A Machine Learning Approach to Post-Earnings Stock Price Prediction

Abstract

Post-Earnings-Announcement Drift (PEAD) is one of the most heavily researched market phenomena, in which cumulative abnormal stock returns continue to drift towards the direction of market surprises in the days and weeks following earnings announcements. We present a machine learning approach for PEAD prediction which incorporates both financial and sentiment feature variables to predict stock returns along several time horizons. The models we implement include OLS, Lasso, Ridge Regression, Logistic Regression, Linear and Non-linear Support Vector Machine, Random Forest, and Gradient Boosting, and we execute a trading strategy based on the predictions made by these models. We find that Linear SVM under a 40 trading-day horizon outperforms the other prediction models, producing a prediction accuracy of 52.75% as well as a Covid-adjusted annualized return and Sharpe ratio of 6.95% and 0.75, respectively. Through further analysis, we show that prediction accuracy can be meaningfully increased by implementing an \(\epsilon\) threshold for regression models, and we also demonstrate that Random Forest considers the CBOE VIX and the EPS-surprise variables as the two most important feature variables for predicting the direction of future stock returns. Given that both Linear SVM and the VIX feature variable have not been significantly studied in the context of predicting PEAD, our results provide two meaningful contributions to existing PEAD prediction literature.