Expected Shortfall (ES) Fundamentals

• Expected Shortfall (ES) is a tail risk measure that averages extreme losses beyond the Value-at-Risk threshold, capturing the severity of rare events.
• It is coherent with properties such as subadditivity, homogeneity, and monotonicity, making it robust for portfolio risk measurement and capital allocation.
• ES is integrated into optimization and regression frameworks via methods like the Rockafellar–Uryasev formula and joint VaR-ES backtesting to address estimation challenges.

Expected Shortfall (ES), also called Conditional Value-at-Risk (CVaR), Tail Value-at-Risk (TVaR), Average Value-at-Risk (AVaR), or superquantile, is a law-invariant tail-risk functional that summarizes not merely the location of a tail threshold, as Value-at-Risk (VaR) does, but the average severity of outcomes beyond that threshold. For a loss random variable $L$ and confidence level $\alpha\in(0,1)$, a standard definition is

$$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$

When $\mathbb P[L=q_\alpha(L)]=0$, this reduces to the tail conditional expectation

$$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$

In return-based conventions, especially in conditional time-series modeling, ES is written for the lower tail,

$$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$

under continuity. Across these conventions, ES is the canonical tail-average risk measure and a regulatory standard precisely because it retains information about tail severity that VaR discards (Emmer et al., 2013, Patton et al., 2017).

1. Definition, sign conventions, and relation to Value-at-Risk

ES is most naturally interpreted as an average of quantiles over the relevant tail. In the loss convention, it averages upper-tail quantiles from level $\alpha$ to $1$; in the return convention, it averages lower-tail quantiles from $0$ to $\tau$. The two forms are mathematically analogous and differ only by the sign convention and the tail under study. This dual usage is standard in the literature summarized here: some papers treat $\alpha\in(0,1)$0 or $\alpha\in(0,1)$1 as losses, where larger values are worse, while others treat $\alpha\in(0,1)$2 as returns and define ES on the lower tail of the conditional return distribution (Emmer et al., 2013, Patton et al., 2017).

The contrast with VaR is foundational. VaR identifies a cutoff exceeded with probability $\alpha\in(0,1)$3 in the loss convention, or $\alpha\in(0,1)$4 in the return convention, but it does not measure how large exceedances are once they occur. ES resolves precisely this limitation. Two portfolios may share the same VaR while having very different tail behavior; ES distinguishes them because it conditions on, or equivalently averages over, the tail beyond the quantile threshold. This is the basic reason ES is repeatedly described as conceptually superior to VaR for tail-risk measurement and regulation (Emmer et al., 2013, Burzoni et al., 2020).

A second representation, central in optimization and robust analysis, is the Rockafellar–Uryasev formula. For an integrable random variable $\alpha\in(0,1)$5 and confidence level $\alpha\in(0,1)$6,

$\alpha\in(0,1)$7

with minimizers equal to the quantile interval $\alpha\in(0,1)$8. This embeds ES in convex optimization through the mean excess loss $\alpha\in(0,1)$9, also called the stop-loss premium, and underlies a substantial part of the computational literature (Guan et al., 2022).

2. Coherence, dependence structure, and axiomatic characterizations

A central reason for the prominence of ES is coherence. In the comparison of standard risk measures, ES satisfies homogeneity, subadditivity, monotonicity, and translation invariance, whereas VaR fails subadditivity in general. Accordingly, ES captures diversification more faithfully and avoids cases in which portfolio aggregation is penalized rather than recognized as risk-reducing. In the same comparison, expectiles are coherent only for $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$0, and for $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$1 they are not comonotonically additive, a property ES retains in the usual law-invariant coherent-risk-measure setting (Emmer et al., 2013).

The dependence sensitivity of ES has also been formalized axiomatically. Under monotonicity, translation invariance, lower semicontinuity, normalization at $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$2, and $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$3-concentration aversion, a risk measure is ES at level $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$4. More generally, concentration aversion implies that the relevant law-invariant functionals are exactly functions of $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$5 and $$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$6, yielding the family of mean-ES criteria. Within that class, coherent risk measures take the explicit form

$$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$7

This gives an economic foundation for mean-ES portfolio selection and isolates ES via lower semicontinuity as the unique monetary member satisfying the full axiom set (Han et al., 2021).

