My thesis contains four articles. Each article investigates the predictive accuracy of one or more behavioral game theory models for different games. Section 2 is based on a joint publication with Johannes Leder and Kinneret Teodorescu (Hariskos et al., 2011). The article is on learning models for predicting decisions from experience in forty market entry games.In Section 3 I investigate and extend the predictions of the social preference model of Fehr and Schmidt (1999) for the public goods game.Section 4 is joint work with Konstantinos Katsikopoulos and Gerd Gigerenzer. The article is on bargaining models for predicting decisions in one-shot ultimatum bargaining games.The main results of our study are that (1) the heuristic mix fits the outcomes observed in the ultimatum bargaining experiment no worse than the inequality aversion models and that(2) the heuristic mix makes better out-of-sample predictions for the three-person ultimatum game. Section 5 is based on joint work with Robert Böhm, Pantelis Pipergias Analytis and Konstantinos Katsikopoulos. The article is on equilibrium models and strategy mix models for predicting behavior in a broad set of one hundred and twenty extensive form games. In our article we investigate four research questions: (1) Is it possible to achieve better fitting and prediction results by specifying and estimating different equilibrium models that were used by Ert et al. (2011)? Our results show that the gap between the seven strategies model and the equilibrium models in fitting the data of the estimation experiment is smaller than expected; however the seven strategies model still predicts the data of the prediction experiment better.(2) How good are the predictions of our submitted models in comparison to the baseline models? Our results show that our second mover models predict the choice behavior in the prediction set of games better than each baseline model and that our first mover models are only outperformed by the seven strategies model. (3) How reliable are the predictions results of the competition? We check how reliable the prediction results of the competition are by comparing them to predictions results of two different cross validations. Our results show that the ranking of the models may change in the cross validations if the prediction results in the competition are close and that only groups of models with similar results that differ considerably between groups do not change ranks. (4) How can we achieve better predictions by combining predictions of different models? Our results show that simple averaging of predictions of good models yields better predictions than each individual model and that optimal predictions are only obtained if predictions of semi-good models that are not highlycorrelated to the predictions of the good models are included.