates a probability or expectation under one probability measure by sampling under another. Let f be a target density, g a proposal density, and h the integrand. If g>0 wherever h f is nonzero, then mu=E_f[h(X)]=E_g[h(X)w(X)] with likelihood ratio w(x)=f(x)/g(x). The estimator mu_hat_IS=n^{-1}sum h(X_i)w(X_i), X_i sampled independently from g, is unbiased when the required integrability conditions hold. Its variance is Var_g(h(X)w(X))/n if the second moment is finite.
This change-of-measure identity is exact, subject to support and integrability. It is developed in Glynn and Iglehart's importance sampling for stochastic simulations (1989), which extends the method beyond simple integrals to stochastic-system paths. In insurance applications, a proposal may make a large loss, a stress scenario, or a rare claims cluster more frequent in the simulated sample. The weighted result must then recover the original target distribution. Counting proposal events without applying their appropriate weights estimates a different probability and cannot be used as the original risk estimate.
For an event A, the ideal zero-variance proposal would be the target law conditional on A, assuming p=P_f(A)>0 and an implementable conditional sampler. Under that proposal, 1_A w equals the constant p. The construction is usually impractical because it requires knowledge of p or an equivalent difficult probability. The conceptual lesson is operational: a good proposal resembles the conditional law of the rare event. For path-dependent systems the relevant likelihood ratio is often a product of transition-density ratios, or a Radon–Nikodym derivative for a path measure. Missing a factor, using the wrong filtration, or applying a terminal-state ratio to a path event can yield silently biased estimates.
An apparently successful importance-sampling run can still fail. A proposal may have insufficient support or produce enormous weights, yielding high or infinite variance. A handful of dominant weights can drive a numerically stable-looking mean. Weight diagnostics, second-moment estimates, independent replications, proposal sensitivity, and simple-case benchmarks are mandatory substantive checks. The commonly reported effective sample size (sum w_i)^2/sum w_i^2 is a descriptive warning indicator for weight concentration; it is not generally a calibrated error bar for rare-event estimators.
Glasserman, Heidelberger, and Shahabuddin (1999) analyze asymptotically optimal importance sampling for path-dependent option simulations under specific assumptions, and their subsequent variance reduction for VaR (2000) combines importance sampling and stratification to estimate loss probabilities. Their finance examples establish methodological precedents, while the applicability to insurance catastrophe or reserve-loss processes requires matching assumptions about tails, dimensions, and event structure.
Importance-Sampling Counterexample and Efficiency
Consider an idealized Bernoulli rare event A with target probability p=10^-4. Under direct simulation, a 10,000-trial run has an expected event count of one. The exact relative standard error of the ordinary event-frequency estimator is sqrt((1-p)/(10,000p))=sqrt(0.9999), approximately 1.0, or 100 percent. The likelihood of observing zero rare events in that run is (1-10^-4)^{10,000}, approximately 36.79 percent. A result of zero from that run would contain little evidence against the posited event probability.
Suppose instead that the simulation uses a Bernoulli proposal Q(A)=q=0.1. For the event indicator estimator, the weighted contribution is p/q=0.001 if A occurs and zero otherwise. Its expectation under Q is q(p/q)=p. The variance of one weighted draw is p^2(1-q)/q, giving relative standard error sqrt((1-q)/(nq))=sqrt(9/n). At n=10,000 this is 0.03, or 3 percent, against approximately 100 percent for direct Monte Carlo. The ratio of standard errors is about 33.3; the ratio of variances is about 1,111. This is a mathematical toy change of measure where p is already known, designed to illustrate the method, not a claimed practical algorithm for discovering an unknown probability.
The example shows that counting rare events is not necessarily the correct unit of information when sampling under an altered law. The proposal generates about 1,000 event observations instead of one, but each observation is weighted to preserve the original estimand. The improvement depends on knowing the exact likelihood ratio and on choosing a useful q. For a complex insurance scenario, constructing such a proposal is the hard scientific work. Glynn and Iglehart's change-of-measure treatment (1989) supplies the methodological foundation for more complicated stochastic systems.
