RESEARCH DOSSIER
Value at Risk and Expected Shortfall
Abstract. Value at Risk (VaR) and Expected Shortfall (ES) summarize different features of a probability distribution of monetary loss. VaR identifies a loss quantile at a chosen confidence level; ES averages quantiles in the remaining upper probability tail. The distinction governs portfolio comparisons, diversification, estimation, insurance capital, banking regulation, and the interpretation of unusually severe outcomes. This monograph develops both measures from a common loss-variable convention, derives their distributional properties, reconciles continuous and discrete formulations, and distinguishes quantile threshold exceedances from properly probability-weighted tail averages. Exact constructions show that two portfolios can reverse their risk ranking when VaR is replaced by ES, that VaR can violate subadditivity for independent rare-event losses, and that ES restores subadditivity under the standard integrability assumptions while remaining sensitive to tail-model error. Further analysis examines elliptical cases where VaR aggregation behaves well, mixtures, truncated claims, heavy-tailed severities, sample-quantile conventions, scenario optimization, backtesting, jointly consistent scoring, capital applications, and model governance. Regulatory evidence distinguishes the one-year 99.5 percent VaR calibration of European insurance solvency from the stressed 97.5 percent ES component of the Basel bank trading-book internal-model framework. The original contribution is a source-audited synthesis, exact worked comparisons, and an evaluation of what each statistic can and cannot support in risk decisions; no new proprietary risk measure, theory, or regulatory doctrine is introduced.
Keywords: Value at Risk; VaR; Expected Shortfall; ES; Conditional Value at Risk; CVaR; risk quantiles; tail expectation; quantile integration; discrete loss distributions; subadditivity; coherent risk measures; risk aggregation; catastrophe loss; heavy-tailed distributions; tail estimation; backtesting; regulatory capital; Solvency II; Basel market risk.
Introduction
The measurement problem begins with a distribution of losses associated with a precisely defined portfolio, observation date, valuation convention, and horizon. A loss quantile answers a threshold question: how large a monetary loss lies at the chosen percentile? The expected loss in an extreme tail answers a different question: after a fraction of the distribution has been selected by rank, what is its average loss? Both questions have practical importance, and their answers can rank the same assets, policies, or portfolios in opposite orders. A measure becomes meaningful only when its random variable, probability law, and aggregation unit are explicit.
The intellectual foundations are well established. The work of Artzner and colleagues (1999) identified axiomatic properties important for evaluating capital risk measures and showed why a quantile statistic may fail to reward diversification. The treatment by Acerbi and Tasche (2002) resolved a recurrent conceptual trap: formulations of expected shortfall that agree for continuous distributions can diverge at an atom. The general-distribution CVaR analysis of Rockafellar and Uryasev (2002) developed the related optimization formulation in the presence of finite scenarios and discontinuities. Together these results permit an exact comparison of quantile and tail-mean risk without conflating their mathematical definitions.
Insurance applications make the contrast especially consequential. Insurance losses often have positive probability at zero, discrete contractual benefit amounts, limits, deductibles, rare catastrophe events, and long-tailed claim severities. A percentile can remain unchanged as a large catastrophe outcome becomes several times worse, provided the catastrophe state stays outside the selected quantile. ES registers the increased tail loss, subject to the information embodied in the assumed distribution. Conversely, percentile thresholds remain useful where a risk decision expressly asks for a coverage or solvency confidence event. Appropriate use depends on the decision's object and on the fidelity of the loss model.
This inquiry uses a single monetary-loss sign convention throughout. All derivations refer to a random variable L for a loss, with positive values indicating adverse outcomes. Sources sometimes use net worth, portfolio profit, or negative returns. Changing a source's sign convention must precede comparison of numerical formulas. The principal confidence parameter is alpha in (0,1); its upper-tail probability is beta = 1 - alpha. When there is no risk-free term or exposure adjustment, quantities are expressed in the same monetary units as L. The symbols are descriptive parameters of standard probability theory, not an introduced proprietary conceptual system.
The method consists of published mathematical definitions, elementary proofs, explicitly reproducible hypothetical loss distributions, and documentary comparison of research and standards. The comparison is international but avoids equating a banking market-risk standard with an insurer's solvency regime. The work does not use confidential market data, assert a new statistical discovery, or infer the adequacy of any particular insurer's capital from one illustrative distribution.
Loss Quantiles
Monetary loss, time horizon, and information set
Let (Omega, F, P) be a probability space and let L be a real-valued loss random variable. Its cumulative distribution is F_L(x) = P(L <= x). When an amount is measured on an insurer's balance sheet, L may be net asset-value deterioration, claims minus planned funding, change in basic own funds with sign adjusted, or net cash requirements. Those quantities can differ even for the same underlying occurrence. Whenever a threshold is stated, the valuation and time horizon must follow it.
A one-day trading loss, one-year change in insurer own funds, and lifetime aggregate of a policy pool are three different random objects. They cannot be compared by changing only alpha. If I_t denotes information available at time t, a forecast may use a conditional law F_{L|I_t} rather than an unconditional historical law. The appropriate distribution depends on the forecast origin, exposure composition, model dynamics, and stress assumptions. The importance of conditioning is supported empirically by the conditional tail-risk estimation study of McNeil and Frey (2000), which combines time-varying volatility estimation with extreme-value modeling.
For insurance losses, L can be represented as the sum of incurred claims, indemnity payments, or liabilities after reinsurance, provided the representation states which recoveries, expenses, and asset changes it includes. Limits and deductibles modify the insured loss distribution even when underlying hazards are unchanged. A cat-event loss might be measured before or after reinstatement premiums and reinsurance counterparty risk. A capital calculation that changes these components while retaining the same VaR or ES label changes the object under study.
