This paper develops a test for forecast bias that is robust to (a) correlated information amongst analysts; (b) common unforecasted industry-wide earnings shocks; (c) information arrival at different points of the forecast horizon; and (d) the possibility that the measure of earnings that analysts forecast differs from that which the econometrician observes. We find no empirical support for herding. On the contrary, analysts systematically issue biased contrarian forecasts that overshoot the publicly-available consensus forecast in the direction of their private information. We find that the forecast bias is economically large and declines with the amount of information at an analyst's disposal. The magnitude of the bias, its systematic variation with analyst following, and the pattern of bias in forecast revisions indicate that the bias is strategically chosen. In particular, the data cannot be explained by analyst myopia or analyst overconfidence