For this simple toy, increasing q toward 1 makes the event-indicator estimate approach a constant, but doing so does not generalize freely: the proposal must retain sufficient support for every outcome contributing to the target quantity, and other integrands may depend heavily on the complement event. Optimizing a proposal for P(A) may worsen the estimate of E[L] or ES_alpha. Every declared gain in efficiency must therefore specify the estimand and the cost model.
Rare Event Computation
Rare Event Estimation and Algorithm Selection
A rare event is defined relative to a probability law, a horizon, and an event threshold. An event with an annual probability of one in 10,000 may be materially relevant for a large insurer, but direct Monte Carlo sampling may be prohibitively noisy at ordinary budgets. Rare-event algorithms aim to concentrate computation on influential regions while preserving an estimand that can be interpreted under the original law. Common approaches include importance sampling, conditional simulation, splitting, multilevel methods, cross-entropy adaptation, and analytic approximations joined to numerical integration.
Rare-event efficiency must be defined carefully. Suppose p(u)=P(L>u) decreases as threshold u increases. For an unbiased estimator Z_u, bounded relative error means sup_u Var(Z_u)/p(u)^2 stays bounded over the stated asymptotic regime. Logarithmic or asymptotic optimality can involve weaker rate comparisons and specific scaling assumptions. Neither label guarantees finite-sample superiority for an insurer's actual loss threshold. Computational budget, parameter uncertainty, and proposal robustness remain part of the evaluation.
A splitting method represents a difficult event as passage through a sequence of levels. Paths reaching an intermediate level may be copied or selectively continued, and weighted conditional transition probabilities are combined. The levels should be chosen to avoid both degeneration and excessive dependence between descendants. An adaptive algorithm that changes levels in response to preliminary data requires its own consistency and uncertainty analysis. The method is useful when rare outcomes develop through sequential states, such as ruin, network outage, or a cumulative catastrophe loss path.
Importance sampling can be especially attractive for light-tailed laws where exponential tilting follows large-deviation structure. Heavy-tailed sums may behave differently: one exceptionally large claim can dominate the event S>u. Exponential moment-generating functions may fail to exist at positive arguments, defeating a naive exponential-tilting construction. The distinction between light-tailed and subexponential rare-event mechanisms is therefore structural rather than merely a question of choosing a bigger sample. Embrechts, Klüppelberg, and Mikosch (1997) develop the underlying tail and extremal-event mathematics.
A valid rare-event study should show the exact event definition, its target probability law, the algorithm's identity or weighting proof, its conditions for finite variance or consistency, a cost-adjusted precision comparison, diagnostic evidence in the actual threshold range, and sensitivity to alternative plausible tail models. It should also report unsuccessful or unstable parameter settings when they materially affect confidence in the chosen method.
Conditional Monte Carlo and Heavy-Tailed Losses
Conditional Monte Carlo uses an expectation of an indicator or payoff given a convenient subset of random variables. If Z=1_A and G is a sigma-field for which E[Z|G] can be evaluated, then E[E[Z|G]]=P(A). By the law of total variance, Var(E[Z|G])<=Var(Z) when second moments exist. The variance reduction follows mathematically, although the cost of computing the conditional expectation and the availability of a tractable conditional law determine whether the method is operationally worthwhile.
For independent nonnegative severities X_1,...,X_n, write S_n=sum_i X_i. Conditioning on the first n-1 severities gives P(S_n>u|X_1,...,X_{n-1})=bar F(u-sum_{i<n}X_i), where bar F(t)=P(X_n>t) and is defined as 1 for t<0 when severities are nonnegative. The resulting random conditional probability can be averaged directly. This is an exact identity; it can turn a binary event indicator into a smoother estimator.
Asmussen and Kroese's heavy-tail rare-event algorithms (2006) exploit additional structure for large sums of identically distributed heavy-tailed variables. Their analyzed conditional estimator uses a symmetry and maximum-claim representation and establishes bounded-relative-error properties in the specified setting. These results should not be transplanted without proof to dependent severities, arbitrary mixed distribution families, complex policy deductibles, or multi-period capital processes. In those applications, the conditional event and tractability conditions must be derived again.
The frequent assertion that a billion independent simulated years necessarily provides a precise assessment of any rare insured loss is false. For a threshold probability of 10^-12, even one billion ordinary independent years has an expected exceedance count of 0.001. The probability of seeing no exceedances is approximately exp(-0.001), about 99.9 percent. A conditional algorithm or defensible tail model can be more informative than simply increasing n, provided its mathematical and empirical assumptions are supported.