Quantiles and probability conventions
Define the lower alpha-quantile of L as
For finite quantiles this is the smallest threshold at which accumulated probability reaches alpha. We define value at risk by VaR_alpha(L) = q_alpha(L). The confidence statement is P[L <= VaR_alpha(L)] >= alpha. With atoms, it may be strictly greater; an equality P[L <= q] = alpha cannot be assumed. Likewise P[L > q] can be smaller than beta.
In many implementations VaR is described as a maximum loss not exceeded with a stated confidence level. That informal statement should be interpreted through the quantile definition: there can be losses greater than VaR, and the probability of exceeding the threshold is no more than beta. The measure says nothing about their magnitude. It should also not be confused with a maximum physically or contractually possible loss; its threshold may be far below a known contractual aggregate limit.
Quantile conventions are consequential around ties. A statistical package may interpolate between adjacent ordered observations, while the mathematical definition above chooses the first threshold with cumulative probability at least alpha. Alternative conventions have uses in statistical estimation, but direct verification of a theoretical example must use its stated convention. If two distributions are compared, alpha, the direction of loss, the horizon, and the quantile convention must all be held fixed.
The geometry of probability mass and tails
For a continuous strictly increasing cumulative distribution, q_alpha has F_L(q_alpha) = alpha, and the event L > q_alpha has probability beta. In this case the average of losses strictly beyond the threshold is well defined when the relevant expectation is finite, and it equals ES_alpha as developed below.
For a discrete loss, there may be a mass m = P[L=q_alpha] > 0. Let r = P[L>q_alpha]. Because the selected upper tail has probability beta but r can be smaller than beta, only beta-r of the probability mass located exactly at q_alpha is needed to complete the upper beta-tail. Averages over all observations satisfying L >= q_alpha instead use more probability than the required tail; averages over L > q_alpha instead use less. Both may differ from ES. This distinction is the central result of the coherence analysis for discontinuous losses by Acerbi and Tasche (2002).
The integrated-quantile expression gives an unambiguous upper-tail averaging operation: every value is weighted by its actual share of the highest beta fraction of probability, including a fractional amount at a threshold atom when required. In simulations this can be implemented by sorting weighted outcomes and taking exactly beta total probability from the worst-ranked outcomes. The operation remains well defined even if several scenarios share the same loss amount.
Economic meanings and decision boundaries
A VaR number is a percentile of a specified loss distribution. Its financial interpretation depends on whether the number represents a negative capital position, an unanticipated deviation from expected loss, or an absolute cash outflow. A threshold can support a question about the probability of breaching a capital buffer. It is not alone an evaluation of the severity of the deficit conditional on breach, an expected funding shortfall, or the welfare implications of failure.
Risk appetite statements often set an upper quantile limit for a portfolio. Such limits can be operationally clear but remain sensitive to the confidence level and distribution. A second statement about ES can illuminate what occurs in the selected tail. The two statistics may be reported together without combining them into a novel metric. Comparisons across lines of business additionally require consistent accounting for insurance liabilities, dependence, reinsurance recoveries, and the aggregation horizon.
An elementary finding follows directly: if two loss distributions have exactly the same values up to their alpha-quantiles but different severities above, their VaR_alpha values can be identical while their ES_alpha values differ. This is a mathematical observation about the information content of a quantile. It does not depend on an institution's choice of regulatory regime.
Value at Risk
Formal definition and elementary properties
For a fixed alpha, VaR is monotone: if L_1 <= L_2 almost surely, then q_alpha(L_1) <= q_alpha(L_2). It is translation equivariant: for deterministic c, VaR_alpha(L+c) = VaR_alpha(L)+c. It is positively homogeneous for nonnegative t: VaR_alpha(tL) = t VaR_alpha(L), with standard qualifications at t=0. These properties follow from distributional transformations and do not require a smooth density.
A general subadditivity assertion, VaR_alpha(L_1+L_2) <= VaR_alpha(L_1)+VaR_alpha(L_2), fails even with independent finite-support nonnegative losses. That failure is analyzed in the research on coherent measures of risk. It is important to state precisely what fails: the inequality is not universal over all joint loss laws. Some restricted model families satisfy it, and diversified portfolios do not automatically produce a VaR increase. A universal rejection and a universal endorsement of quantile aggregation would each obscure the conditional mathematics.
If alpha is chosen at a level near one, two different models may both return a modest VaR for portfolios dominated by rare events. The probability of an event that lies beyond the quantile can change, and its size can change, without altering the reported threshold until a probability boundary is crossed. A discontinuous quantile response is mathematically natural at mass points. In risk governance it makes confidence-level sensitivity an important complementary calculation.
Location-scale and normal-distribution results
Suppose L = mu + sigma Z, sigma > 0, and Z has a fixed reference distribution with alpha-quantile z_alpha. Then VaR_alpha(L) = mu + sigma z_alpha. For a standard normal Z, z_alpha = Phi^{-1}(alpha). At alpha = 0.95, z_alpha is approximately 1.64485; at alpha = 0.99 it is approximately 2.32635. These values are distribution properties, not universal multipliers of all insurance losses.
For a normal loss with mean mu and standard deviation sigma, both the mean and the dispersion matter. If L has mean 40 and standard deviation 20 monetary units, VaR_0.95 is approximately 72.897. This calculation presumes the normal law and includes whatever negative-loss outcomes the normal distribution permits. Modeling a nonnegative insurance claim amount by an unrestricted normal distribution can therefore misrepresent low-loss behavior as well as a catastrophic right tail.
Location-scale invariance yields useful analytical checks of code. If a deterministic expense k is added to all simulated claims, an implementation should raise both VaR and ES by k. If all monetary amounts are multiplied by a positive exchange-rate or indexation factor, the reported measures should scale by that factor. Failure of either check indicates a convention, implementation, or unit problem, not a nuanced disagreement about risk theory.
