We investigate the potential for skillfully predicting the number of daily temperature extremes over 3 month (seasonal) periods. We use retrospective forecasts from the Met Office seasonal forecasting system, GloSea4, nominally initialized 1 month ahead of the target season. Initially, we define daily extremes to be events outside either the upper or lower deciles of the daily temperature distribution from the relevant season. This definition provides a threshold that is sufficiently “extreme” to be of interest to many users but moderate enough to allow a sufficient sample for verification and to be of regular use to users. We show that skill reduces slightly at more extreme thresholds. Correlations of predicted and observed numbers of upper or lower decile extreme days over a season are significantly greater than zero over much of the globe and, in general, are better than a persistence forecast. Forecast skill for seasonal mean temperature is similar to, but generally greater than, the skill of predictions of the number of extreme days. Observations have a strong relationship between the seasonal mean and the number of extreme days. We show that the skill in predicting the number of extreme days is largely a consequence of this relationship and occurs primarily through a shift in the distribution of the daily data rather than a change of its shape. The ability to predict the El Niño–Southern Oscillation and climate change are both significant contributors to the skill in predicting temperature extremes. In summer, significant skill also comes from initializing soil moisture.