Extreme-value modeling and simulation complement each other. An extreme-value fit estimates tail behavior from actual data under regularity assumptions; a rare-event simulator computes tail consequences under a specified probability law. Neither automatically validates the other. Fitted tail-index uncertainty, threshold selection, temporal change, and claims truncation are separate from the numerical error of a simulation based on fixed fitted parameters. A complete uncertainty statement must preserve both levels.
Quantile and Tail Estimates
Risk Quantiles and Tail Functionals
For confidence level alpha in (0,1), the generalized loss quantile is q_alpha=inf{x:F_L(x)>=alpha}. Value-at-risk at alpha is often identified with this loss quantile, subject to sign and convention. With independent samples and a continuous, strictly increasing local distribution, an empirical quantile can be estimated from an appropriate order statistic. Its asymptotic variance under regular conditions is alpha(1-alpha)/(n f(q_alpha)^2), where f is the density at the true quantile and f(q_alpha)>0. Thus a small tail density makes even a large Monte Carlo sample produce an imprecise quantile estimate.
A normal approximation for sample quantiles requires local smoothness and sufficient effective tail observations. It is dangerous to carry that approximation over to a distribution with point masses or a flat cumulative-distribution segment. Such features occur naturally in insurance through zero claims, coverage limits, deductibles, and clustered reinsurance attachment levels. A distribution's 99.5th quantile can shift discontinuously under very small changes to probability mass close to the selected level. Quantile sensitivity should therefore be studied at the mathematical level, not inferred from simulation convergence alone.
Expected shortfall has several representations and convention-dependent details. For integrable loss L, a robust quantile-integral representation is ES_alpha(L)=(1/(1-alpha))int_alpha^1 q_u(L) du. This representation handles atoms consistently and is associated with the coherence analysis in Acerbi and Tasche's study of expected shortfall (2002). For sufficiently continuous distributions it coincides with the mean loss beyond the alpha quantile, but a simple conditional average E[L|L>=q_alpha] need not have the intended interpretation when there is a mass point at the quantile.
Rockafellar and Uryasev's optimization of conditional value-at-risk (2000) provides an optimization representation that avoids treating quantile selection as the only numerical route to a tail-sensitive criterion. For an integrable loss, ES_alpha(L)=min_t {t+(1/(1-alpha)) E[(L-t)_+]}, with the familiar qualification that minimizers may be nonunique for discrete distributions. Scenario averages yield a tractable convex empirical objective in t. This is a mathematical property of the stated loss functional; it does not itself prescribe which insurer capital threshold should be adopted.
Importance sampling complicates the estimation of empirical quantiles and expected shortfall: ordinary unweighted order statistics from the proposal distribution generally estimate the wrong quantile. Weighted empirical distribution estimators, likelihood-ratio moment estimators, or a specialized optimization estimator are needed, along with proper error analysis. A variance-reduction method proven efficient for one tail probability need not be efficient for a range of quantiles or for expected shortfall across all levels.
Discontinuous Losses and Expected Shortfall
Consider a hypothetical annual loss variable with P(L=0)=0.990, P(L=100)=0.009, and P(L=10,000)=0.001. Its cumulative probabilities are F(0)=0.990, F(100)=0.999, and F(10,000)=1. Its 99th percentile under the generalized inverse definition is q_0.99=0. That number alone communicates little about the size of the nonzero tail. At 99.5 percent, q_0.995=100, since F(0)<0.995<=F(100).
The 99th-percentile expected shortfall under the quantile-integral definition averages the worst 1 percent of probability mass: ES_0.99=[0.009(100)+0.001(10,000)]/0.01=1,090. The 99.5th-percentile expected shortfall is ES_0.995=[0.004(100)+0.001(10,000)]/0.005=2,080. The different figures are consequences of exact probability mass allocation rather than simulation error. A naive rule that simply averages values strictly greater than the quantile would mishandle the mass at 100 for the 99.5-percent calculation.