Empirical quantiles and quantile estimation
Given n ordered observed losses L_(1) <= ... <= L_(n), a direct empirical quantile for the left-continuous inverse distribution is L_(ceil(n alpha)). It estimates a population quantile under appropriate sampling assumptions; the sample definition is not a proof of a forecasting model's validity. Temporal dependence, regime change, changing underwriting standards, claim development, and incomplete catastrophe records can invalidate simple empirical interpretation.
With 250 independent forecasts assessed at alpha = 0.99, the nominal expected number of threshold exceptions is 250 times 0.01 = 2.5. Under the idealized Bernoulli-exception model the probability of zero exceptions is 0.99^250, approximately 0.081. Thus zero observed exceptions alone does not demonstrate exact calibration. At more extreme insurance quantiles, the amount of effective data in the relevant tail is even smaller, and estimation typically depends on extrapolation or explicit modeling assumptions.
Estimation should distinguish a historical distribution of past realizations from the prospective conditional law of the next loss. The extreme-value conditional estimation approach of McNeil and Frey (2000) illustrates why volatility dynamics and tail shape both matter for financial-risk estimation. For insurance, analogous challenges arise from sparse catastrophe observations, censoring, inflation, changing exposure, risk mitigation, and dependence among claims.
VaR as a ranking and capital statistic
The quantile permits a clear threshold ordering: a portfolio with VaR_0.95 = 10 requires a threshold of 10 to achieve at least 95 percent coverage of its modeled monetary loss, while a portfolio with VaR_0.95 = 0 has at least 95 percent probability at zero loss or below. The ranking does not specify which portfolio has a more severe remaining 5 percent tail. A decision-maker applying a quantile-only criterion is selecting a particular feature of the distribution.
If VaR is used as a proxy for capital, one must specify whether expected loss is funded separately, whether capital absorbs loss from zero or only the unexpected portion, and whether insolvency is defined by net asset value rather than claims paid. The 99.5 percent insurance capital criterion in the European Solvency II Article 101 refers to the VaR of basic own funds under the legal framework, not a generic quantile of gross claim severity. Numerical examples in this monograph use generic losses and should not be transplanted directly into supervisory solvency calculations.
Portfolio aggregation and dependence
A portfolio loss is S = sum_i L_i. Its VaR depends on the entire joint distribution of the component losses. Marginal quantiles do not suffice to determine the distribution of S. In particular, knowing VaR_alpha(L_i) for each component without their dependence or scenario pairing generally cannot determine VaR_alpha(S). This is particularly important for geographic insurance concentrations, shared infrastructure, and correlated asset-liability stress.
Even when marginal claim probabilities are identical, two lines may aggregate very differently under independence, positive association, or a common catastrophe shock. The aggregation question should be separated from a question of capital allocation: the latter assigns a portfolio-wide amount to components and can depend on differentiability, decomposition conventions, and incentives. That broader allocation problem belongs to a separate Research Dossier, while the present work uses aggregation to compare VaR and ES.
For jointly normal losses with finite covariance matrix and alpha >= 0.5, the quantile coefficient z_alpha is nonnegative. The standard deviation of a sum is at most the sum of component standard deviations by the triangle inequality in L2. It follows that VaR_alpha(L_1+L_2) <= VaR_alpha(L_1)+VaR_alpha(L_2) for jointly normal losses at those confidence levels. This restricted result coexists with the finite-support failure below. The relevant modeling assumption, not the name VaR alone, determines whether the desired aggregation property holds.
Expected Shortfall
The integrated-quantile definition
For an integrable monetary loss L and alpha in (0,1), expected shortfall is
The integral of the upper quantile curve has the units of loss. Its mathematical value is an average over a probability mass of exactly beta = 1-alpha, chosen from the worst-ranked outcomes. This definition accommodates continuous, discrete, and mixed distributions. If positive-tail integrability fails, ES can be infinite even when VaR at every strictly subunit confidence level is finite.
When the loss distribution is continuous at q_alpha and the event L > q_alpha has probability beta, the expression equals E[L | L > q_alpha] under ordinary finite-expectation conditions. The conditional-exceedance interpretation is therefore a special case, not a general definition. In an insurance setting with point masses, the integrated-quantile definition is the safer starting point. The foundational comparison in Acerbi and Tasche (2002) explicitly investigates this issue.
ES_alpha(L) >= VaR_alpha(L) for an integrable loss because q_u(L) is nondecreasing in u and the upper beta average cannot fall below its starting quantile. This inequality compares measures of the same loss variable at the same confidence alpha. A numerical comparison of VaR_0.995 and ES_0.975 involves different confidence levels, and neither numerical ordering is guaranteed solely by this inequality.
Atoms, conditional tails, and equivalent formulas
Let q = VaR_alpha(L), r = P(L > q), and beta = 1-alpha. Suppose E[L_+] is finite. The upper-tail average is
The second term supplies precisely the portion of the probability mass at q needed to complete the selected beta tail. An equivalent formula is
Both expressions remain valid for mass points, provided q is the defined lower quantile and expectations exist. They show why a simple sample average of all losses at or above q can be wrong: its denominator is not necessarily beta. The general loss-distribution treatment of Rockafellar and Uryasev (2002) provides the coherent CVaR formulation for precisely such distributions.
A practical weighted-scenario algorithm sorts scenarios from the greatest loss to the smallest. It then accumulates exactly beta total probability, taking part of the final tied scenario probability if necessary. If scenarios have equal probability and beta n is an integer, the result is the average of the worst beta n simulated outcomes. If beta n is not an integer, a fractional weight is required at the boundary to match the empirical quantile integral. The estimate remains subject to sample uncertainty and sensitivity to how scenarios were generated.
Subadditivity and risk-measure structure
Under standard integrability assumptions, ES is monotone, translation equivariant, positively homogeneous, and subadditive for losses: ES_alpha(X+Y) <= ES_alpha(X)+ES_alpha(Y). These are the familiar coherence properties stated in monetary-loss notation. The abstract risk-measure framework was developed in Coherent Measures of Risk, while Expected Shortfall as a coherent alternative and the distributionally precise work of Acerbi and Tasche (2002) establish the relevance of ES.
