Machine Learning Based Portfolio Selection Under Systemic Risk

This paper aims to enhance the classical mean-variance portfolio selection by using machine learning techniques and accounting for systemic risk. The optimal portfolio is solved through a three-step supervised learning model. Firstly, the Smooth Pinball Neural Network is employed to predict return distributions of individual assets and the market. Secondly, we use copula to model dependence between assets and the market, based on which we simulate return scenarios. Lastly, we maximize an ex-ante conditional Sharpe ratio conditioning on systemic events. We run a large-scale comparative study using nearly 600 US individual stocks over 37 years. Our set of predictors includes 94 firm characteristics, 14 macroeconomic variables, and 74 industry dummies. The backtesting results demonstrate the superiority of our proposed approach over popular benchmark strategies including a GARCH-based model. This outperformance is statistically significant and robust to the inclusion of transaction costs