This counterexample illustrates a broader research lesson. For portfolios with contractual caps, zero-claim masses, or discrete catastrophe layers, stating a confidence level does not fully specify an estimator. One must specify the quantile convention, the ES definition, the treatment of ties, and whether any smoothing or extrapolation is used. The issue is directly addressed by Acerbi and Tasche (2002) in their analysis of discontinuous distributions, and by Artzner et al. (1999) in the risk-measure coherence problem.
For a heavy-tailed or mixed loss model, calculating a very high ES additionally requires enough information about losses beyond the quantile, not just a precise threshold location. It is possible for a simulation to estimate the 95th percentile with relative stability while severely underrepresenting the contribution of a very low-probability catastrophic layer to expected shortfall. This is a structural feature of tail averaging, not a numerical anomaly that can be removed by displaying more decimal places.
Model Calibration, Structural Uncertainty, and Sensitivity
Sampling precision is conditional on a fixed model. Calibration adds parameter uncertainty, often estimated from finite historical observations. For a fitted frequency lambda_hat or tail index xi_hat, simulating very many scenarios with that parameter held fixed approximates the fitted model's output more closely; it does not narrow the confidence interval for lambda or xi established by the original data. A principled study separates conditional simulation variance from estimation uncertainty and from structural disagreement among model classes.
One approach to parameter uncertainty is nested simulation: draw plausible parameter values from a documented inferential distribution, then simulate losses conditionally on them. Its unconditional output reflects the chosen parameter-uncertainty representation. The resulting predictive distribution may be useful for decisions, but its interpretation depends on the statistical paradigm and prior assumptions, the way dependence is handled, and the difference between parameter risk and process risk. Reporting a combined simulation percentile without documenting the layers can disguise what has actually been estimated.
Sensitivity analysis must examine the variables that matter for the decision. A tail probability can be sensitive to claim inflation, frequency overdispersion, catastrophe correlation, reinsurance recovery impairment, and low-data severity tails. A mean estimate can be relatively insensitive to certain copula choices while the 99.5th-percentile aggregate loss changes sharply. It is therefore useful to report both local parameter sensitivity and comparisons among plausible structural alternatives. The assessment of dependence uncertainty in risk aggregation by Embrechts, Puccetti, and Rüschendorf (2013) supports treating joint-model choice as a substantive part of risk estimation.
A further distinction is between forecasting and conditional stress evaluation. A stress scenario may have an economically consequential trajectory with an unknown, scientifically unjustified probability. It can still be simulated to analyze consequences or liquidity needs, but assigning an arbitrary probability and incorporating the result into a probabilistic VaR estimate creates an ungrounded precision claim. Scenario analysis, historical stress, sensitivity testing, and probability-based simulation answer connected but distinct questions.
Calibration bias can arise from omitted unreported claims, threshold reporting, large-loss censoring, changing coverage, reinsurance reconstruction, or exposure-selection effects. When observations reflect claims payments rather than underlying hazard intensities, contractual changes can be mistaken for a hazard trend. A reproducible random seed does not correct such biases. Good documentation retains the source population, loss definition, observation period, transformations, exclusions, censoring treatment, and reasons for choosing the tail model.
Heavy-tail extrapolation is especially dependent on assumptions outside the observed range. Embrechts, Klüppelberg, and Mikosch's extreme-value methods (1997) provide rigorous limiting structures for maxima and threshold exceedances, while their use in a specific insurance sample requires verification of threshold stability and other regularity assumptions. A graph showing fit to central observations cannot establish reliability at a probability level far below the inverse sample size.
Insurance Capital, Risk Governance, and Interpretation
Simulated losses can inform capital adequacy, pricing, reserving, reinsurance retention, liquidity planning, or contingency scenarios. These uses have different targets. A pricing calculation may prioritize expected discounted covered payments plus expenses and a risk load; a solvency assessment may focus on adverse aggregate outcomes over a defined horizon; a liquidity assessment asks when cash is due, even when ultimate discounted liability is acceptable. It is unsound to treat one annual VaR output as a complete answer to all three.
The International Association of Insurance Supervisors' Insurance Core Principles and ComFrame include ICP 16, which describes an enterprise risk management framework for solvency purposes and the role of quantitative risk techniques, an own risk and solvency assessment, and scenario analysis. The standard establishes supervisory expectations at the international level; it does not prescribe a universal Monte Carlo sample size, variance-reduction algorithm, or specific insurer model. Local legal force and supervisory implementation depend on jurisdiction. The appropriate institutional link between a numerical study and its governance decision should therefore be explicitly established for each use.