A useful proof route employs an upper-tail optimization representation. For integrable L,
The bounded density Z selects and weights an adverse upper fraction of the distribution. For any X and Y, the expectation E[Z(X+Y)] equals E[ZX]+E[ZY], and its supremum over a common admissible set is no greater than the sum of the separate suprema. This proves subadditivity. Translation equivariance follows because E[Z]=1, and positive homogeneity follows for nonnegative scaling. The representation also clarifies that ES is a worst-tail mean under a constrained probability reweighting, rather than an unconditional expectation of all claims.
Subadditivity prevents the sum's ES from exceeding the sum of stand-alone ES values in a consistently modeled system. It does not guarantee that adding a new exposure lowers absolute portfolio tail loss, and it does not provide insurance policyholders with a guaranteed reduction in premium or insolvency probability. It is a structural property of the functional applied to a given joint loss law.
Scenario optimization and dual interpretations
Rockafellar and Uryasev's CVaR optimization approach expresses ES as the minimum over a threshold eta of
For a suitable integrable L, min_eta g_alpha(eta) = ES_alpha(L); alpha-quantiles are minimizers under the usual generalized quantile conditions. For finite scenarios with probability weights w_j and loss functions L_j(x), this becomes
When the decision-dependent scenario losses L_j(x) are affine, these constraints form a linear-programming-compatible objective. This is a substantive advantage for optimizing exposures and hedges. The 2002 general-distribution analysis explains how the representation survives the discreteness common to scenarios and finite samples.
The representation must not be mistaken for automatic tractability of all insurance decisions. Nonlinear claims functions, discontinuous reinsurance triggers, endogenous behavior, constraints on indemnity amounts, and nonconvex portfolio requirements can defeat a direct linear-program formulation. The mathematical convenience derives from the joint structure of the loss functions and the chosen decision set, not solely from the name ES.
Statistical existence and heavy-tail behavior
A finite quantile may coexist with an infinite upper-tail expectation. For a Pareto-type loss whose tail has shape exponent k, an upper tail with k <= 1 has infinite mean even though every fixed alpha-quantile with alpha < 1 is finite. Therefore VaR_alpha can remain a finite number while ES_alpha = infinity. This distinction matters for insurance catastrophe modeling where severity distributions can be heavy-tailed and fitted tail indices uncertain.
A generalized Pareto approximation above threshold u uses scale sigma_u > 0 and shape xi. If xi < 1 and a quantile q lies in the modeled exceedance region, the conditional expected excess beyond q is (sigma_u + xi(q-u))/(1-xi); hence, under the continuous exceedance interpretation, ES at a qualifying upper quantile equals
The expression diverges as xi approaches 1 from below. This calculation assumes the generalized Pareto tail remains valid beyond q, an assumption requiring scrutiny. The empirical merits of combining volatility filtering and extreme-value estimation are discussed by McNeil and Frey (2000). Insurance severity data introduce additional complications including changing claims inflation, deductibles, coverage limits, and selection in recorded claims.
The existence condition is not mere mathematical decoration. In heavy-tailed models a small change in a fitted tail-shape parameter can make the estimated ES very large or infinite while leaving a fixed quantile finite. A numerical system may return a finite sample ES simply because no infinitely large loss was observed; that finite output does not establish finite population ES.
Continuous and Discrete Losses
Continuous uniform and normal examples
Take L uniformly distributed on [0,100]. Its alpha-quantile is 100 alpha, and integration gives ES_alpha(L) = 50(1+alpha). At alpha = 0.95, VaR = 95 and ES = 97.5. Both values are directly checkable without simulation. The tail expectation is the midpoint of 95 and 100 because the top five percent occupies a uniform interval.
For L = mu + sigma Z with standard-normal Z, where sigma > 0, the ES formula is
where phi and Phi denote the standard normal density and distribution and z_alpha = Phi^{-1}(alpha). At alpha = 0.95, phi(z_alpha)/0.05 is approximately 2.0627, while z_alpha is approximately 1.64485. For the preceding example with mu = 40 and sigma = 20, VaR_0.95 is approximately 72.897 and ES_0.95 is approximately 81.254. Both calculations rely on finite normal moments.
These examples establish an intuitive relation: ES averages losses in the region beyond a quantile, and generally therefore exceeds the threshold. The size of the gap depends on tail dispersion and shape. A thin-tailed distribution and a heavy-tailed distribution may have identical VaR at one alpha but radically different ES values.
Discrete insurance claim cases
Consider a single policy with L_A = 0 with probability 0.95 and L_A = 100 with probability 0.05. At alpha = 0.95 the lower quantile is VaR_0.95(L_A)=0 because F_A(0)=0.95. The largest five percent of the distribution consists of losses of 100, so ES_0.95(L_A)=100. The naive conditional average E[L_A | L_A >= VaR] is E[L_A]=5, which is not ES. The conditional average E[L_A | L_A > VaR] does happen to equal 100 here, but that agreement is contingent on the event L_A > 0 having probability exactly beta.
Take a second claim L_B=10 with certainty. Then VaR_0.95(L_B)=10 and ES_0.95(L_B)=10. VaR ranks A ahead of B because 0 < 10, whereas ES ranks A behind B because 100 > 10. The ranking reversal is exact, with no estimates or simulation error. It demonstrates the intrinsic distinction between a probability-threshold decision and a tail-severity decision, rather than a universal preference rule for every risk context.
Now consider a third loss L_C with P(L_C=0)=0.96, P(L_C=100)=0.04. At alpha = 0.95, VaR_C=0, but ES_C=(0.04 times 100 + 0.01 times 0)/0.05=80. The strict-exceedance conditional mean E[L_C|L_C>0] equals 100, not 80. This example shows why the two conditional formulations need not agree even when the threshold is zero. Exactly one percentage point of zero-loss mass must complete the worst five percent tail.
