We study time variation in the shape of the distribution of stock returns. In a global sample covering 17 countries, returns are more left skewed and fat tailed during good times than during bad times. This pattern creates pro-cyclical variation in conditional tail risk, which is the risk of losing several conditional standard deviations of returns. The variation in higher-order moments is hard to reconcile with the idea that disaster risk is elevated in bad times, which is otherwise a basic premise of leading disaster-based asset pricing models.