Spatial Clustering of Low Probability Events: Implications for Decision Support Services

Providing weather information to decision makers in a form that is understandable and actionable is critical to preparing and keeping the public safe from hazardous weather. Decision support services for hazardous weather increasingly rely on providing decision makers probabilistic weather information. Probabilistic information can provide both details on most likely as well as less likely outcomes that may impact the public or events. When providing decision support briefings, forecasters must decide what information to provide to partners. For example, the National Weather Service Probabilistic Winter Weather web page provides information on the most likely snowfall as well as the 10th and 90th percentile snowfall. Yet, if the probabilistic forecast is reliable, 20% of observations will be outside of this range.

The National Blend of Models (NBM) ingests multiple deterministic models and ensemble systems in order to create probabilistic snowfall and precipitation forecasts (in addition to other weather variables). The NBM is planned to be the NWS’s flagship guidance source and serve as the starting point for much of the official forecast information produced by the NWS. Probabilistic snowfall forecasts from the NBM were examined for several storms in the winter of 2023-2024 to learn how observations outside of the 10th-90th percentile (outlier observations) were spatially distributed. The spatial distribution of outlier observations was not random (or isolated) across the domain. Most outlier observations clustered in areas that encompassed several hundred kilometers even for reliable ensemble forecasts. When ensemble forecasts were underdispersive and/or skewed, outlier observations could be in larger clusters. Within some clusters, the observations were several centimeters above (below) the 10th (90th) percentile snowfall. Some clusters also increased in area as the snow event approached.

Outlier events can cover large portions of states resulting in impacts that are greater or less than forecast. Unforecast impacts can negatively affect decision makers and the public. A survey of NWS forecasters asked how they brief partners when expecting an outcome that is greater (less) than the 10th (90th) percentile. These results show that strategies will need to be developed that provide decision makers information about the low probability of different impacts due to outlier events. To anticipate outlier events and alert decision makers before they happen, forecasters will also need to analyze observations and numerical model data as the precipitation event approaches.