Tail severity and sensitivity experiments
Keep P(L_C=100)=0.04 but increase the loss of the adverse state to 500. VaR_0.95 remains zero, because the cumulative probability at zero remains 0.96. ES_0.95 rises from 80 to 400. The comparison isolates severity sensitivity: probability is fixed, quantile threshold is unchanged, and the mean of the worst five percent rises fivefold. No conceptual inference about real-world claim size follows from a fabricated monetary example.
Alternatively keep severity at 100 and vary adverse-event probability p. At alpha = 0.95, VaR is zero for p <= 0.05 and jumps to 100 for p > 0.05 under the specified two-point distribution. ES is 100 p/0.05 when p <= 0.05 and 100 when p > 0.05. ES changes continuously in p through the boundary, although its derivative changes there; VaR exhibits a discontinuity. This is one reason model perturbations around a mass-point threshold should be separately reported.
A simple stress test can also move the confidence level. For P(L=0)=0.96 and P(L=100)=0.04, VaR_0.95=0, VaR_0.99=100, ES_0.95=80, and ES_0.99=100. A change from alpha=0.95 to alpha=0.99 is not a harmless relabeling: the selected tail shrinks from five percent to one percent. Comparing risk estimates without stating alpha obscures both what was measured and which outcomes were averaged.
Mixed distributions, deductibles, and limits
Many claim distributions contain a probability mass at zero and a continuous distribution of positive payment amounts. For example, a policy may have no paid claim during a year with probability 0.97, while a claim state has a continuous severity distribution. For alpha=0.95, VaR=0 in this model, but ES is positive whenever the claim severity has finite positive mean. Only three percentage points of positive-claim outcomes fill the worst five percent tail, and the remaining two points are zero-loss outcomes. A calculation based solely on the severity conditional on claim incidence overstates the correct ES for this alpha if it neglects the zero portion.
A deductible transforms underlying severity X into a payment Y=(X-d)+. The payment distribution acquires an atom at zero equal to P(X<=d). A policy limit M further transforms Y into min((X-d)+,M), generating a possible mass at the maximum payout M. Both atoms affect percentile and tail calculations; the integrated-quantile definition automatically handles them. A loss model incorporating reinsurance attachment and exhaustion has analogous discontinuities or mass concentrations.
These transformations explain why continuous approximations should be audited against the actual contract's payment function. The gross severity and the insured payment can have different VaR, ES, tail indices, and dependence. A statistical model may be scientifically valid for gross losses while unsuitable for net insurer capital if its contract layer is implemented incorrectly.
Finite samples and Monte Carlo implementations
For equal-weight n-scenario samples, define the empirical distribution assigning mass 1/n to each observed loss. VaR_alpha is the observation with rank ceil(n alpha). ES_alpha is the integral of the empirical quantile curve; implementation must weight the highest observations so that the total selected mass is exactly beta. If n beta is not an integer, the boundary observation receives a fractional weight. If observations are tied, the selected share can be drawn from any of the tied identical amounts without changing the result.
For instance, an equally weighted sample of 20 observations with 19 zeros and one loss 100 has VaR_0.95=0 and ES_0.95=100. If an implementation instead returns the mean of observations satisfying L >= VaR, it produces 5; this is an identifiable algorithmic defect relative to the selected ES definition. The same defect can remain hidden in continuous data where exact ties are rare.
Monte Carlo sample size should be assessed against tail probability. An estimate at alpha=0.995 based on 10,000 independent equally weighted scenarios has an average of only 50 observations in the selected half-percent tail. Its stability depends heavily on the behavior of the underlying severity distribution. Increasing sample size helps with numerical error, but cannot by itself correct the wrong loss law, omitted common-shock states, or unrepresentative event sampling.
The computational record should preserve the seed, scenario probabilities, data transformation, quantile convention, handling of scenario ties, estimation horizon, inflation basis, gross/net definition, and uncertainty assessment. When analytical formulas are available, they provide regression tests for a simulation pipeline. This is particularly valuable for isolated policy-layer and catastrophe-event models.
Diversification Counterexamples
Two independent rare-event claims
Let X and Y be independent claims with P(X=100)=P(Y=100)=0.04 and P(X=0)=P(Y=0)=0.96. At alpha=0.95 each has VaR_0.95=0 and ES_0.95=80. The sum S=X+Y has the exact distribution
| Aggregate loss S | Probability |
|---|---|
| 0 | 0.9216 |
| 100 | 0.0768 |
| 200 | 0.0016 |
The probability at or below zero is 0.9216 < 0.95, and the probability at or below 100 is 0.9984 >= 0.95. Hence VaR_0.95(S)=100, while VaR_0.95(X)+VaR_0.95(Y)=0. The subadditivity inequality fails by 100 monetary units. This is a direct, fully enumerated counterexample under independence, with finite nonnegative losses. It illustrates the mathematical issue motivating Artzner and colleagues' coherence framework.
For ES_0.95(S), the worst five percent includes all 0.0016 probability at 200 and another 0.0484 at 100. Consequently
By contrast, ES_0.95(X)+ES_0.95(Y)=160. ES satisfies subadditivity in this example: 103.2 <= 160. This does not say the portfolio's expected aggregate loss falls when claims are added. Indeed E[S]=8, twice the single-claim mean of 4. It says that a coherent tail-risk functional of the combined loss does not exceed the sum of the same functional applied to the components.
Dependence and common shocks
Now replace the independence assumption by perfect concurrence: set X=Y almost surely, with the same marginal probabilities as above. Then S takes values 0 with probability 0.96 and 200 with probability 0.04. At alpha=0.95, VaR(S)=0 and ES(S)=160. The stand-alone ES values remain 80 each, and ES(S)=ES(X)+ES(Y). Dependence removes the ES diversification benefit that appeared under independence, even though a quantile of the perfectly concurrent sum happens to be zero at the specified alpha.
