Scoring Systems for Probabilistic Forecasts of Ranked Categories based on the RPS (Invited Presentation)

In 1969 Edward Epstein introduced the ranked probability score (RPS) into the meteorological literature. From its humble beginnings as a note in the Journal of Applied Meteorology, the RPS evolved to become a major tool in modern forecast evaluation. The score has assumed particular importance since the advent of ensemble forecasting. Ensemble forecasting systems provide routine probabilistic predictions for virtually any variable predicted by numerical models. The RPS and closely related scores have become an important tool to meet the demands of the evaluation of this probabilistic information.

This paper will review the evolution of the RPS from its origins based on utility theory to the form of the score commonly used today. The paper will discuss the partitioning of the score into components measuring reliability, resolution, and uncertainty, as well as its extension to the continuous ranked probability score, used to evaluate forecasts expressed as continuous probability distributions. A brief comparison of the RPS to other evaluation tools commonly used evaluation of probabilistic forecast information will be also made.