A related actuarial characterization comes from efficient insurance contracts. In a Pareto-optimal insurance framework, if deductible-type ceded loss functions are efficient, then the insured and insurer must evaluate risk through an ES/mean mixture,

$$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$8

With the additional requirement of lower semicontinuity with respect to almost sure convergence, this mixture collapses to ES itself. This gives an insurance-design justification for the special role of ES, not merely a portfolio-theoretic one (Wang et al., 2021).

The geometry of ES also depends on the operation under consideration. ES is convex with respect to linear portfolio aggregation,

$$\mathrm{VaR}_\alpha(L)=q_\alpha(L)=\inf\{l:\mathbb P(L\le l)\ge \alpha\}, \qquad \mathrm{ES}_\alpha(L)=\frac{1}{1-\alpha}\int_\alpha^1 q_u(L)\,du.$$9

but concave with respect to finite mixtures of probability distributions,

$\mathbb P[L=q_\alpha(L)]=0$0

This establishes a sharp distinction between diversifying by holding a convex combination of positions and randomizing among their distributions, and it extends to spectral risk measures (Tselishchev, 2019).

3. Elicitability, identifiability, and backtesting

The main classical caveat attached to ES is that it is not elicitable on its own: there is no strictly consistent scoring function for ES alone. This obstructs direct forecast comparison and simple univariate backtesting in the manner available for VaR. At the same time, ES is conditionally elicitable, and one practical workaround proposed in the comparison of standard measures is to approximate ES by a linear combination of four quantiles,

$\mathbb P[L=q_\alpha(L)]=0$1

so that standard VaR backtesting machinery can be leveraged indirectly. That paper also recommends careful manual inspection of extreme upper-tail observations and notes a related Basel Committee approach based on two quantiles, $\mathbb P[L=q_\alpha(L)]=0$2 and $\mathbb P[L=q_\alpha(L)]=0$3 (Emmer et al., 2013).

A decisive refinement is that the pair $\mathbb P[L=q_\alpha(L)]=0$4 is jointly elicitable. A strictly consistent class of joint scoring functions exists, and this permits comparative backtesting via Diebold–Mariano-type tests, model ranking, and benchmark-relative evaluation even though ES alone is not elicitable. The joint score is intrinsically non-separable into an isolated VaR score plus an isolated ES score, reflecting the fact that ES becomes statistically tractable only together with its associated quantile (Fissler et al., 2015).

This insight underpins regression-based ES backtesting. In the ESR framework, a stand-alone ES regression is infeasible, so estimation is performed through a joint VaR–ES regression, after which Wald-type backtests are constructed. The proposed variants are the Auxiliary ESR backtest, the Strict ESR backtest, and the Intercept ESR backtest. The strict and intercept versions require only ES forecasts as input, and the intercept version yields a one-sided ES-only backtest focused on underestimation of risk. Misspecification-robust covariance estimation is essential in finite samples and materially improves performance (Bayer et al., 2018).

Forecast comparison has also been extended from backtesting to encompassing. ES encompassing tests use a joint loss for $\mathbb P[L=q_\alpha(L)]=0$5 and examine whether one forecast already contains the relevant information in another forecast, or whether forecast combination improves performance. Three variants are proposed: a joint VaR and ES encompassing test, an auxiliary ES encompassing test, and a strict ES encompassing test designed for settings in which only ES forecasts are reported. The asymptotic theory is explicitly misspecification-robust (Dimitriadis et al., 2019).

More recent work has expanded the backtesting toolkit beyond classical scoring-based procedures. E-backtesting constructs model-free, sequential, anytime-valid e-processes for ES and VaR. Because ES alone is not backtestable in that framework, VaR enters as auxiliary information, and the canonical ES backtest e-statistic is

$\mathbb P[L=q_\alpha(L)]=0$6

This yields one-sided, optional-stopping-valid procedures aligned with the regulatory emphasis on underestimation of risk (Wang et al., 2022). A different development decomposes ES validation into inter-violation durations and violation severities and uses shifted Meixner and shifted Legendre polynomials to derive orthogonal moment conditions. The resulting Wald test nests unconditional and conditional coverage backtests for both VaR and ES and is explicitly diagnostic: it separates misspecification in violation frequency, tail severity, and temporal dependence (Hué et al., 2024).

