Machine Learning Guided to Avoid Overfitting

What modeling techniques help avoid biases/overfitting in use of machine learning to predict stock returns? In his July 2023 paper entitled “Less is More? Reducing Biases and Overfitting in Machine Learning Return Predictions”, Clint Howard explores how modeling choices affect machine learning as applied to predicting next-month stock returns, as follows:

• He considers 11 machine learning methods encompassing ordinary least squares, elastic net, random forest, gradient boosted regression trees, deep neural networks with one to five layers, an ensemble of the five neural networks and an ensemble of all methods.
• Initially, he uses the first 18 years of his sample (March 1957 to December 1974) for model training and the next 12 years (January 1975 to December 1986) for validation. Each December, he retrains with the training sample expanded by one year and the validation sample rolled forward one year.
• He trains all 11 machine learning models either on all firm/stock data together or separately on distinct groups of large, medium-sized and small firms, with size-based predictions subsequently merged.
• For each of the two sets of predictions each month, he sorts stocks into tenths, or deciles from highest to lowest predicted excess return and reforms a hedge portfolio that is long (short) the tenth, or decile, of stocks with the highest (lowest) predicted excess returns.

He calculates breakeven portfolio frictions (zero alpha) for multi-factor models of stock returns, including a 6-factor (market, size, book-to-market, profitability, investment, momentum) model. Using a database of 206 monthly firm/stock characteristics during March 1957 through December 2021, he finds that: (more…)