Robustness and Exploitability of Intraday Stock Return Prediction

Are intraday stock market exchange-traded funds (ETF), stock sector ETFs and individual stock returns exploitably predictable at short horizons? In their June 2023 paper entitled "Intraday Stock Predictability Everywhere", Fred Liu and Lars Stentoft study intraday U.S. equity return predictability using machine learning methods. Specifically, they:

• Consider the market portfolio represented by SPDR S&P 500 ETF (SPY), sector portfolios represented by the nine Select Sector SPDR ETFs and individual S&P 500 constituent stocks. For portfolios, return predictors are lagged returns of the portfolio itself and its constituents. For individual stocks, return predictors are the lagged returns of SPY and its constituents.
• Consider intraday return horizons of 1, 5, 10, 15 and 30 minutes.
• Train 17 machine learning methods based initially on the first ten months of data, validate on the next month and evaluate out-of-sample predictive power on the ensuing month. Each month, they repeat these steps by rolling all data by one month (142 test months).
• Test statistical significance via the power of predictions to explain actual future stock returns (R-squared).
• Test gross economic value of predictions via portfolios that buy and sell assets according to predicted returns.
• Test net economic value of predictions by trading only when predicted long or short returns exceed trading frictions (estimated as the bid-ask spread) and debiting frictions from actual returns.

Using intraday transaction data for the specified ETFs and S&P 500 stocks during February 2004 through October 2016, they find that: