Can deep neural networks outperform Fama-MacBeth regression and other supervised learning approaches in stock returns prediction with asset-pricing factors?

In asset pricing, most studies focus on finding new factors, such as macroeconomic factors or firm characteristics, to explain risk premiums. Investigating whether these factors help forecast stock returns remains active research in finance and computer science. This paper conducts an extensive comparative analysis using a large set of pricing factors. It compares out-of-sample stock-level and portfolio-level prediction performance among neural networks, the traditional Fama-MacBeth regression, and other supervised learning algorithms such as regression and tree-based algorithms. Our analysis shows the benefit of employing neural networks, and deeper neural networks enjoy marginal improvements in terms of prediction.