Intraday Stock Market Predictability with Machine Learning
Log in to collectAcademic paper
Intraday Stock Predictability Everywhere
Fred Liu; Lars Stentoft
- CAUniversity of Guelph
- CAWestern University
- ?University of Western Ontario, Department of Economics
- CACenter for Interuniversity Research and Analysis on Organizations
- DKAarhus University
- ?Aarhus University - CREATES
- ?Center for Interuniversity Research and Analysis on Organization (CIRANO)
- ?Department of Economics, University of Western Ontario
Strategy in a nutshell
Universe: S&P 500 stocks and sector portfolios (NYSE/AMEX/NASDAQ) using intraday TAQ data (2004–2016). Predict returns using lagged intraday portfolio and stock returns with linear (AR1, ridge, lasso, elastic net) and nonlinear (random forests, gradient-boosted trees, neural networks) machine learning models. Portfolios are constructed from model forecasts, rebalanced intraday, and evaluated for out-of-sample R² and Sharpe ratios. Linear models show higher statistical predictability; nonlinear models outperform economically after transaction costs, especially for short horizons (1-minute).
Economic rationale
Intraday predictability arises from stock characteristics, liquidity, and slow-moving capital. Machine learning models exploit these patterns for economically significant returns, with nonlinear models capturing complex relationships and enabling profitable intraday strategies despite transaction costs.