A statistical learning method called random forests is applied to the prediction of transitions between weather regimes of wintertime Northern Hemisphere (NH) atmospheric low frequency variability. A dataset composed of 55 winters of NH 700-mb geopotential height anomalies is used in the present study. A mixture model finds that the three Gaussian components that were statistically significant in earlier work are robust; they are the Pacific North America (P N A) regime, its approximate reverse (the reverse P N A, or RN A), and the blocked phase of the North Atlantic Oscillation (BN AO). The most significant and robust transitions in the Markov chain generated by these regimes are P N A -> BN AO, P N A -> RN A and BN AO -> P N A. The break of a regime and subsequent onset of another one is forecast for these three transitions. Taking the relative costs of false positives and false negatives into account, the random-forests method shows useful forecasting skill. The calculations are carried out in the phase space spanned by a few leading empirical orthogonal functions of dataset variability. Plots of estimated response functions to a given predictor confirm the crucial influence of the exit angle on a preferred transition path. This result points to the dynamic origin of the transitions.