Conditional VaR and Expected Shortfall

Value at Risk answers "how bad is a typical bad day?" but is silent on "how bad are the days that are worse than that?" Conditional VaR, also called Expected Shortfall (ES) or CVaR, answers the second question by averaging losses in the tail beyond the VaR threshold. It is the risk measure serious risk managers use when VaR is not enough.

The definition

For a confidence level $\alpha$ (say 95%), VaR is the $\alpha$-quantile of the loss distribution. Expected Shortfall is the conditional expectation of loss given that VaR is breached:

$$ES_{\alpha} = E[L \mid L > \text{VaR}_{\alpha}]$$

For continuous loss distributions this simplifies to:

$$ES_{\alpha} = \frac{1}{1-\alpha}\int_{\alpha}^{1} \text{VaR}_u , du$$

Expected Shortfall averages the loss across all quantiles from $\alpha$ to 1, capturing the entire severity of the tail.

Why ES fixes VaR's failures

Tail sensitivity: Two portfolios with identical 95% VaR can have very different 99% losses. VaR cannot distinguish them; ES does, because the average beyond the threshold reflects the actual depth of the tail.

Coherence: ES satisfies the four axioms of a coherent risk measure — monotonicity, sub-additivity, positive homogeneity, and translation invariance. VaR violates sub-additivity, meaning VaR can suggest a merged portfolio is riskier than the sum of its parts, contradicting diversification. ES never does.

Convexity: ES is a convex function of portfolio weights, so it can be minimized with standard optimization tools, making it usable in portfolio construction, not just risk reporting.

Calculating ES

Historical method: Sort losses, take the worst $(1-\alpha)$ fraction, and average them. Simple and assumption-free, but depends on having enough tail observations — which is exactly where data is scarcest.

Parametric method: Under a normal distribution:

$$ES_{\alpha} = \mu + \sigma \cdot \frac{\phi(z_{\alpha})}{1-\alpha}$$

where $\phi$ is the standard normal density. The normal assumption understates ES for fat-tailed distributions, so use a Student-t or extreme value distribution when tails matter.

Monte Carlo method: Simulate the portfolio under an assumed return model, sort simulated losses, and average the tail. Flexible across non-normal distributions and option positions.

Practical use for traders

• Tail-aware position limits: Set ES limits in addition to VaR. A position that fits within VaR but breaches ES carries hidden tail risk.
• Strategy comparison: Two strategies with equal mean return and equal VaR can be ranked by ES — the lower ES strategy is safer in the tail.
• Capital allocation: Allocate risk budget using ES rather than VaR when tails are fat, which is most of the time in real markets.

ES depends on the same tail data that is hardest to estimate. Bootstrap the sample and report the ES range, not a point estimate. Never confuse an ES estimate with the actual worst case.