Intraday Stock Return Prediction under COVID-19: A Machine Learning Approach with High-Frequency Factor Design

This research examines the constraints of the Fama-French Five-Factor (FF5) model within high-frequency trading contexts and suggests the amalgamation of Machine Learning (ML) methodologies with newly developed intraday features to more effectively capture return dynamics at the index level. We use high-frequency data from all the stocks that make up the S&P 500 to see how well linear and nonlinear predictive frameworks work when the market is very volatile and when structural changes happen because of the COVID-19 pandemic. In predicting next-day open-to-close returns, our results demonstrate that machine learning modelsin particular, XGBoostperform noticeably better than linear benchmarks, especially when enhanced with unique intraday factors like realized volatility and volume-price dynamics. In the long run, the model produces economically significant excess returns and exhibits high predictive accuracy with positive and statistically significant alpha under standard factor models. These findings suggest that traditional asset pricing models fall short in capturing short-term market dynamics, and that integrating high-frequency signals can substantially improve return forecasting in turbulent periods.