Belief Formation: An Experiment With Outside Observers

In this paper we investigate the necessary ingredients for an accurate model of belief formation. Using experimental data from a previous experiment, we bring in a new group of subjects whose job it is to predict the action choices of the subjects from the previous experiment. While the rules we consider are all, strictly speaking, adaptive (being based on past observables), some of the variables we uncover represent fairly sophisticated behaviour. Going from less to more sophisticated, we find that the following are important components of the belief formation process: the history of play, payoffs (whether real or ``imagined" in the sense of Camerer and Ho (1999)) of the player whose actions our subjects are predicting and the payoffs of the other player. The paper also documents the presence of subject-specific heterogeneity in both initial beliefs and, to varying degrees, almost all of the variables found to influence beliefs.