This contrast is a warning against interpreting a VaR zero as a claim of negligible systemic exposure. The event's modeled probability remains four percent, its joint loss increases to 200 when it occurs, and the VaR threshold remains zero because alpha is below the no-loss probability. A catastrophe model incorporating shared windstorm, flood, earthquake, cyber service, or asset-market factors may create analogous accumulations, although empirical magnitudes and dependence patterns must be estimated rather than borrowed from this binary construction.
At the opposite extreme, scenarios with mutually exclusive claims may place aggregate probability mass differently and yield different quantiles. Marginal distributions alone permit multiple couplings. A valid assessment of capital needs therefore requires either a joint probability model or a documented range of aggregation scenarios. The general mathematics of risk-measure aggregation does not identify the correct dependence law.
Conditions for normal-portfolio VaR aggregation
For jointly normally distributed losses X and Y, write their means mu_X and mu_Y and their standard deviations sigma_X and sigma_Y. The aggregate variance equals sigma_X^2 + sigma_Y^2 + 2 Cov(X,Y), and the aggregate standard deviation is at most sigma_X + sigma_Y. At alpha>=0.5, z_alpha>=0. Thus
This proof uses a shared nonnegative quantile coefficient and a joint normal model. Its validity does not extend automatically to a mixture of catastrophe losses and routine attritional claims or to a distribution with rare discontinuous liabilities. VaR's behavior in such portfolios must be derived from the actual aggregate loss law.
The normal result exposes a recurring reasoning error: observing subadditivity in a spreadsheet calibrated to joint normality is not evidence that VaR satisfies the property for every loss distribution. Conversely, observing the binary counterexample is not evidence that VaR always punishes diversification. Mathematical counterexamples disprove universal claims; they do not dictate all empirical portfolio outcomes.
Ranking reversals and insurance decisions
The first discrete case ranked a rare 100-unit claim favorably under VaR_0.95 and unfavorably under ES_0.95 relative to a certain 10-unit loss. A different decision maker may rank these contracts under expected loss, utility, ruin probability, concentration, or an explicit risk budget. Those orderings answer different questions. In particular E[L_A]=5, whereas E[L_B]=10; a risk-neutral expected-loss criterion favors A, consistent with VaR but inconsistent with ES. There is no contradiction because the objectives differ.
This comparison also illustrates why tail-risk reporting should not silently replace actuarial expected claims in pricing. Premium formation considers expected indemnity, expenses, risk loading, policyholder behavior, competition, and financial capacity. ES can be informative about tail-capital exposure but is not an arithmetic estimate of expected claim cost. Likewise, a high ES does not directly imply a high probability of insolvency unless capital, reserve, asset, and claim-payment dynamics have been specified.
Risk transfer changes these relationships. A catastrophe reinsurance layer may leave the frequency of small claims unaffected and sharply reduce the insurer's extreme net loss, while introducing premium costs and counterparty recovery risk. A VaR or ES comparison that ignores premium payments, timing of recoveries, and default terms cannot establish the economic value of the layer. It can, however, identify the distributional difference between gross and net modeled losses under stated assumptions.
Counterexample audit
Every numerical construction above is exactly specified and can be checked using the same three operations: enumerate state probabilities, build cumulative probabilities, and integrate the upper quantile step function. No simulated sample or external dataset is needed. The two-point ranking reversal uses {0,100} with probabilities {0.95,0.05} compared against constant 10. The subadditivity example uses two independent Bernoulli severities of 100 with event probability 0.04, while its dependence comparison uses perfectly matched events.
The independence calculation has three aggregate states, with probabilities 0.96^2=0.9216, 2(0.04)(0.96)=0.0768, and 0.04^2=0.0016. They sum to one. The ES numerator 0.0016(200)+0.0484(100)=5.16 and division by 0.05 yields 103.2. The two stand-alone ES values are each (0.04 times 100)/0.05=80. These identities serve as a reproducible checklist for any software implementation.
A small conceptual refinement is necessary. An ES subadditivity proof assumes that all losses are evaluated under one internally consistent probability space and that the relevant expectations are finite or otherwise handled in an extended-valued framework. Adding stand-alone ES numbers computed from mutually inconsistent marginal forecasts does not automatically generate a valid aggregate bound for an operational portfolio. Consistency of valuation date, measure, model, and units remains indispensable.
Model Validation and Capital Uses
Estimation error and conditional distributions
A risk number is a functional of an estimated probability law. The uncertainty of that estimate can exceed the numerical difference between two portfolios' reported VaR or ES. ES involves an average of potentially extreme losses and can be particularly sensitive to rare observations, tail-shape assumptions, and dependence modeling. VaR can be more stable under changes in far-tail severity, but that same stability can conceal variation in losses beyond its threshold. Each feature is a reason for a targeted diagnostic rather than a reason to discard one statistic in all applications.
Forecast validation should distinguish the uncertainty of a future realized loss from parameter uncertainty, process variation, sampling error, and model-class error. Historical resampling can estimate some sampling effects under assumptions about stationarity and dependence. Parametric simulation may be informative when the model family is credible. Extreme-value methods extrapolate beyond observations and therefore require threshold choice, tail-index diagnostics, and extrapolation sensitivity. These are not interchangeable cures for sparse insurance catastrophe data.
The foundational evidence in McNeil and Frey (2000) demonstrates that conditional volatility and heavy tails matter for risk estimates in their financial time-series study. Their result supports the methodological proposition that ignoring dynamic dependence and tail structure can worsen financial-risk forecasts; it does not by itself establish an insurance catastrophe model's calibration. Transferring a method from traded-market returns to insurance losses requires independent data and hazard-specific validation.
