Man versus Machine Learning: Earnings Expectations and Conditional Biases

We construct a statistically optimal unbiased benchmark for the expectations of firms' earnings using machine learning. We show that analyst expectations are on average biased upwards and this bias exhibits substantial time-series and cross-sectional variation. The bias is increasing in the forecast horizon (on average) and analysts revise their expectations downwards as earnings announcement dates approach. We further find that the analysts' biases are associated with negative cross-sectional return predictability, and the short legs of multiple anomalies are comprised of firms for which the analysts' forecasts are too optimistic relative to our benchmark. Managers of companies for which analysts' upward biases are the largest, seem to take advantage of these biases by issuing stocks.