Expectations Matter : When (not) to Use Machine Learning Earnings Forecasts

We comprehensively examine the superiority of various statistical methods and analysts’ forecasts in predicting earnings to provide an updated view on the best surrogate of earnings expectations. We also evaluate the impact of specification choices used in recent studies that implement ML models to forecast earnings by exhaustively comparing 3,024 models derived from the full combination of nine sets of specification choices and six machine learning (ML) algorithms. Our findings show that only a few key specification choices significantly affect the accuracy of ML model forecasts, while the others have negligible effect. Our analysis, along with codes and estimates, provide a much needed bridge between the earlier literature and recent ML earnings forecasting studies. We also find that the extent to which ML models are better at forecasting earnings than analysts is nuanced. Contrary to the impression created by recent studies, we find no compelling advantage to shifting from the well-established and readily available benchmark, analysts’ forecasts, to ML earnings forecasts for applications that require near-term earnings expectations. Even for two-year ahead earnings, ML superiority over analysts is economically insignificant among large-cap stocks. Furthermore, ML superiority has been decreasing over time. Finally, ML models outperform analysts only when the model uses analysts’ forecasts as an input. Overall, our results suggest that analysts’ forecasts are likely to remain the preferred earnings forecast in situations where ML superiority is negligible: calculating next month’s earnings surprise and among large-cap stocks. However, ML forecasts can offer significant improvements in longer-term forecasts for small-cap stocks