Comparing Learning Models with Ideal Micro-experimental Data Sets

This paper compares three learning models. The models compared are the Reinforcement Learning Model of Erev and Roth, the EWA model of Camerer and Ho and a Belief Learning model in which beliefs are elicited from subjects using a proper scoring rule (The Stated Belief Model of Nyarko and Schotter). We Þnd that the Stated Belief Model outperforms both of the others in dramatic fashion and is capable of predicting not only the behavior of subjects but the period-to-period changes in their behavior in a far superior manner. We suggest that the reason for this performance has to do with the fact that eliciting beliefs and using them as inputs into the belief learning model provides us with the ”ideal data set” upon which to compare these models.