Profitable Machine Learning Stock Picking Strategies?

Can machine learning models pick stocks that unequivocally generate alpha out-of-sample? In their November 2023 paper entitled “The Expected Returns on Machine-Learning Strategies”, Vitor Azevedo, Christopher Hoegner and Mihail Velikov assess expected net returns and alphas of machine learning-based anomaly trading strategies. They use nine machine learning models to predict next-month stock returns based on inputs for up to 320 published anomalies, added to the mix according to respective publication dates:

• Ordinary Least Squares with Huber Loss Function (OLS-HUBER).
• Elastic Net (ENET).
• Feedforward Neural Network (FFNN) with two to five hidden layers (FFNN2, FFNN3, FFNN4 and FFNN5).
• Long Short-Term Memory (LSTM) with one or two hidden layers (LSTM1 and LSTM2).
• ENSEMBLE, a combination of all other models.

They train the models using an expanding window, with the last seven years reserved for six years of validation and one year of out-of-sample-testing. During the test year, they each month reform a portfolio that is long (short) the value-weighted tenth, or decile, of stocks with the highest (lowest) predicted next-month returns. They then calculate actual next-month gross returns and 6-factor (market, size, value, profitability, investment and momentum) alphas during the test year. To calculate net returns and alphas, they multiply trading frictions estimated from historical bid-ask spreads times monthly portfolio turnovers. Using returns and firm characteristics for a broad sample of U.S. common stocks having data covering at least 20% of the 320 anomalies during March 1957 through December 2021, with out-of-sample tests starting January 2005, they find that:

(more…)