On Random-Forest-Based prediction intervals

Abstract

In the context of predicting continuous variables, many proposals in the literature exist dealing with point predictions. However, these predictions have inherent errors which should be quantified. Prediction intervals (PI) are a great alternative to point predictions, as they permit measuring the uncertainty of the prediction. In this paper, we review Quantile Regression Forests and propose five new alternatives based on them, as well as on classical random forests and linear and quantile regression, for the computation of PIs. Moreover, we perform several numerical experiments to evaluate the performance of the reviewed and proposed methods and extract some guidelines on the method to choose depending on the size of the data set and the shape of the target variable.