Snow much uncertainty: Probabilistic snowfall range perceptions, preferences, and understanding among the public

Snowfall forecasts are inherently uncertain. Small changes in temperature or mesoscale banding can lead to large variation in snowfall accumulation over short distances. For this reason, snowfall forecasts are often presented as a range of potential values (i.e., 2-4 inches). While snowfall ranges provide more detail than a single-value deterministic forecast, they do not fully characterize and explicitly communicate the uncertainty inherent to the forecast. Such probabilistic messaging has been a focus of the Weather Enterprise for decades (National Research Council 2006, Morss et al. 2008, Hirschberg et al. 2011, Novak et al. 2023), as well as the subject of empirical research. This research has demonstrated that probabilistic messaging can be understood by different end users and can improve decision outcomes (Joslyn & LeClerc 2012, Ripberger et al. 2022).

However, questions remain about when and how to effectively use probabilistic messaging (NOAA Science Advisory Board, 2021), especially in the context of snowfall forecasts. For instance, NWS can now create snowfall range forecasts using probabilistic thresholds of the ensemble forecast guidance, such as the 10th and 90th or 25th and 75th percentiles. However, this can sometimes lead to forecast snowfall ranges that can be quite wide (i.e., 2-8”), even if they do capture a higher likelihood that the actual snowfall will be in the forecast range. Some initial evidence suggests that key NWS partners (Rogers et al. 2022) and members of the public (Prestley et al. 2024) tolerate and even prefer this trade-off, especially when provided with explicit numeric confidence intervals (i.e., 80%). However, more evidence is needed to assess whether there are limits to how wide a range can be before the forecast is no longer useful to end users, or whether hybrid approaches like “goalpost” ranges that present a narrower range together with a reasonable best and worst-case scenario are more preferred in different contexts. These questions are all the more relevant due to the real-world nature of snowfall forecasting, where the distribution of forecast snowfall changes from event to event, and where zero might be a plausible low-end value for a snowfall range forecast.

In this presentation, we report on our efforts to empirically test how public preferences for, perceptions of, and understanding of probabilistic snowfall ranges varies based on these messaging and meteorological dimensions of probabilistic snowfall ranges. We investigated these dimensions as part of the Winter Weather and Society survey, a demographically representative sample of U.S. residents focused on winter weather perceptions and information usage. Our results augment existing empirical research to further inform the design and operationalization of probabilistic messaging for winter weather hazards.