A numerical capital result can be sensitive to the decision rule. Suppose a high loss quantile determines a capital buffer: the appropriate output is a distribution of annual net loss under the applicable capital framework and valuation assumptions. An expected-shortfall measure instead weights the magnitude of losses across the upper tail. The coherent-risk analysis of Artzner et al. (1999) and the expected-shortfall properties of Acerbi and Tasche (2002) explain why these measures can differ in diversification behavior and interpretation. Neither functional should be treated as evidence of loss probability model quality.
A risk committee needs a decision-relevant report rather than an unqualified number. A useful report identifies the estimand, exposure and time horizon, source data date, fixed or uncertain model assumptions, central estimates and tail metrics, numerical confidence, plausible model range, stress outcomes, implemented mitigations, limitations, and explicit decisions requiring review. If one assumption dominates the result, it deserves more explanatory space than a long table of Monte Carlo digits. If a model's largest exposure is a thinly observed catastrophe tail, reporting that fact may be more valuable than extending a low-sensitivity ordinary-loss simulation.
The governance distinction between a computed risk and a chosen risk tolerance is fundamental. Simulation can estimate consequences of a specified risk appetite or retention threshold under a model. It cannot establish a socially, legally, or financially legitimate appetite from mathematics alone. Such choices additionally require policyholder obligations, financing costs, contractual rights, governance accountability, and supervisory requirements. Risk management remains a decision process informed by numerical evidence.
Convergence and Reproducibility
Simulation Validation and Reproducibility
Validation begins with identities the program must satisfy. If a simulated compound-Poisson mean disagrees materially with lambda E[X] after appropriate allowance for sampling error, the discrepancy must be investigated before attention shifts to a remote quantile. Simplified limiting cases should also behave predictably: zero frequency implies zero aggregate claims, a zero deductible recovers the base payment transformation, a limit below a deductible produces zero covered payment under the specified formula, and an independent model should reproduce factorization identities where assumed.
Verification and validation answer different questions. Verification asks whether the algorithm and implementation faithfully compute the specified mathematical model. Validation asks whether that model and its assumptions are suitable for the intended real-world risk question. Unit tests, independent programming implementations, numerical convergence studies, and analytic checks support verification. Data-quality investigation, comparison with held-out periods, expert scrutiny of unobserved extremes, and scenario plausibility assessments support model validation. A model can pass extensive code tests and still be unsuitable for a changed climate, legal, or market regime.
A reproducibility record should contain model purpose; mathematical and contractual specifications; input datasets or an immutable, governed description of restricted inputs; data transformations; distribution and copula versions; generator and seed; code and dependency versions; hardware-sensitive numerical options; number of replications; sampling scheme and weights; estimator definitions; intermediate diagnostic summaries; and final outputs with uncertainty. Access controls may protect confidential claims data, but restricted access should not be confused with the absence of provenance. Independent reviewers should be able to recreate a synthetic test even when live policyholder data are unavailable.
Consistent use of random streams is essential for parallel simulations and comparisons. Jamroz and Williams (2020) discuss reproducible Monte Carlo replicates for distributed uncertainty studies. Their scientific setting is measurement uncertainty, so the transfer here is methodological rather than an assertion of insurance-specific empirical findings. When two alternatives are evaluated under common random numbers, the paired difference may have a smaller variance than the difference of independent sample means, but its standard error must use the pairing structure rather than a false independence assumption.
Numerical diagnostics should include sample means and variances where moments exist, an event-count table for relevant thresholds, coefficient of variation or relative error for target estimators, a check of importance-sampling weight concentration, variation across independent replications, and sensitivity to sample size or algorithmic tuning. For quantiles and tail averages, diagnostic work should examine exceedance counts and the effect of one or a few largest observations. A single narrow-looking confidence interval generated by a questionable estimator is a prompt for investigation, not evidence of success.
The underlying software must be tested in the same execution environment in which the result is prepared. Numerical issues such as overflow, cancellation, finite precision in extreme likelihood ratios, random-stream collisions, and non-deterministic reductions can affect results. For extremely small probabilities, log-likelihood calculations and numerically stable normalization may be necessary. A calculation reported to six decimal places should not imply that its true accuracy exceeds its error analysis.
