Using the #CARET package by Max Kuhn, we find out that #randomForest is better than the mean. During the annual meeting of the DVPW #methods section, Andreas Graefe argued, the mean would beat more complex ensemble methods in forecasting. At least, when it comes to a classical prediction scenario, this is not true. I created a simulated data set, defined different models (linear and tree-based) relying on different sets of information. Then I averaged the results. Against this “mean-ensemble”, I send a plain and simple #randomForest in the beauty contest. In addition, I build another model that takes the mean of all models including the random forest. Finally, I build a model that takes all the results of all other models as predictors but combines them again with a random forest. The whole procedure was repeated 100 times. As we can see in the graphic, the simple random forest beats the mean in every error type. For the mean absolute percentage error (MAPE), the lin