Predicting Stock Return Movement Directions with Sentiment Analysis of News Headlines: A Machine Learning Approach

This chapter combines the sentiment features of news headlines, stock market data, and macroeconomic factors to predict the direction of stock return movements in one month after the release of the news article. The sentiment features are extracted with Flair, an advanced library for Natural Language Processing (NLP). The stock market data of a company contains a comprehensive collection of 94 variables used by a prior literature (Gu et al., 2020) for the prediction of stock return. To construct the prediction model, this chapter applies seven machine learning algorithms (including Gradient Boosting, XGBoosting, Random Forest, Artificial Neural Networks, Support Vector Machine, Naïve Bayes, and Logistic Regression). The out-of-sample tests show that tree-based ensemble methods (i.e., Random Forest, XGBoosting, and Gradient Boosting) provide the most accurate predictions, with the maximum AUC-ROC of 0.74. Furthermore, this study provides evidence for the effectiveness of the sentiment features of news headline for the prediction of future stock returns as the median of the sentiment score of the news headline is listed as one of the most important predictors in the model.