Worst-Case CVaR

Ordinary conditional value-at-risk (CVaR) is the average loss in the worst $\alpha\%$ of outcomes, computed under one assumed loss distribution. The problem is that the distribution itself is estimated from limited historical data, and a small error in that estimate can produce a large error in the tail risk it implies, exactly where you can least afford to be wrong. Worst-case CVaR addresses this by not committing to a single distribution at all: instead it defines a whole set of plausible distributions (an "uncertainty set", say, all distributions with a given mean and covariance, or within some statistical distance of the historical one) and reports the CVaR of whichever distribution in that set is worst.

Portfolios optimized against worst-case CVaR are therefore more conservative than ones optimized against a single historical or parametric estimate, they're explicitly hedged against the possibility that the true distribution isn't quite the one you fitted. This comes at a real cost: because the optimization protects against the worst plausible case, the resulting portfolio typically gives up some expected return relative to one optimized against the single best point-estimate.

If a standard CVaR estimate based on the historical sample says the worst-5% average loss is 8%, a worst-case CVaR computed over an uncertainty set of distributions sharing the same mean and covariance might report 11%, the extra 3 points reflecting how much worse the tail could plausibly be if the true distribution differs from the historical sample in ways consistent with the same first two moments.