Economic Models or Machine Learning Techniques? Evidence from Asset Pricing Models

Abstract

Previous comparisons of econometric and machine learning asset pricing models have focused on statistical measures such as the r-squared. In this paper I compare popular machine learning models with traditional factor models using the level of mispricing, as measured by the model's alpha, as the primary evaluation metric. In making this comparison I highlight where machine learning models cannot be implemented in the traditional asset pricing framework. For comparison to previous studies I also compare models based on forecast accuracy as measured by the out-of-sample r-squared. Using the 30 industry portfolios as test assets traditional factor models achieve smaller levels of mispricing and more accurate forecasts of asset returns. The benefits of deep learning recently documented in the literature appear limited to test assets with highly nonlinear returns.

Keywords: Asset Pricing, Machine Learning, Neural Networks, Mispricing

JEL Classification: G12