Tree-based methods are convenient and powerful machine learning tools that can be seen as alternatives to classical regression and prediction models such as generalized linear models, see for example. The most standard procedures are designed to estimate the expectation of a random variable, that is, when it comes to risk, a central scenario (or a best estimate using the Solvency II terminology). In this work, we propose an extension of these tree methods to the study of extreme events, which are of particular interest when it comes to investigate the tail of the distribution and design reinsurance policies. We propose a detailed description of our adaptation of decision trees and support the methodology with new consistency results on these topics. We illustrate the performance of the procedure on simulated data and on a cyberdata base.