Comparative Evaluation of Numerical Methods
| Method | Primary target and advantage | Essential conditions | Main diagnostic vulnerability |
|---|---|---|---|
| Ordinary Monte Carlo | Means, frequencies, path outputs; simple unbiased estimators | Correct draws, integrability, independent replicates as required | Sparse exceedances in remote tails |
| Stratification / Latin hypercube | More even input-space coverage, often lower variance | Known strata and valid weights; suitable response structure | Tail concentration not assured |
| Antithetic or control variates | Lower cost-adjusted variance via known structure | Valid pairing or known control expectation | Misspecified control or invalid error calculation |
| Importance sampling | Rare probabilities or targeted integrals | Proposal support, exact likelihood weights, finite useful moments | Extreme weights, proposal mismatch, code bias |
| Conditional Monte Carlo | Smoothed event/payoff estimates | Tractable, correct conditional expectation | Wrong conditional law or costly conditioning |
| Splitting / sequential rare-event design | Paths to rare system states | Valid level/weight construction and path dependence accounting | Particle degeneration, dependent offspring |
| Quantile order statistics | Empirical loss quantiles | Adequate tail mass; regular local CDF for normal approximations | Low density, atoms, unstable threshold |
| Expected-shortfall scenario optimization | Tail average and capital optimization | Integrable losses, correct mass and weight treatment | Underobserved extreme severity |
| Extreme-value extrapolation | Inference beyond observed thresholds | Tail regularity, stable threshold and calibration | Structural extrapolation and parameter uncertainty |
There is no estimator that dominates for every risk quantity and distribution. The source-based contributions of Rubinstein and Kroese (2016), Glasserman (2003), Glynn and Iglehart (1989), and Asmussen and Kroese (2006) establish a family of methods with distinct mathematical performance criteria. An appropriate method is chosen by the target quantity, the rarity and structure of the event, available analytic conditioning, tail class, dependence mechanism, implementation cost, and the ability to diagnose error.
For an initial model, elementary analytic checks and a small independent simulation may be the most informative first step. For a moderately remote quantile, ordinary Monte Carlo with explicit exceedance counts may be sufficient. For a more distant threshold, importance sampling or conditioning may reduce numerical uncertainty, but only after source-to-model assumptions are re-established. Where the tail law itself is weakly identified, additional observations, alternative model comparisons, or explicit decision robustness may have greater scientific value than further computational acceleration.
Research Limitations and Priorities
This study examines numerical estimation under specified distributions and draws internationally relevant examples from insurance and risk management. It uses an original controlled synthetic case rather than proprietary insurer data, and it does not purport to estimate real insurer solvency probabilities, prove the accuracy of any operational catastrophe model, or provide legal guidance for particular capital regimes. Original research cited here was checked against available publisher pages, author or institutional records, and official standard-setter materials. Publisher previews and abstracts are not represented as full-text examination of subscription works; substantive claims are confined to supported results and elementary independently derived mathematics.
Several research questions follow from the analysis. One is the development and independent validation of importance-sampling proposals for portfolios with heterogeneous severe claims, contractual nonlinearities, and common shocks. Another is the comparison of tail quantile estimators under distributional atoms and low observed exceedance counts. A third concerns combined numerical and model uncertainty in the presence of nonstationary hazard intensities and evolving exposure. A fourth is the challenge of making advanced variance-reduction diagnostics intelligible to boards, auditors, and regulators without flattening their mathematical assumptions.
These are open research priorities within established probability, statistics, actuarial science, and risk governance. They are not declarations of a new theoretical system. Scientifically credible progress would require published assumptions, tractable benchmarks or controlled experiments, access to appropriate exposure data, independent replication, and rigorous comparison with competing estimators and model specifications.
Conclusion
Simulation and numerical risk evaluation constitute a disciplined method for computing consequences of uncertainty under an explicit probabilistic description. A model determines the loss process; an estimator targets a defined functional; computation produces an approximation; and diagnostics characterize its sampling precision. Importance sampling, conditioning, stratification, and specialized rare-event methods improve numerical efficiency when their mathematical requirements are met. Quantiles and expected shortfall require additional care at discontinuities and in regions supported by very few observations.
