Messy Asset Pricing: Can AI Models Lead to a Consensus?

Abstract

We investigate the extent to which modern academic machine learning models agree on which factors are priced in the cross-section of returns. We find that models disagree to a surprising extent. The majority of average returns attributed to factor exposure by any given model is generally deemed pure-alpha by other models. We provide a theoretical model with many predictors that provides an impossibility theorem for creating a consensus model using machine learning and available data.

Keywords: asset pricing, machine learning, factor models, stochastic discount factors, mean-variance efficiency, random forests, neural networks, prediction, Sharpe ratio optimization

JEL Classification: G10, G12