Verification of the NWS Probabilistic Storm Total  Snow Forecasts
Dave Radell, NOAA/NWS/Eastern Region Headquarters, Bohemia, NY
Phil Schumacher, NOAA/NWS/WFO Sioux Falls
Jeff Waldstreicher, NOAA/NWS/Eastern Region Headquarters

During winter 2016-2017, 44 Weather Forecast Offices (WFOs) conducted an experiment in producing probabilistic storm total snowfall information as a means to communicate uncertainty in the forecast.  The forecast technique utilizes NCEP's Weather Prediction Center's (WPC) probabilistic snowfall percentile accumulation values to create a cumulative probability distribution function at every grid point, utilizing the official NWS WFO storm total snowfall forecast as the modal value.  Graphics of various snowfall accumulation percentiles, probability of exceeding different snowfall thresholds, and the probability of the snowfall being within specific ranges are output and made available to end users for decision making. The 10th and 90th percentile values are communicated as ”expect at least this much” and the “potential for this much” storm total snowfall, respectively, while the official NWS WFO storm total snowfall is communicated as the most likely snowfall amount.

Given that individual cumulative distribution functions (CDFs) were created at each grid point within a WFO Country Warning Area (CWA), a unique opportunity exists to verify these forecast CDFs directly with observed distributions.  Numerous point locations east of the Rocky Mountains, for total of approximately 1000 forecast-observed pairs, were used for verification over the NWS Central and Eastern Region WFO CWAs for several events during the 2016-2017 snow season.  Forecasts were verified against observed storm total snowfall reports from WFO Public Information Statements, cooperative observer reports via xmACIS2, and/or computed from CF6 reports. Several statistical metrics were used to assess forecast skill from both a deterministic and probabilistic perspective, for both synoptic and lake effect events.  Metrics including bias and mean error, percent correct within the 10th-90th range, cumulative reliability diagrams, cumulative ranked probability scores and probability integral transform (PIT) histograms were generated and used to access accuracy, reliability and over/under dispersiveness of the underlying ensemble distribution.  Verification results from this past season’s winter experiment will be presented.