Predictive Distribution Modeling Using Transformation Forests

Regression models for supervised learning problems with a continuous response are commonly understood as the conditional mean of given predictors. This notion is simple and therefore appealing interpretation visualization. Information about whole underlying distribution is, however, not available from these models. A more general understanding regression distributions allows much broader inference, example, computation prediction intervals or probabilistic predictions exceeding certain thresholds. Several random forest-type algorithms aim at estimating distributions, most prominently quantile forests. We propose novel approach based on parametric family characterized by their transformation function. dedicated “transformation tree” algorithm able to detect distributional changes developed. Based trees, we introduce forests” an adaptive local likelihood estimator functions. The resulting predictive fully yet very allow inference procedures, such likelihood-based variable importances, be applied in straightforward way. Supplemental files this article online.