Presenter: Zihao Chen, PhD candidate in Statistics

Title: Portfolio Construction Using Predicted Extreme Stock Returns in the Cross Section: Classification Task with Machine Learning 

Abstract:In Finance, traditional investment portfolios are constructed based on the predicted means of stock returns. However, the predicted expected returns cannot enable us to gain knowledge about tail behaviors of the stock returns, which potentially bring greater profits. In this project, we propose to form portfolios based on the predicted probabilities of extreme stock returns using a large panel of cross-sectional data. Machine learning (ML) methods allow us to capture the potential nonlinear dependence of such probabilities on a large number of firm characteristics. We also adopt Bayesian optimization with Gaussian process to significantly improve the computational efficiency in tuning hyperparameters of ML methods. Simulation studies confirm the accuracy of our methods, and an empirical analysis using 4000+ firms in the U.S. market and the time series observations of their returns and firm characteristics from 1975 to 2016 shows that our approach can identify stocks with extreme positive or negative returns and achieve superior performance relative to the traditional portfolios in long-short investing.