Predicting human behavior in unrepeated, simultaneous-move games

It is commonly assumed that agents will adopt Nash equilibrium strategies; however, experimental studies have demonstrated that this is often a poor description of human players' behavior in unrepeated normal-form games. We analyze five widely studied models of human behavior: Quantal Response Equilibrium, Level-k, Cognitive Hierarchy, QLk, and Noisy Introspection. We performed what we believe is the most comprehensive meta-analysis of these models, leveraging ten datasets from the literature recording human play of two-player games. We first evaluated predictive performance, asking how well each model fits unseen test data using parameters calibrated from separate training data. The QLk model (Stahl and Wilson, 1994) consistently achieved the best performance. Using a Bayesian analysis, we found that QLk's estimated parameter values were not consistent with their intended economic interpretations. Finally, we evaluated model variants similar to QLk, identifying one (Camerer et al., 2016) that achieves better predictive performance with fewer parameters.