Intraday Stock Predictability Everywhere

Abstract

With approximately 900 million observations we conduct, to our knowledge, the largest study ever of intraday stock return predictability using machine learning techniques finding consistent out-of-sample predictability across market, sector, and individual stock returns at various time horizons. While linear models have the strongest statistical predictive power, nonlinear models economically dominate them and machine learning intraday long-short portfolios based on their forecasts attain Sharpe ratios of 4 after transaction costs. Predictability is short-lived, highest in the middle of the day, and more pronounced for less liquid firms, which indicates that slow-moving capital is an economic source of mispricing.

Keywords: Machine Learning, Intraday Return Prediction, High-Frequency, Equity Market, Big Data, Lasso, Elastic Net, Random Forest, Gradient Boosting, Neural Networks, Deep Learning, Fintech

JEL Classification: C45, C52, C55, C58, G0, G1, G17