Stock Return Prediction and Anomaly Detection by Regression Trees

Noisy markets need extensive descriptions that are noisy themselves, such as deep regression trees that capture many data-local nonlinear anomalies and that do not require arbitrary weighting schemes like traditional linear multifactor models often do. Simple tools allow extraction of general and stock specific insightful information from such regression trees and facilitate factor selection. Large tree-models show excellent out-of-sample predictive power, that we measure and discuss extensively. The method is described by a seven-factor example model for the worldwide industrial sector that delivers 20% quintile-separation excess return out-of-sample and that reveals for instance how following price falls analysts react belatedly and investors overreact and thus create reversal outperformance opportunities