We use machine learning to construct a statistically optimal and unbiased benchmark for firms' earnings expectations. We show that analyst expectations are on average biased upwards, and that this bias exhibits substantial time-series and cross-sectional variation. On average, the bias increases...