I am developing a marketing (Churn) model that has an event rate of 0.5%. So i thought to perform oversampling. I mean making the number of events equal to number of non-events by reducing non-events (50-50 after sampling). After oversampling, we need to adjust predicted probabilities as it inflates intercept. I always do it in logistic regression. Does the same adjustment require for decision tree, random forest or other ensemble techniques? I am following "Applied Predictive Modeling with R (Caret Package)". The author has used the same technique (down sampling) but he did not adjust predicted probability when he score validation and test data. Any help would be highly appreciated! [[alternative HTML version deleted]]