What Makes a "Good" Probabilistic Forecast?

In his 1993 Weather and Forecasting essay, Allan Murphy described three types of forecast "goodness": consistency, quality, and value. In recent years, multi-model numerical ensemble data and post-processing techniques have facilitated the ability to publish probabilistic information spanning a wide range of forecast time scales, from convective phenomena to seasonal forecasts. There are numerous possible ways to visualize and disseminate this information in the form of a probabilistic forecast product. For example, a probabilistic snow forecast might take the form of a table of reasonable low-end to high-end values at a list of forecast points (e.g., 10th to 90th percentile ranges), or might take the form of a graphic quantifying the probability of exceeding various thresholds at every forecast point over an area (e.g., more than 4 inches of snow in 12 hours). In some cases, the former product may have the higher value to users, while in other cases the latter product may have higher value. How do we measure Murphy's three types of "goodness" of these forecasts to determine the most effective probabilistic forecast product for a given scenario? This presentation will explore ways of measuring the "goodness" of a variety of proposed and actual probabilistic forecast products at various temporal scales, as well as its dependence on the forecast scenario. How the three types of forecast "goodness" interrelate as applied to probabilistic products will also be discussed. Finally, implications for effective communications of forecast and hazard information will be considered.