Variability of the North Atlantic Oscillation (NAO) drives wintertime temperature anomalies in the Northern Hemisphere. Dynamical seasonal prediction systems can skilfully predict the winter NAO. However, prediction of the NAO-dependent air temperature anomalies remains elusive, partially due to the low variability of predicted NAO. Here, we demonstrate a hidden potential of a multi-model ensemble of operational seasonal prediction systems for predicting wintertime temperature by increasing the variability of predicted NAO. We identify and subsample those ensemble members which are close to NAO index statistically estimated from initial autumn conditions. In our novel multi-model approach, the correlation prediction skill for wintertime Central Europe temperature is improved from 0.25 to 0.66, accompanied by an increased winter NAO prediction skill of 0.9. Thereby, temperature anomalies can be skilfully predicted for the upcoming winter over a large part of the Northern Hemisphere through increased variability and skill of predicted NAO.