Equity Return Prediction with Augmented Regression and Factor Models

Abstract

In this study, we seek to empirically explore the effectiveness of various regression and factor models in predicting the returns of stocks from different sectors. There has been plenty of research performed in equity return prediction, and the work in this study builds on a lot of the classical framework of financial econometrics, multiple factor models, and regression theory. We present two initial frameworks for constructing the stock return predictions within a given sector: a "baseline" approach and a "new"approach. We analyze and compare the amount of prediction variability explained by both frameworks by looking at adjusted R-squared as a measure of prediction performance. We then augment the models by introducing simple sentiment variables derived from text data such as financial news, statements, and social media. The goal of the inclusion of sentiment features is to see if such text analysis data improves the overall prediction power of our initial models based on adjusted R-squared.

In terms of the two frameworks, we generally see that the new framework models outperform the baseline framework model on the basis of mean adjusted R-squared, especially with predictions in the financial services and health care industries. We also see that the inclusion of our sentiment features leads to an overall positive impact on mean adjusted R-squared across most sectors and return frequencies, leading us to believe that inclusion of a sentiment feature is worthwhile in trying to predict equity returns.