Dynamical models of various centres have shownin recent years seasonal prediction skill of the North AtlanticOscillation (NAO). By filtering the ensemble members onthe basis of statistical predictors, known as subsampling, it ispossible to achieve even higher prediction skill. In this studythe aim is to design a generalisation of the subsampling ap-proach and establish it as a post-processing procedure.Instead of selecting discrete ensemble members for eachyear, as the subsampling approach does, the distributions ofensembles and statistical predictors are combined to createa probabilistic prediction of the winter NAO. By compar-ing the combined statistical–dynamical prediction with thepredictions of its single components, it can be shown that itachieves similar results to the statistical prediction. At thesame time it can be shown that, unlike the statistical predic-tion, the combined prediction has fewer years where it per-forms worse than the dynamical prediction.By applying the gained distributions to other meteorolog-ical variables, like geopotential height, precipitation and sur-face temperature, it can be shown that evaluating predic-tion skill depends highly on the chosen metric. Besides thecommon anomaly correlation (ACC) this study also presentsscores based on the Earth mover’s distance (EMD) and theintegrated quadratic distance (IQD), which are designed toevaluate skills of probabilistic predictions. It shows that byevaluating the predictions for each year separately comparedto applying a metric to all years at the same time, likecorrelation-based metrics, leads to different interpretations ofthe analysis.