In this paper, we study the economic interpretability of machine learning models. We find that nonlinearity of machine learning models helps capture mispricing signals accentuated by information uncertainty. In the cross-section, nonlinear machine learning models outperform linear models in firms with higher information uncertainty proxied by stock return volatility, earnings volatility, analyst dispersion, and firm age. The results are robust after excluding microcap stocks. In the time-series, those superior performances are more pronounced following heightened investor sentiment periods.

JEL classification: To be included

Keywords: Information Uncertainty, Machine Learning, Conditional Autoencoder, Deep Learning