The paper examines the potential of a novel data mining method, the random forest classifier, to support managerial decision making in complex forecasting applications. A modelling paradigm is proposed that embraces a learning curve analysis and grid-search to analyse the model’s sensitivity towards the number of training examples and parameter settings, respectively, and, eventually, produce a final classifier with high predictive accuracy. The effectiveness of the approach is evidenced by experimental evaluation using the data of the 2008 data mining cup competition.