Expected Shortfall
Expected Shortfall is the average of all portfolio losses that exceed the Value at Risk threshold at a specified confidence level, providing a more complete measure of tail risk than VaR by quantifying the expected severity of extreme losses, not just their likelihood.
Expected Shortfall (ES) and Conditional Value at Risk (CVaR) are the same measure referred to by different names — ES is the term more commonly used in academic literature and regulatory documents, while CVaR is more common in practitioner contexts. Both describe the mean of the loss distribution beyond the VaR cutoff.
The practical motivation for Expected Shortfall arose from the observed failures of VaR during financial crises. During the 2008 financial crisis, many banks reported that their portfolios were suffering losses at the one-in-10,000-day level or worse — statistically impossible if returns were normally distributed and VaR models were accurate. The problem was not just model error in estimating VaR; it was that VaR said nothing about what happened when the threshold was crossed. ES fills that gap.
From a regulatory perspective, the Basel Committee on Banking Supervision formally replaced VaR with Expected Shortfall as the core market risk metric under the FRTB rules adopted as part of the Basel III framework. US banking regulators, through the Federal Reserve and OCC, incorporate these standards into the capital rules for systemically important financial institutions. Using a 97.5% ES instead of 99% VaR, regulators argued, produces a more conservative and crisis-aware capital requirement.
For fund managers, Expected Shortfall can be incorporated into risk limits, mandate constraints, and reporting. A manager might set a limit that the 95% ES of the portfolio cannot exceed 8% of NAV over any rolling 20-day period, providing more meaningful tail risk governance than a pure VaR limit. Risk systems from vendors like BlackRock Aladdin, FactSet, and Bloomberg PORT calculate ES alongside VaR and drawdown statistics.
ES is not without criticism. It is more data-hungry than VaR — accurately estimating the mean of the tail requires more observations in the tail — and it can be sensitive to the assumed return distribution. Simulation-based approaches (historical simulation or Monte Carlo) tend to be more reliable for ES estimation than analytical approaches when the true distribution has fat tails.