Consider a two-player discounted infinitely repeated game. A player's belief is a probability distribution over the opponent's repeated game strategies. This paper shows that, for a large class of repeated games, there are no beliefs that satisfy three properties: learnability, a diversity of belief condition called CSP, and consistency. Loosely, if players learn to forecast the path of play whenever each plays a strategy that the other anticipates (in the sense of being in the support of that player's belief) and if the sets of anticipated strategies are sufficiently rich, then neither anticipates any of his opponent's best responses. This generalizes results in Nachbar (1997).

Alternating-offer and demand bargaining models of legislative bargaining make very different predictions in terms of both ex ante and ex post distribution of payoffs, as well as in the role of the order of play. The experiment shows that actual bargaining behavior is not as sensitive to the different bargaining rules as the theoretical point predictions, whereas the comparative statics are in line with both models. We compare our results to studies that attempt to distinguish between these two approaches using field data, finding strong similarities between the laboratory and field data regardless of the underlying bargaining process. Copyright The Econometric Society 2005.