A tail-risk statistic equal to the expected loss conditional on losses exceeding the Value at Risk (VaR) threshold at a specified confidence level for a given loss distribution and loss definition; it quantifies the average severity of losses in the tail beyond the VaR cutoff.

Principle

By conditioning on exceedance of the VaR threshold, CVaR captures average tail severity and thus provides a coherent, convex measure of downside risk that is sensitive to the shape of the loss tail rather than only to a quantile.

Demonstration

Illustrative computation: for a portfolio loss distribution and confidence level 97.5%, identify the VaR such that 2.5% of losses exceed it, then compute the mean of losses conditional on being greater than that VaR; this conditional mean is the CVaR (expected shortfall) at 97.5%.

Misapplication

Confusing VaR with CVaR or treating CVaR as a complete description of tail risk without specifying the distributional model, aggregation method, or whether losses are liquidatable; the semantic error is assuming CVaR alone characterizes all relevant tail dependence or capital needs.

Consequence

CVaR is used for regulatory and internal capital allocation, optimization and stress assessments because it penalizes extreme losses more consistently than VaR; however, its numeric value depends on modeling choices, loss aggregation and the confidence level, affecting capital and portfolio decisions.

Reversal

For discrete, bounded or poorly estimated tails, or when loss distributions are subject to model uncertainty, the conditional expectation beyond a quantile may be unstable or undefined under some definitions; alternative tail metrics or robustification may be required.

Boundary

Applies to loss distributions and requires specification of confidence level, loss definition (e.g., mark-to-market, liquidation), and aggregation method; it does not by itself prescribe allocation rules, regulatory treatment or recovery assumptions.

Semantic Tension

Conservatism versus model sensitivity: CVaR is more conservative and informative about tail severity than VaR but is more sensitive to tail-model assumptions and estimation error, creating a trade-off between prudence and numeric stability.

Synthesis

CVaR provides a principled, expectation-based summary of tail losses beyond a chosen quantile, making it preferable to VaR for optimization and coherent-risk frameworks, but its practical use requires careful specification of model, aggregation and confidence choices and attention to estimation risk.