Forecasting multinomial stock returns using machine learning methods

Abstract

In this thesis, the daily returns of the S&P 500 stock market index are predicted using a variety of different machine learning methods. We propose a new multinomial classification approach to forecasting stock returns. The multinomial approach can isolate the noisy fluctuation around zero and allows us to focus on predicting the more informative large absolute returns. Our in-sample and out-of-sample forecasting results indicate significant return predictability from a statistical point of view. Moreover, all the machine learning methods considered outperform the benchmark buy-and-hold strategy in a real-life trading simulation. The gradient boosting machine is the top-performer in terms of both the statistical and economic evaluation criteria.