The author uses extreme value theory to study time-varying idiosyncratic tail risk for a large panel of US stocks. The author demonstrates a significant performance gain by using forward-looking information extracted from implied volatilities and nonlinear models, compared to linear models that use only backward-looking information. Extreme value theory plays a key role in predicting the distribution of return realizations conditional on the occurrence of a tail event. The author finds that, surprisingly, out-the-money calls (respectively, puts) contain important information about lower (respectively, upper) tails. Furthermore, the author finds evidence that the asymmetric nature of the negative tail distribution in comparison to the positive tail is captured by nonlinear models only.