Evolutionary Selection of Individual Expectations and Aggregate Outcomes∗

In recent ‘learning to forecast’ experiments with human subjects of Hommes, et al. (2005), three different patterns in aggregate price behavior have been observed: slow monotonic convergence, permanent price oscillations and dampened price fluctuations. These different aggregate outcomes are at odds with a representative agent who is fully rational or employs a single adaptive learning rule. We construct a simple model of individual learning, based on performance based evolutionary selection or reinforcement learning among heterogeneous expectations rules, explaining these different aggregate outcomes. Agents are boundedly rational and choose from a small set of simple price prediction rules, such as naive, adaptive or trend following expectations, similar to individual rules estimated in the experiment. Agents update their active rule by evolutionary selection based upon forecasting performance. Simulations show that after some initial learning phase, coordination on a common rule occurs. Out-of-sample predictive power of our model is higher than of the rational or other homogeneous expectations benchmarks. Our results show that heterogeneity in expectations is crucial to describe individual forecasting behavior as well as aggregate price behavior.