DescriptionThis dissertation explores applications of machine learning methods in empirical asset pricing. The focus of machine learning methods is predictability, which is an important task in empirical asset pricing; therefore, it is natural to apply machine learning methods to solve predictability problems in empirical asset pricing. In my dissertation, one chapter examines predictability in intraday stock returns and the other chapter in mutual fund performance with the help of machine learning methods.
The first chapter predicts intraday stock returns via machine learning methods. This chapter introduces SHapley Additive exPlanations (SHAP) to overcome the challenge of economic interpretability faced by researchers in applying machine learning methods. In particular, we adopt tree-based machine learning methods to predict intraday stock returns with predictors basing on past trading information and news data and then interpret the model predictions with SHAP. Our forecasts generate an out-of-sample R2 of 3.93% for predicting last half hour stock returns and of 9.88% for predicting first half-hour stock returns, which could be translated into huge economic gains; in addition, our results indicate that intraday returns are dominated by short term reversal effect.
The second chapter examines the predictability of mutual fund performance. I aim to find out the relations between predictors and future long term fund performance. Predictive panel regressions show that the linear relations between predictors and the fund performance are not stable and robust. Tree-based methods could generate relatively accurate out-of-sample forecasts, which shows the relatively strong power of nonlinear methods in modeling the relation between fund performance and the corresponding predictors and could be translated into large economic gains. In addition, I find that the intrinsic value, liquidity and profitability of the stock holdings in the fund portfolio are important predictors of long term fund performance. My findings help complement the applications of machine learning methods to predict fund performance.