Expected Shortfall

Expected shortfall estimates average loss in the modeled tail beyond a selected confidence level and complements value at risk.

Expected shortfall (ES) estimates the average loss in the worst portion of a modeled loss distribution beyond a selected confidence level. At 95% confidence, it summarizes the average loss in the worst 5% of modeled outcomes.

Expected shortfall is also commonly called conditional value at risk (CVaR), average value at risk, or conditional tail expectation (CTE). Terminology and calculation conventions can differ, especially for discrete distributions, so a report should define the measure rather than relying on the acronym.

Key Takeaways

  • Value at Risk identifies a cutoff; expected shortfall estimates average loss farther into the tail.
  • ES is not a worst-case loss and does not describe every scenario beyond the cutoff.
  • The estimate depends on portfolio scope, horizon, confidence level, data, distribution, dependence, valuation, and liquidity assumptions.
  • In a finite scenario set, observations at the VaR boundary may need weighting to represent the intended tail probability correctly.
  • ES can reveal severity hidden by VaR, but sparse tail data makes it highly sensitive to model and scenario choices.
  • Stress testing remains necessary because a modeled tail may omit structural breaks, market closures, defaults, or forced liquidation.

Expected Shortfall vs. Value at Risk

QuestionValue at riskExpected shortfall
What does it measure?A loss quantile or cutoffAverage loss in the modeled tail beyond a selected confidence level
Does it show tail severity?NoPartly, through the tail average
Is it the maximum loss?NoNo
Main sensitivityLocation of the selected quantileShape and severity of the entire selected tail
Typical useLimits, reporting, backtesting, risk comparisonTail-risk comparison, limits, capital models, stress-oriented analysis

Two portfolios can have the same VaR and very different expected shortfall. If one portfolio’s losses stop just beyond the cutoff while another has rare but very large losses, VaR may not distinguish them, but ES usually will.

Mathematical Definition

Let \(L\) be a loss variable, with larger positive values representing larger losses. A general quantile-based definition is:

$$ \operatorname{ES}_{\alpha}(L) = \frac{1}{1-\alpha} \int_{\alpha}^{1} \operatorname{VaR}_{u}(L)\,du $$

For a continuous loss distribution, this is commonly expressed as the conditional average of losses in the tail at or beyond the VaR cutoff:

$$ \operatorname{ES}_{\alpha}(L) = \mathbb{E} \left[ L \mid L \geq \operatorname{VaR}_{\alpha}(L) \right] $$

The conditional expression is convenient, but the precise treatment of equality at the cutoff matters when many scenarios have the same loss. The quantile-integral definition avoids giving too much weight to a large group of observations exactly at VaR.

Some reports use returns or profit and loss rather than positive loss values. Signs and tail direction must be stated explicitly.

Worked Example

Assume a one-day portfolio model reports:

  • confidence level: 95%
  • VaR: $2 million
  • expected shortfall: $3.4 million

The model places the 95th-percentile loss at $2 million. Within the modeled worst 5% of outcomes, average loss is $3.4 million.

This does not mean every loss beyond VaR will equal $3.4 million. Some tail losses may be close to $2 million and others may be much larger. ES compresses that range into one average.

If a revised model leaves VaR at $2 million but increases ES to $5 million, the cutoff is unchanged while the modeled tail becomes more severe. That change could affect limits, hedging, liquidity planning, or capital analysis.

Calculating Expected Shortfall

Historical Simulation

Revalue the current portfolio under historical risk-factor changes, sort the losses, and average the selected worst tail. If the number of scenarios does not divide cleanly at the confidence cutoff, use appropriate weighting.

Historical ES is easy to explain but cannot include a shock that is absent from the sample. The lookback window and observation weighting can dominate the result.

Parametric Estimation

Estimate ES from an assumed distribution and its parameters. Closed-form formulas exist for some distributions and simple portfolios.

Parametric ES can be efficient, but a normal or otherwise thin-tailed assumption may understate jump, skew, and fat-tail behavior. Nonlinear instruments may also require approximation or full revaluation.

Monte Carlo Simulation

Generate scenarios from a stochastic model, revalue the portfolio, and average the selected tail. Monte Carlo can represent complex portfolios, changing volatility, and nonlinear payoffs.

Its flexibility does not make it assumption-free. Dependence, distribution, calibration, number of simulations, random error, and pricing models all affect the tail.

Why ES Can Improve Tail-Risk Analysis

VaR looks at one point on the loss distribution. ES uses the range of losses beyond the selected confidence level. This can make it more sensitive to:

  • options and nonlinear payoffs
  • concentrated positions
  • fat-tailed risk factors
  • dependence that strengthens in stress
  • jump-to-default or gap risk
  • strategies with frequent small gains and rare large losses

Under standard mathematical definitions, expected shortfall is subadditive: the ES of a combined portfolio should not exceed the sum of stand-alone ES measures. That property generally supports recognition of diversification without the aggregation problems VaR can exhibit for some distributions.