Quantile backtesting and conditional coverage
For a predicted one-period VaR threshold v_t, define exception I_t = 1{L_t > v_t}. Under ideal continuous conditional quantile calibration P(I_t=1 | I_{t-1}, I_{t-2}, ... , information at forecast time) should match beta in the appropriate sense. A basic unconditional test examines the frequency of exceptions; a conditional test also examines clustering and predictability of exceptions. The classical coverage-testing work of Kupiec (1995) is central to the former, while Christoffersen (1998) emphasizes conditional interval forecast evaluation.
When a forecast distribution contains atoms, the exception rate for the strict event L>VaR may be lower than beta even under perfect model calibration. Randomized treatment of ties or appropriately defined interval probabilities may be needed for a statistical equality test. Applying a continuous-model exception test mechanically to a claims distribution with a dominant zero state is therefore potentially misleading.
Testing very high-quantile insurance capital forecasts also faces a small-number problem. At a nominal half-percent annual tail probability, one independent observation per year yields only 0.005 expected exceptions per year. A long historical record still provides limited evidence, and the underlying exposure and climate-economic environment may change over the observation period. This does not eliminate the need for validation; it shifts the design toward complementary tail diagnostics, model review, scenario analysis, and explicit uncertainty communication.
Evaluating expected shortfall forecasts
An ES point forecast is not naturally judged by the same exception-count statistic as a quantile. ES describes the average magnitude of a selected tail, so validation can examine tail-loss realizations and the consistency of accompanying VaR forecasts. The expected shortfall backtesting analysis by Acerbi and Szekely (2014) proposes nonparametric methods for the ES assessment problem. Its institutional research note supports the existence and purpose of those proposed tests, without making them a universal regulator-mandated procedure.
Forecast evaluation has a deeper statistical dimension. Gneiting (2011), in Making and Evaluating Point Forecasts, explains why consistent scoring rules must match the target functional and why ES alone is not generally elicitable as an isolated point functional. The later joint elicitability result of Fissler and Ziegel (2016) establishes that the pair (VaR, ES) is jointly elicitable under appropriate assumptions. That result permits strictly consistent joint loss functions and informed comparative forecast rankings. It is essential to distinguish a mathematical possibility for joint scoring from automatic evidence of satisfactory real-world forecast calibration.
In applied financial econometrics, dynamic semiparametric ES and VaR models developed by Patton, Ziegel, and Chen (2019) illustrate joint modeling and scoring-based estimation. These contributions weaken simplistic assertions that ES cannot be empirically evaluated because it is not separately elicitable. At the same time, no scoring methodology eliminates data scarcity, distribution shift, or the need for model-specific testing.
Insurance and banking regulatory comparison
Different regulatory systems impose different loss objects, confidence levels, and institutional responsibilities. Under Solvency II Article 101, the European Union Solvency Capital Requirement is calibrated to the 99.5 percent one-year VaR of basic own funds. Article 101 identifies the risks to be covered and the broad calibration object, while Article 104's module framework sets out risk modules under the standard formula, including a 99.5 percent one-year VaR calibration for individual modules. These are legal statements about a particular solvency regime, not an empirical claim that 99.5 percent is uniquely optimal in all insurance markets.
The Basel Committee's 2019 market-risk standard sets out an internal models approach for bank trading-book market risk using expected shortfall measures and a treatment of risk-factor liquidity horizons. The specified ES one-tailed confidence level under the internal-model approach is 97.5 percent. This framework is not a rule requiring every insurer to substitute ES for its solvency capital measure, nor does it state that ES is the sole component of bank market-risk capital. The standard distinguishes internal models, standardized approaches, and nonmodellable risk-factor requirements.
The 2019 Basel publication states an international standard timetable and has been incorporated into the consolidated Basel Framework; local legal implementation, transition, and applicable text must be checked by jurisdiction and date. Its stated rationale includes more explicit sensitivity to tail loss and market illiquidity than the earlier VaR-centered treatment. The correct scholarly comparison is therefore between the objects and calibrations actually prescribed by each framework: insurer basic-own-funds solvency over one year in one case, bank trading-book market-risk capital with specified internal-model and liquidity machinery in the other. Substituting their numeric thresholds into a single generic aggregate-claims formula would misrepresent both.
Stress testing, governance, and capital interpretation
Stress testing evaluates severe but defined adverse scenarios that may be poorly represented by historical estimation. A stress result complements VaR and ES by exposing concentrations, nonlinear insurance treaty effects, reserve deterioration, liquidity needs, and recovery timing. Stress testing is not equivalent to computing ES unless the scenarios are assigned a probability law and the required tail integration is performed. Likewise a deterministic maximum contract payout may be relevant without being a quantile or a probability-weighted tail expectation.
Good model governance first fixes the exposure definition, decision purpose, valuation basis, and reporting horizon. It then reports selected loss statistics, sensitivity to alpha, data coverage, tail behavior, dependence structure, contract transformations, estimation uncertainty, and credible scenario extremes. For multi-line insurance portfolios, related questions include whether catastrophe-event catalogues capture joint hazard occurrences, whether reinstatement and reinsurance counterparty effects are modeled, and whether asset-side stresses coincide with liability losses.
A particular capital requirement includes legal recognition of eligible capital, loss-absorbing instruments, valuation rules, and supervisory constraints. One may calculate an economically informative VaR or ES on an illustrative loss distribution without thereby computing a statutory solvency requirement. The legal calibration of a metric is an important empirical fact, while the metric's mathematical properties follow separately from probability theory.
Methods, provenance, and research limitations
The mathematical core uses distribution functions, generalized quantiles, elementary expectation identities, quantile integration, and exact enumeration of Bernoulli claim states. The proofs and examples are authored specifically for this monograph, while the underlying methods and coherence results are attributed to their established literature. References identify the publication version and link to checked publisher records or primary institutional pages. Several journal publishers expose bibliographic descriptions and abstracts rather than unrestricted complete texts; evidence notes distinguish that scope from a claim of full-text assessment.