4. Optimization, estimation error, and statistical instability

The optimization literature treats ES as a portfolio-selection criterion analogous to variance in Markowitz optimization, but the tail focus induces extreme data hunger. In the Gaussian i.i.d. large-$\mathbb P[L=q_\alpha(L)]=0$7, large-$\mathbb P[L=q_\alpha(L)]=0$8 regime, the relevant control parameter is the aspect ratio

$\mathbb P[L=q_\alpha(L)]=0$9

and the estimation error $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$0 depends primarily on $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$1 and the confidence level $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$2. The contour map of ES estimation error identifies level sets $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$3 in the $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$4-plane and an uppermost phase boundary beyond which ES optimization has no feasible solution. At $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$5 and target estimation error $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$6, the required ratio is

$$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$7

For $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$8, this implies $$\mathrm{ES}_\alpha(L)=\mathbb E[L\mid L\ge q_\alpha(L)].$$9 daily observations, about $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$0 years of daily returns. The same work emphasizes that parametric estimation does not eliminate the fundamental tail-data problem, since reliable tail estimation remains difficult in realistic fat-tailed markets (Kondor et al., 2015).

The instability is especially severe because ES depends only on the worst tail fraction of the sample. At $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$1, only the worst $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$2 of observations contribute; at $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$3, only the worst $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$4 do. As $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$5, finite-$$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$6 contour lines bend toward smaller admissible $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$7, so the required sample length grows further. This makes unregularized high-confidence ES optimization problematic even before model misspecification is considered (Kondor et al., 2015).

Regularization alters this landscape substantially. Under an asymmetric $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$8 penalty, with no-short-selling as the limiting case $$\mathrm{ES}_\tau(Y_t\mid \mathcal F_{t-1}) =\frac{1}{\tau}\int_0^\tau F_t^{-1}(s)\,ds =\mathbb E_t\!\left[Y_t\mid Y_t\le Q_\tau(Y_t\mid \mathcal F_{t-1})\right]$$9, $\alpha$0, replica-based analysis yields explicit formulas for in-sample ES, out-of-sample ES, estimation error, and the optimal weight distribution. The no-short constraint acts as a high-volatility cutoff: high-volatility assets are set to zero with higher probability than low-volatility assets. This induces sparsity and renormalizes the effective aspect ratio to

$\alpha$1

where $\alpha$2 is the density of eliminated assets. In that sense, regularization extends feasibility by shrinking the effective dimension of the optimization problem, although beyond the physically meaningful region the solution may become dominated by the constraint rather than by information in the data (Papp et al., 2021).

The computational theory of ES also has a dual side. The reverse ES optimization formula shows that for a fixed threshold $\alpha$3,

$\alpha$4

with maximizers $\alpha$5. Together with the Rockafellar–Uryasev minimization formula, this reveals a symmetry between the ES curve $\alpha$6 and the mean excess curve $\alpha$7, and links both to Fenchel–Legendre transforms and optimized certainty equivalents (Guan et al., 2022).

5. Conditional, dynamic, and high-dimensional ES regression

Modern ES regression proceeds by exploiting either joint elicitability with VaR or a two-step construction in which the quantile is treated as a nuisance parameter. In dynamic semiparametric modeling, the key step was to specify $\alpha$8 directly and estimate parameters by minimizing the FZ0 loss,

$\alpha$9

rather than specifying a full conditional density. This led to one-factor and two-factor GAS-type ES-VaR models, as well as a hybrid GAS/GARCH model, and empirical evidence on major equity indices showed that the one-factor GAS and hybrid specifications outperform rolling-window and many GARCH-based benchmarks for moderate tail levels (Patton et al., 2017).

A closely related but fully autoregressive proposal is Conditional Autoregressive Expected Shortfall (CAESar). Its three-step design first estimates VaR via CAViaR regression, then formulates ES in an autoregressive manner, and finally jointly estimates VaR and ES under a monotonicity constraint ensuring $1$0. CAESar accommodates heteroskedasticity and dynamic patterns without a distributional assumption on price returns, and reported simulations and empirical exercises indicate strong performance relative to regression-based competitors, particularly in deep tails (Gatta et al., 2024).

