CVaRs

CVaR, short for Conditional Value at Risk, is a risk measure used in finance to assess tail risk. It is also known as Expected Shortfall (ES). CVaR at a given confidence level α focuses on the losses that occur in the tail beyond the VaR threshold and represents the expected loss given that a loss has exceeded that threshold. It is widely used in risk management, capital allocation, and performance evaluation because it provides information about the severity of extreme losses.

Formally, for a loss random variable L and a confidence level α in (0,1), VaRα is the α-quantile of L, the value such that losses exceed it with probability (1−α). CVaRα is the average of losses in the tail beyond VaRα, i.e., the expected loss conditional on L being at least VaRα. In continuous distributions, CVaRα can be written as E[L | L ≥ VaRα]. This makes CVaR a more informative measure of tail risk than VaR alone, since it accounts for how bad losses can be beyond the threshold.

CVaR has several desirable theoretical properties. It is a coherent risk measure for α in (0,1), satisfying sub-additivity, positive homogeneity, translation invariance, and monotonicity. In contrast, VaR can fail to be sub-additive, which means it may not always reflect diversification effects. Because CVaR concentrates on the tail, it provides a consistent picture of extreme losses and is sensitive to the shape of the tail distribution.

In practice, CVaR is estimated from data via historical simulation, Monte Carlo simulation, or parametric models. A widely used computational form, introduced by Rockafellar and Uryasev, expresses CVaRα as the minimum over z of z plus a tail-average term: CVaRα = inf_z { z + (1/(1−α)) E[(L − z)⁺] }. In empirical work this expectation is approximated by a sample average, often within convex optimization frameworks that enable direct portfolio optimization to minimize CVaR.

Applications of CVaR include portfolio optimization, risk budgeting, and regulatory capital requirements. It is favored for its tail-sensitivity and coherence, which align with concerns about severe losses. However, CVaR also depends on accurate tail estimation and data availability in the far tail; model risk and non-stationarity can limit its reliability, and practitioners must be cautious about assumptions underlying tail behavior and the chosen confidence level.