A maximum likelihood approach to combining forecasts

Abstract

We model an individual who wants to learn about a state of the world. The individual has a prior belief and has data that consist of multiple forecasts about the state of the world. Our key assumption is that the decision maker identifies explanations that could have generated this data and among these focuses on those that maximize the likelihood of observing the data. The decision maker then bases her final prediction about the state on one of these maximum likelihood explanations. We show that in all the maximum likelihood explanations, moderate forecasts are just statistical derivatives of extreme ones. Therefore, the decision maker will base her final prediction only on the information conveyed in the relatively extreme forecasts. We show that this approach to combining forecasts leads to a unique prediction, and a simple and dynamically consistent way to aggregate opinions.