The exact compound-Poisson case demonstrated that severe losses can represent a small share of claim counts while contributing a large share of annual variance; its simulation provided an auditable comparison of analytic and random estimates. The Bernoulli change-of-measure example quantified the potential efficiency gain of a correctly weighted estimator. The three-point loss counterexample established that a stable percentile can coexist with substantial upper-tail severity and that the treatment of probability mass at a quantile is essential to expected-shortfall calculation. These findings are direct deductions from stated models and independent numerical checks.
The principal synthesis is methodological: a larger simulation does not repair an incorrect event definition, unsupported dependence assumption, wrong likelihood ratio, defective tail model, or uncertain exposure data. Risk estimates become fit for scholarly and institutional interpretation when numerical precision, structural validity, empirical calibration, and the decision's time and contractual context are documented separately and assessed together. The practical standard of evidence is traceability from the real-world question to the mathematical model, from the model to the computation, and from the computation to a clearly delimited conclusion.
References
Acerbi, C., & Tasche, D. (2002). On the coherence of expected shortfall. Journal of Banking & Finance, 26(7), 1487–1503. https://doi.org/10.1016/S0378-4266%2802%2900283-2
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
Asmussen, S., & Kroese, D. P. (2006). Improved algorithms for rare event simulation with heavy tails. Advances in Applied Probability, 38(2), 545–558. https://doi.org/10.1239/aap/1151337084
Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling Extremal Events for Insurance and Finance. Springer. https://doi.org/10.1007/978-3-642-33483-2
Embrechts, P., Puccetti, G., & Rüschendorf, L. (2013). Model uncertainty and VaR aggregation. Journal of Banking & Finance, 37(8), 2750–2764. https://doi.org/10.1016/j.jbankfin.2013.03.014
Glasserman, P. (2003). Monte Carlo Methods in Financial Engineering. Springer. https://doi.org/10.1007/978-0-387-21617-1
Glasserman, P., Heidelberger, P., & Shahabuddin, P. (1999). Asymptotically optimal importance sampling and stratification for pricing path-dependent options. Mathematical Finance, 9(2), 117–152. https://doi.org/10.1111/1467-9965.00065
Glasserman, P., Heidelberger, P., & Shahabuddin, P. (2000). Variance reduction techniques for estimating value-at-risk. Management Science, 46(10), 1349–1364. https://doi.org/10.1287/mnsc.46.10.1349.12274
Glynn, P. W., & Iglehart, D. L. (1989). Importance sampling for stochastic simulations. Management Science, 35(11), 1367–1392. https://doi.org/10.1287/mnsc.35.11.1367
International Association of Insurance Supervisors. (2024; online tool accessed 2026). Insurance Core Principles and Common Framework (ComFrame), ICP 16: Enterprise risk management for solvency purposes. https://www.iais.org/icp-online-tool/
Jamroz, B. F., & Williams, D. F. (2020). Consistency in Monte Carlo uncertainty analyses. Metrologia. https://www.nist.gov/publications/consistency-monte-carlo-uncertainty-analyses
Joint Committee for Guides in Metrology. (2008). JCGM 101:2008: Evaluation of measurement data—Supplement 1 to the Guide to the expression of uncertainty in measurement—Propagation of distributions using a Monte Carlo method. BIPM. https://doi.org/10.59161/JCGM101-2008
McKay, M. D., Beckman, R. J., & Conover, W. J. (1979). A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics, 21(2), 239–245. https://doi.org/10.1080/00401706.1979.10489755
McNeil, A. J., Frey, R., & Embrechts, P. (2015). Quantitative Risk Management: Concepts, Techniques and Tools (revised ed.). Princeton University Press. https://press.princeton.edu/books/hardcover/9780691166278/quantitative-risk-management
Rockafellar, R. T., & Uryasev, S. (2000). Optimization of conditional value-at-risk. Journal of Risk, 2(3), 21–41. https://doi.org/10.21314/JOR.2000.038
Rubinstein, R. Y., & Kroese, D. P. (2016). Simulation and the Monte Carlo Method (3rd ed.). Wiley. https://doi.org/10.1002/9781118631980