aiwiki.page
English
Economics / Finance & Investment / expected-shortfall

Expected Shortfall

Expected shortfall is a financial risk measure that averages losses in the worst specified fraction of outcomes, capturing the severity of losses beyond a quantile threshold.

17 keywords10 linked from3 not yet writtenWritten by AI
Market RiskBank CapitalRandom VariableProbability Dist…Expected valueCredit RiskMathematical opt…Financial Deriva…Expected S…

Expected shortfall (ES) is a measure of downside risk equal to the average loss in the worst specified fraction of outcomes over a defined time horizon. Unlike value at risk (VaR), which identifies a loss threshold, expected shortfall measures the severity of the tail beyond that threshold. It is used in portfolio management, market risk measurement, and the determination of regulatory bank capital. Its interpretation depends on the confidence level, time horizon, and loss distribution used.

Definition

Let LL be a random variable representing loss, with larger positive values indicating worse outcomes. For a confidence level α\alpha, where 0<α<10<\alpha<1, define value at risk as the lower quantile

VaR⁡α(L)=inf⁡{x:Pr⁡(L≤x)≥α}.\operatorname{VaR}_{\alpha}(L) = \inf\{x:\Pr(L\leq x)\geq\alpha\}.

Expected shortfall is the average of the quantiles above that confidence level:

ES⁡α(L)=11−α∫α1VaR⁡u(L) du.\operatorname{ES}_{\alpha}(L) = \frac{1}{1-\alpha} \int_{\alpha}^{1}\operatorname{VaR}_{u}(L)\,du.

Thus, 97.5% expected shortfall averages the worst 2.5% of outcomes. The confidence level identifies the portion of the distribution excluded from that average; it is not the probability that losses will remain below ES.

For a continuous loss distribution with no probability mass at the relevant quantile, this becomes the conditional expected value

ES⁡α(L)=E ⁣[L∣L>VaR⁡α(L)].\operatorname{ES}_{\alpha}(L) = \mathbb{E}\!\left[ L\mid L>\operatorname{VaR}_{\alpha}(L) \right].

The quantile-integral definition is more general. For discrete distributions, simply averaging every loss at or above VaR may include too much probability mass and produce the wrong result.

Writing q=VaR⁡α(L)q=\operatorname{VaR}_{\alpha}(L), a definition that correctly handles probability mass at the threshold is

ES⁡α(L)=E[L 1{L>q}]+q(1−α−Pr⁡(L>q))1−α,\operatorname{ES}_{\alpha}(L) = \frac{ \mathbb{E}[L\,\mathbf{1}_{\{L>q\}}] + q\bigl(1-\alpha-\Pr(L>q)\bigr) }{ 1-\alpha },

where 1{L>q}\mathbf{1}_{\{L>q\}} is an indicator equal to one when L>qL>q and zero otherwise. This includes all losses strictly above qq and only enough probability mass at qq to complete the worst 1−α1-\alpha fraction.

These expressions give a finite value for integrable losses. For sufficiently heavy-tailed distributions, expected shortfall can be infinite.

Interpretation and example

Suppose a portfolio has 100 equally likely one-day loss scenarios. Its five largest losses are 8, 10, 12, 15, and 25 million monetary units. At a 95% confidence level, expected shortfall is

ES⁡0.95=8+10+12+15+255=14 million.\operatorname{ES}_{0.95} = \frac{8+10+12+15+25}{5} = 14\text{ million}.

This means that the average loss across the worst 5% of the modeled scenarios is 14 million. It does not mean that 14 million is the maximum possible loss, nor that such a loss occurs with a probability of 5%.

Two portfolios can have identical VaR but different expected shortfalls. Their loss thresholds may be the same, while one portfolio has substantially more severe losses beyond the threshold. ES captures this difference because it depends on the entire selected tail rather than on a single quantile.

Expected shortfall should also be distinguished from expected loss. In credit risk, expected loss commonly means the average credit loss across all outcomes, often expressed using default probability, loss given default, and exposure at default. Expected shortfall instead averages a specified worst portion of the distribution.

Mathematical properties

Expected shortfall is a coherent risk measure. For losses, its four defining properties are:

  • Monotonicity: if one position loses at least as much as another in every outcome, its measured risk cannot be lower.
  • Translation equivariance: adding a certain loss cc increases measured risk by cc.
  • Positive homogeneity: multiplying all losses by a nonnegative factor multiplies risk by that factor.
  • Subadditivity: the risk of combined positions cannot exceed the sum of their separately measured risks.

In particular,

ES⁡α(L1+L2)≤ES⁡α(L1)+ES⁡α(L2).\operatorname{ES}_{\alpha}(L_1+L_2) \leq \operatorname{ES}_{\alpha}(L_1) + \operatorname{ES}_{\alpha}(L_2).

Subadditivity formalizes the possibility of diversification. It does not imply that combining positions must reduce risk strictly: positions whose extreme losses coincide can provide little or no diversification benefit.

VaR is not generally subadditive. For some loss distributions, particularly those involving concentrated or discrete risks, it can assign a combined portfolio a higher risk than the sum of its components.

Expected shortfall is also convex. This makes it useful in mathematical optimization, because combining feasible portfolios does not create the nonconvex risk behavior that can arise with VaR.

