Optimal Forecasting Groups

Abstract

This paper characterizes the optimal composition of a group for making a combined forecast. In the model, individual forecasters have types defined according to a statistical criterion we call type coherence. Members of the same type have identical expected accuracy, and forecasters within a type have higher covariance than forecasters of different types. We derive the optimal group composition as a function of predictive accuracy, between- and within-type covariance, and group size. Group size plays a critical role in determining the optimal group: in small groups the most accurate type should be in the majority, whereas in large groups the type with the least within-type covariance should dominate. This paper was accepted by Peter Wakker, decision analysis.