Summary Verification Measures and Their Interpretation for Ensemble Forecasts

Ensemble prediction systems produce forecasts that represent the probability distribution of a continuous forecast variable. Most often, the verification problem is simplified by transforming the ensemble forecast into probability forecasts for discrete events, where the events are defined by one or more threshold values. Then, skill is evaluated using the mean-square error (MSE; i.e., Brier) skill score for binary events, or the ranked probability skill score (RPSS) for multicategory events. A framework is introduced that generalizes this approach, by describing the forecast quality of ensemble forecasts as a continuous function of the threshold value. Viewing ensemble forecast quality this way leads to the interpretation of the RPSS and the continuous ranked probability skill score (CRPSS) as measures of the weighted-average skill over the threshold values. It also motivates additional measures, derived to summarize other features of a continuous forecast quality function, which can be interpreted as descriptions of the function's geometric shape. The measures can be computed not only for skill, but also for skill score decompositions, which characterize the resolution, reliability, discrimination, and other aspects of forecast quality. Collectively, they provide convenient metrics for comparing the performance of an ensemble prediction system at different locations, lead times, or issuance times, or for comparing alternative forecasting systems.