We examine the return predictability of time-varying extreme-event risk at the different points on the return distribution using quantile regression. We find evidence of strong predictive power at the lower quantiles for forecast horizons of up to one year. At the higher quantiles, however, our results show no association between tail risk and the excess stock market returns. Taken together, the evidence explains the abnormally large equity premium, observed during periods of sharp falls in stock prices when market sentiment is bearish.