Comparing Probability Forecasters: Basic Binary Concepts and Multivariate Extensions

Abstract : In the applied forecasting literature much attention has been lavished on questions about the evaluation of probability forecasts, and the subjectivist view of probability has been invoked to aggregate probability forecasts over a diverse set of events or statements. One criterion invoked in such evaluations is that of calibration: a set of statements or events is considered and we ask if x percent of those assigned probability x of being correct prove to be correct, for each value of x. From this perspective, weather forecasters generally have been found to perform well. What is especially helpful in the evaluation of such probability forecasters is that they make forecasts about a long sequence of events (e.g. rain on a given day), and thus it makes sense to think about probability functions associated with the forecasts. In this paper we focus on a criterion for comparing forecasters, refinement, which goes beyond that of calibration.