A series of experiments suggest that, compared to the Bayesian benchmark, people may either underreact or overreact to new information. We consider a setting where agents repeatedly process new data. Our main result shows a basic distinction between the long-run beliefs of agents who underreact to information and agents who overreact to information. Like Bayesian learners, non-Bayesian updaters who underreact to observations eventually forecast accurately. Hence, underreaction may be a transient phenomenon. Non-Bayesian updaters who overreact to observations eventually forecast accurately with positive probability but may also, with positive probability, converge to incorrect forecasts. Hence, overreaction may have long-run consequences.
©2011 Walter de Gruyter GmbH & Co. KG, Berlin/Boston