High-dimensional linear ES regression has developed along several lines. One approach uses a robust two-step procedure with a first-stage quantile regression and a second-stage adaptive Huber regression based on a Neyman-orthogonal score. The orthogonality condition

$1$1

makes ES inference first-order insensitive to first-stage quantile error, while Huberization addresses the highly skewed and heavy-tailed generated responses that arise in ES regression. The resulting theory provides explicit non-asymptotic bounds and Gaussian approximation error control when the number of regressors grows with sample size (He et al., 2022).

A sparse version is ES LASSO. There the conditional ES is written as a linear function of high-dimensional regressors after defining an auxiliary dependent variable from the response and a pre-estimated quantile,

$1$2

and estimating

$1$3

The nonasymptotic analysis explicitly tracks first-stage quantile error, accommodates $1$4-mixing heavy-tailed time series, and supports applications such as CoES with nonlinear regressor dictionaries (Barendse, 2023).

Panel and nonparametric ES regression extend the same template. Expected Shortfall Panel Regression introduces a latent factor structure,

$1$5

thereby modeling tail risk jointly across units with common latent downside factors, and establishes consistency, asymptotic normality, and non-asymptotic normal approximation for the two-stage estimator (Hou et al., 14 Apr 2026). In a nonparametric direction, deep neural ES regression uses a first-stage deep quantile regression estimator $1$6, a second-stage ES estimator based on the surrogate response

$1$7

and a Huberized second-stage loss

$1$8

The resulting deep robust ES estimator is first-order insensitive to quantile error and achieves improved non-asymptotic behavior under heavy tails (Yu et al., 11 Nov 2025).

Taken together, these developments suggest a common architecture: conditional quantile recovery in the first stage, ES-specific tail-mean estimation in the second, and explicit robustness mechanisms—joint scoring, orthogonality, Huberization, sparsity control, or latent factors—tailored to the fact that ES is both more informative and more statistically fragile than VaR.

6. Capital allocation, ES contributions, and generalizations of the ES paradigm

ES is not only a scalar portfolio risk measure; it is also a basis for capital allocation. Under the Euler principle, the contribution of component $1$9 to portfolio ES is

$0$0

and these ES contributions satisfy the full allocation property

$0$1

In earlier capital-allocation terminology, for a homogeneous risk measure $0$2, risk contributions are defined by

$0$3

and for ES the contribution is the expected loss of the component in the portfolio tail scenario. This makes ES especially natural for Euler allocation and for analyzing diversification indices and marginal diversification effects (Emmer et al., 2013, Koike et al., 2024).

The statistical treatment of ES contributions inherits the same joint-evaluation logic as ES itself. ESC forecasts are jointly identifiable with total VaR, and they are multi-objective elicitable under a lexicographic order in which VaR accuracy is primary and ESC accuracy is secondary. This has produced traditional and comparative backtests for tuples of ES contributions, Murphy diagrams robust to scoring-function choice, and semiparametric compositional regression models in which allocation weights live on the simplex and automatically satisfy the full allocation property (Koike et al., 2024).

A different generalization is Adjusted Expected Shortfall. Instead of controlling ES at a single level, adjusted ES controls the entire ES profile $0$4 against an increasing benchmark curve $0$5,

$0$6

Ordinary ES is a special case for a particular piecewise benchmark $0$7. The resulting family is monetary, monotone, cash additive, convex, and law invariant, but generally not coherent; positive homogeneity holds only when the measure collapses to standard ES. When $0$8 for a benchmark loss $0$9, adjusted ES becomes the minimal capital shift needed to ensure second-order stochastic dominance over $\tau$0, thereby connecting ES-based capital regulation to SSD in an exact way (Burzoni et al., 2020).

These extensions do not displace standard ES so much as refine its domain of use. ES remains the canonical tail-average benchmark; ES contributions localize that benchmark across components; adjusted ES replaces single-level tail control by profile-level tail control. A plausible implication is that the contemporary literature increasingly treats ES not as a single isolated functional, but as the core object in a broader family of tail-profile, allocation, and dominance-based constructions.