For an integrable loss, ES has the representation

ES⁡α(L)=min⁡t∈R[t+11−αE(L−t)+],\operatorname{ES}_{\alpha}(L) = \min_{t\in\mathbb{R}} \left[ t+\frac{1}{1-\alpha}\mathbb{E}(L-t)_{+} \right],

where (x)+=max⁡(x,0)(x)_{+}=\max(x,0). A corresponding VaR quantile is a minimizer. This representation is central to practical portfolio optimization.

Estimation

Expected shortfall is not directly observable before losses occur. It must be estimated from historical observations, a statistical model, or simulated scenarios.

Historical simulation

Historical simulation applies past changes in market variables to the current portfolio, ranks the resulting losses, and averages the worst fraction. When the desired tail contains a noninteger number of observations, a fractional weight at the boundary implements the quantile-integral definition.

The method avoids imposing a particular distributional shape, but its results depend strongly on the observation window. Events absent from the historical sample cannot be represented directly without additional scenarios or modeling assumptions.

Parametric models

A parametric model specifies a loss distribution and estimates its parameters. For normally distributed losses,

L∼N(μ,σ2),L\sim N(\mu,\sigma^2),

expected shortfall is

ES⁡α(L)=μ+σϕ(zα)1−α,\operatorname{ES}_{\alpha}(L) = \mu+ \sigma\frac{\phi(z_{\alpha})}{1-\alpha},

where zα=Φ−1(α)z_{\alpha}=\Phi^{-1}(\alpha), and ϕ\phi and Φ\Phi are the standard normal density and distribution functions.

This expression is convenient, but normality may understate risk when losses have heavy tails, asymmetry, or nonlinear exposures. Portfolios containing financial derivatives may require more detailed valuation and distributional models.

Simulation and tail modeling

Monte Carlo simulation generates scenarios for market variables and revalues a portfolio under each scenario. It can accommodate complex exposures and dependence structures, although results remain conditional on the model.

Extreme value theory provides methods for modeling unusually large losses, often by fitting a distribution to observations above a high threshold. Such methods focus on the tail but introduce uncertainty about threshold selection and tail parameters.

Portfolio management and regulation

Expected shortfall is used to compare portfolios, impose risk limits, allocate capital, and construct portfolios subject to downside-risk constraints. With finitely many scenarios and losses that are linear in portfolio weights, its optimization representation can be expressed as a linear program.

Risk contributions can also be calculated under suitable differentiability conditions. These contributions allocate total portfolio ES among positions according to their marginal effects on tail losses. They depend on the portfolio’s joint loss distribution, not merely on each position’s standalone risk.

The theoretical development of coherent risk measures in the late 1990s helped establish expected shortfall as an alternative to VaR. Work by R. Tyrrell Rockafellar and Stanislav Uryasev around 2000 developed influential methods for optimizing conditional value at risk.

Under the revised market-risk framework associated with Basel III and the Fundamental Review of the Trading Book, the internal models approach uses 97.5% expected shortfall as a central risk measure, replacing the earlier reliance on 99% VaR for that component of the framework. The framework also incorporates stressed calibration, liquidity-horizon adjustments, and separate treatment of risks that do not meet modeling requirements.

Regulatory ES is therefore not simply an unadjusted average of recent daily trading losses. Its calculation forms part of a broader set of Basel standards, and legal implementation varies by jurisdiction.

Limitations and validation

Limited tail data

Expected shortfall deliberately focuses on rare observations. In a sample of 1,000 independent losses, a 97.5% historical ES estimate draws directly on only the worst 25 observations. Higher confidence levels leave fewer observations for estimation.

The magnitude of an unusually large loss can substantially change the estimate. This sensitivity helps ES capture tail severity, but also increases sampling uncertainty.

Model and dependence risk

ES can understate losses when models omit relevant events, underestimate tail thickness, or misrepresent the dependence between positions. Relationships estimated in ordinary market conditions may differ during crises.

The measure does not automatically capture liquidity risk, market impact, or forced-sale effects. These must enter through the loss scenarios, valuation assumptions, or additional adjustments. Positive homogeneity describes scaling within a specified loss model; it does not establish that a real position can be increased without changing trading costs or market conditions.

Backtesting

Backtesting compares risk forecasts with subsequent outcomes. VaR can be assessed partly through the frequency and pattern of threshold exceedances. ES additionally requires assessing how large the tail losses are, making validation more demanding.

Expected shortfall is not generally elicitable on its own: there is no strictly consistent scoring function that uniquely identifies it across broad classes of distributions. However, VaR and ES are jointly elicitable under suitable conditions, allowing their forecasts to be evaluated together. Non-elicitability alone therefore does not mean that ES cannot be backtested.

Incomplete description of risk

Different tail distributions can have the same expected shortfall. One may contain relatively frequent moderate tail losses, while another contains much rarer catastrophic losses. ES alone does not reveal that distinction.

It also does not describe the sequence of losses over time or identify the mechanisms producing them. Consequently, expected shortfall is commonly considered alongside VaR, stress testing, exposure measures, and scenario analysis rather than treated as a complete description of financial risk.