Evolutionary Switching between Forecasting Heuristics: An Explanation of an Asset-Pricing Experiment

In this paper we propose an explanation of the findings of a recent laboratory market forecasting experiment. In the experiment the participants were asked to predict prices for 50 periods on the basis of past realizations. Three different aggregate outcomes were observed in an identical environment: slow monotonic price convergence, persistent price oscillations, and oscillatory dampened price fluctuations. Individual predictions exhibited a high degree of coordination, although the individual forecasts were not commonly known. To explain these findings we propose an evolutionary model of reinforcement learning over a set of simple forecasting heuristics. The key element of our model is the switching between heuristics on the basis of their past performance. Simulations show that such evolutionary learning can reproduce the qualitative patterns observed in the experiment.