We revisit La Porta’s (1996) finding that returns on portfolios of stocks with most optimistic analyst long term earnings growth forecasts are substantially lower than those for stocks with the most pessimistic forecasts. We document that the finding still holds, and present several further facts: a) firms with most optimistic forecasts have grown very fast in the past, but growth subsequently reverts to the mean; b) expectations of earnings growth also follow a boom-bust cycle, so extreme expectations are too extreme and mean revert; c) predictable low returns are associated with declines in fundamentals and expectations. Similar findings obtain, though more weakly, for firms with most pessimistic expectations. We develop a learning model in which analysts forecast future fundamentals based on the history of earnings growth, but beliefs are shaped by the representativeness heuristic: analysts update excessively in the direction of states of the world whose objective likelihood rises the most in light of the news. Intuitively, good earnings growth news predicts future googles but not as many as analysts believe. We analyze this learning model with diagnostic rather than rational expectations using Kalman filter techniques. The model delivers the empirical findings we initially document. It also has additional empirical implications that distinguish it from both the Bayesian learning and earlier behavioral models. We examine these implications empirically and find supportive evidence. Joint with with P. Bordalo, R. La Porta, and A. Shleifer.