Suppose that you ask several experts about their evaluation of a new project. Should you be more or less confident in your assessment if the expert’s opinions vary a lot? We show that the answer depends on the distribution of opinions. Predictions based on variable opinions are more accurate for light-tailed distributions but less accurate for heavy-tailed distributions. Using this result, we characterize how the accuracy of a collective prediction - based on the average of several forecasts - varies with the variability of these forecasts. More variable forecasts are less accurate when the distribution of forecasts is strongly peaked around a mode close to the truth. More variable forecasts are more accurate only when the distribution of forecasts is flat-topped or skewed, or when there is heterogeneity in how correlated forecasts are.