Subadditivity does not prove that modeled diversification will survive a crisis. If the dependence model is wrong, portfolio ES can still understate loss.

Confidence Level, Horizon, and Liquidity

An ES statement should identify:

  • portfolio and legal-entity scope
  • valuation date and currency
  • horizon
  • confidence level
  • loss or profit-and-loss convention
  • scenario-generation method
  • lookback period and weighting
  • valuation treatment
  • liquidity assumptions
  • included and excluded risks

A higher confidence level examines a narrower and more extreme tail. That can increase the estimate while reducing the number of observations available for estimation.

The horizon should reflect the intended decision. Mechanical square-root-of-time scaling inherits assumptions that may fail during volatility clustering, position changes, nonlinear payoffs, margin calls, and illiquid markets.

Backtesting and Model Validation

Expected shortfall is harder to assess from a simple breach count than VaR because ES concerns the size of tail losses, not only whether a cutoff was crossed. Practical validation can combine:

  • VaR exception and clustering analysis
  • comparison of predicted and realized tail losses
  • benchmark or challenger models
  • sensitivity to confidence level and lookback window
  • stress and reverse-stress scenarios
  • data and valuation reconciliation
  • stability by desk, product, and risk factor
  • review of model changes and overrides

Validation should consider the use of the measure, not only its calculation. A technically sound ES can be misused if it is applied to an unapproved portfolio, interpreted as a worst case, or used without liquidity analysis.

Basel Market-Risk Context

The Basel Framework’s internal models approach uses expected shortfall for specified bank trading-book market-risk capital calculations. The framework sets its own confidence level, liquidity-horizon, stress-calibration, desk-approval, modellability, and aggregation requirements.

Those regulatory specifications should not be generalized to every investment portfolio or internal risk report. A non-bank asset manager, corporate treasury, insurer, or investor may use a different ES methodology for a different purpose.

Risks and Limitations

  • Sparse evidence: at high confidence levels, relatively few historical observations determine the estimate.
  • Model dependence: distribution, correlation, volatility, and simulation choices can dominate tail results.
  • Regime change: historical relationships may fail during crises or structural shifts.
  • Liquidity: modeled marks may not reflect executable exit prices, market impact, or time needed to hedge.
  • Valuation: complex instruments can add pricing-model error inside every scenario.
  • Aggregation: apparently diversified exposures may share hidden factors or become dependent in stress.
  • Averaging: the tail mean can hide the range between moderately severe and catastrophic outcomes.
  • Coverage: market ES may omit default, funding, legal, operational, settlement, and strategic risks.

Common Mistakes

  • Calling ES the maximum or worst possible loss.
  • Calculating a simple average of observations above VaR without handling the boundary correctly.
  • Comparing ES numbers with different confidence levels, horizons, scopes, or methods.
  • Assuming CVaR, CTE, and ES always use identical sign and discrete-sample conventions.
  • Treating a higher ES as proof that one portfolio is universally worse without checking expected return, use, and scope.
  • Ignoring scenario weights and the number of tail observations.
  • Using ES instead of stress testing.
  • Reporting more decimal precision than the data and model justify.

Official Sources

These sources describe specified international bank market-risk standards and terminology. They do not prescribe one universal ES method for all portfolios, firms, or jurisdictions.

  • Value at Risk: The loss quantile that defines where the selected modeled tail begins.
  • Tail Risk: The wider set of extreme-event mechanisms and capacity constraints that a tail average cannot fully describe.
  • Downside Risk: Adverse outcomes measured relative to zero, a target, a minimum acceptable return, or another threshold.
  • Stress Testing: Scenario-based analysis that complements expected shortfall when history and modeled distributions omit relevant shocks.
  • Model Risk: The risk that tail estimates mislead because of weak data, assumptions, implementation, validation, or use.

FAQs

Is expected shortfall the same as CVaR?

The terms are commonly used interchangeably. However, sign conventions and treatment of discrete probability at the VaR cutoff can differ, so the calculation should be defined explicitly.

Is expected shortfall always larger than VaR?

For the same loss distribution, horizon, confidence level, and standard upper-tail definitions, expected shortfall is at least as large as VaR because it averages losses at and beyond the cutoff.

Does expected shortfall show the worst possible loss?

No. It is an average across the modeled tail. Individual losses can be much larger, and events absent from the model are not represented.

Educational Use

This article provides general financial education. It is not personalized investment, trading, banking, actuarial, regulatory, model-validation, capital, liquidity, or risk-management advice.

Browse Risk Management