Abstract

A plethora of models of learning has been developed and studied in macro-economic models in recent years. In this paper we will try to discriminate between these learning models by running laboratory experiments with incentivized human subjects. Participants predict inflation rates for 50 successive periods in a standard overlapping generations model and are rewarded on the basis of their forecasting accuracy. The information set for each participant contains the past inflation rates and the participant's own past predictions which, in turn, determine the actual inflation rate. We consider two treatments, with a low and a high level of monetary growth, respectively. We find that the level of convergence to the monetary steady state is significantly lower and volatility of inflation rates higher in the second treatment. Constant gain learning algorithms, such as adaptive expectations with a low adjustment parameter, seem to provide a better description of the experimental data than decreasing gain algorithms, such as (ordinary) least squares learning. Moreover, many participants switch between prediction strategies during the experiment on the basis of poor performance of their initial prediction strategy.