Building portfolios based on machine learning predictions

Abstract

This paper demonstrates that portfolio optimization techniques represented by Markowitz mean-variance and Hierarchical Risk Parity (HRP) optimizers increase the risk-adjusted return of portfolios built with stocks preselected with a machine learning tool. We apply the random forest method to predict the cross-section of expected excess returns and choose n stocks with the highest monthly predictions. We compare three different techniques—mean-variance, HRP, and 1/N— for portfolio weight creation using returns of stocks from the S&P500 and STOXX600 for robustness. The out-of-sample results show that both mean-variance and HRP optimizers outperform the 1/N rule. This conclusion is in the opposition to a common criticism of optimizers’ efficiency and presents a new light on their potential practical usage.