Performance attribution of machine learning methods for stock returns prediction

We analyze the performance of investable portfolios built using predicted stock returns from machine learning methods and attribute their performance to linear, marginal non-linear and interaction effects. We use a large set of features including price-based, fundamental-based, and sentiment-based descriptors and use model averaging in the validation procedure to get robust out-of-sample predictions. We find that the superiority of regression trees and neural networks comes from two points: their strong regularization mechanism and their capacity to capture interaction effects. The non-linear component of the marginal predictions on the other hand has no predictive power. Thanks to our methodology, we manage to isolate and study in detail the interaction component. We find that it has significative long term performance independent from the linear modeling and is stable through time.