We examine the efficacy of machine learning in a central task of fundamental analysis: forecasting corporate earnings. We find that machine learning models not only generate significantly more accurate and informative out-of-sample forecasts than the state-of-the-art models in the literature but also perform better compared to analysts’ consensus forecasts. This superior performance appears attributable to the ability of machine learning to uncover new information through identifying economically important predictors and capturing nonlinear relationships. The new information uncovered by machine learning models is of considerable economic value to investors. It has significant predictive power with respect to future stock returns, with stocks in the most favorable new information quintile outperforming those in the least favorable quintile by approximately 34 to 77 bps per month on a risk-adjusted basis.