Agnostic Fundamental Analysis via Machine Learning

ABSTRACT We construct stock mispricing signals by estimating fair values from financial statements using machine learning, minimizing data‐snooping bias. Signals derived from boosted regression trees and neural networks predict future stock returns and outperform linear benchmarks. We demonstrate that machine learning‐based signals capture nonlinear relationships between financial variables and firm values, providing incremental information beyond existing mispricing factors. This advantage improves valuation performance, particularly during volatile periods and for small firms. Our findings lend support to the validity of fundamental analysis and contribute to the growing literature on machine learning in finance.