We decompose conditional variance into the sum of two components: one driven by uncertainty about cash-flow growth, or bad variance, and a second driven by uncertainty about discount rates, or good variance. We develop a simple theoretical model with time-varying second conditional moments and document empirical evidence that suggests bad variance earns a risk premium that is statistically and economically significant, whereas good variance does not. In out-of-sample tests we find that bad variance dominates other predictors of market returns and can help identify the gap between the lower bound of Martin (2016) and the actual market risk premium.