In this paper we study the belief formation processes of a group of outside observers making predictions about the actions of a player involved in a repeated game. We document four main results. First, there is substantial heterogeneity in the accuracy of our observers, with average accuracy being quite poor. Second, while there is no difference between the most and the least accurate observer in their initial beliefs, there are striking differences in their belief updating rules. The most accurate observers have a well-formulated model of player behavior, are good at best responding and quickly incorporate new information to their beliefs. The worst observers behave in an opposite manner on all three fronts. Third, when the game does not converge, subjects look beyond historical actions to make predictions and place