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

We comprehensively examine the extent to which machine learning technology can improve earnings forecasts, and if so, whether market expectations appear to reflect those superior forecasts. We find consistent evidence that machine forecasts generate smaller forecast errors than analysts’ forecasts. However, this relation only holds when the machine uses the analysts’ forecast as an input, and is greatest when the machine focuses on correcting predictable analysts’ forecast biases. Furthermore, this relation is decreasing over time, and it is economically small, except for firms in the smallest size quintile as well as for forecasts of longer horizons. Finally, we find that investors’ earnings expectations largely align with the more accurate machine forecasts than with analysts’ forecasts, but investors nevertheless appear to place more weight on analysts’ forecasts than the machine forecast does. Furthermore, investors’ overweighting on analysts’ forecasts appears to be smaller for firms with high institutional ownership and is decreasing for all firms over time. Our results have important implications for the literature on market expectations and forecasting, as well as for researchers seeking guidance on when to use machine forecasts for earnings expectations.