Regulatory passages are tied to official Basel and EIOPA pages. They establish institutional requirements only for the cited standard or jurisdiction and do not imply implementation uniformity across the world. The study contains no fitted proprietary catastrophe model, no insurer balance sheet, no original cross-country empirical dataset, and no independent external referee report. Its conclusions are therefore analytical and comparative rather than a statement that a particular market's tail has been statistically measured.
The link and evidence protocol treats every externally sourced scientific assertion as a claim requiring a direct, identifying inline hyperlink or a clearly associated full bibliographic entry. Short anchor text connects a specific statement with the particular work most relevant to it. Repeated source use is intentional where the same study supports different arguments; long parenthetical source chains are avoided. Exact references and the source-to-claim ledger remain separate from the reader-facing flow, enabling later audit and correction without inventing a new citation system.
Conclusion
VaR and ES are distinct, mathematically established measures of a common modeled loss distribution. VaR returns a threshold for achieving a chosen probability level. ES averages the upper fraction of losses defined by the same confidence level. Their divergence is substantive in insurance contexts with rare severe claims, zero-loss atoms, contract-induced limits, and tail-distribution uncertainty.
The first exact finding is a ranking reversal: a policy with a five percent chance of losing 100 and otherwise losing zero has VaR_0.95 equal to zero and ES_0.95 equal to 100, whereas a certain loss of 10 yields VaR and ES both equal to 10. The second is a diversification counterexample: two independent four-percent probability claims of 100 each have stand-alone VaR_0.95 values of zero, but their aggregate VaR is 100. Their aggregate ES is 103.2, below the sum of their stand-alone ES values, 160. These calculations expose the information omitted by a quantile and the structural value of integrated tail averaging.
The third finding concerns statistical and institutional interpretation. ES is coherent under the standard integrability conditions, but can be highly sensitive to tail-shape uncertainty and can be infinite when upper-tail means do not exist. VaR can satisfy subadditivity within a sufficiently restricted distributional family, including jointly normal losses at upper confidence levels, while failing it for perfectly ordinary finite discrete risks. Mathematical properties must be assessed under the distributional assumptions actually used.
The fourth finding is that evaluation and regulation require specific context. Quantile exception tests, conditional coverage analysis, ES backtesting, and joint VaR-ES scoring investigate different aspects of forecast performance. European insurer solvency VaR and Basel trading-book stressed ES correspond to separate regulated loss concepts, horizons, and institutional responsibilities. Neither calibration can be generalized into a universal rule for all insurance or banking systems.
A rigorous risk decision accordingly begins with the modeled loss object and the action for which a statistic is requested, retains source-specific definitions for discrete distributions, validates tail assumptions, and reports the appropriate combination of quantile, tail expectation, sensitivity, and stress information. This synthesis advances the explanatory integration of established results and worked examples within insurance risk science without asserting the creation of a new risk-measure theory.
References
Acerbi, C., & Szekely, B. (2014). Backtesting expected shortfall. MSCI Research Insight, December 2. https://www.msci.com/research-and-insights/paper/research-insight-backtesting-expected-shortfall-december-2014
Acerbi, C., & Tasche, D. (2002a). Expected shortfall: A natural coherent alternative to value at risk. Economic Notes, 31(2), 379-388. https://doi.org/10.1111/1468-0300.00091
Acerbi, C., & Tasche, D. (2002b). On the coherence of expected shortfall. Journal of Banking & Finance, 26(7), 1487-1503. https://www.sciencedirect.com/science/article/abs/pii/S0378426602002832
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
Basel Committee on Banking Supervision. (2019). Minimum capital requirements for market risk (corrected February 2019; integrated into the consolidated Basel Framework). Bank for International Settlements. https://www.bis.org/publications/201901-standards-minimum-capital-requirements-market-risk
Christoffersen, P. F. (1998). Evaluating interval forecasts. International Economic Review, 39(4), 841-862. https://ideas.repec.org/a/ier/iecrev/v39y1998i4p841-62.html
European Union. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 101: Calculation of the Solvency Capital Requirement. EIOPA Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2188_en
European Union. (2009, as amended). Directive 2009/138/EC (Solvency II), Article 104: Design of the Basic Solvency Capital Requirement. EIOPA Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2191_en
Fissler, T., & Ziegel, J. F. (2016). Higher order elicitability and Osband's principle. Annals of Statistics, 44(4), 1680-1707. https://doi.org/10.1214/16-AOS1439
Gneiting, T. (2011). Making and evaluating point forecasts. Journal of the American Statistical Association, 106(494), 746-762. https://doi.org/10.1198/jasa.2011.r10138
Kupiec, P. H. (1995). Techniques for verifying the accuracy of risk measurement models. The Journal of Derivatives, 3(2), 73-84. https://doi.org/10.3905/jod.1995.407942
McNeil, A. J., & Frey, R. (2000). Estimation of tail-related risk measures for heteroscedastic financial time series: An extreme value approach. Journal of Empirical Finance, 7(3-4), 271-300. https://www.sciencedirect.com/science/article/pii/S0927539800000128
Patton, A. J., Ziegel, J. F., & Chen, R. (2019). Dynamic semiparametric models for expected shortfall (and Value-at-Risk). Journal of Econometrics, 211(2), 388-413. https://www.sciencedirect.com/science/article/pii/S030440761930048X
Rockafellar, R. T., & Uryasev, S. (2000). Optimization of conditional value-at-risk. The Journal of Risk, 2(3), 21-41. https://doi.org/10.21314/JOR.2000.038
Rockafellar, R. T., & Uryasev, S. (2002). Conditional value-at-risk for general loss distributions. Journal of Banking & Finance, 26(7), 1443-1471. https://www.sciencedirect.com/science/article/abs/pii/S0378426602002716