2016-2019 Verification of the National Weather Service Probabilistic Snowfall Program

In the past, the National Weather Service has provided deterministic snowfall forecasts to its core partners and the public at large. These forecasts included small, subjective ranges in snowfall and were not based upon objective forecast guidance, such as ensemble spread. In many cases, observed snowfall was outside the forecast range. Also, many users require additional information about snowfall such as the probability of receiving at least an inch of snow. In 2015, the National Weather Service began experimenting with producing probabilistic snowfall information. Probabilistic snowfall forecasts were created by finding cumulative distribution functions (CDF) from the model ensemble spread. The CDF is then modified with both the Weather Prediction Center (WPC) and local weather forecast office (WFO) deterministic snowfall forecasts. Information from these probabilistic forecasts are then displayed on participating WFO web pages and used in support of local WFO Impact-based Decision Support Services (IDSS).

From winters 2016-2017 through 2018-2019, over half of the WFOs in the National Weather Service participated in the probabilistic snowfall experiment. As part of the experiment, observed snowfall and probabilistic forecasts were saved from both the WFO and WPC. This robust data set was then verified to determine the reliability of the probabilistic forecasts. A comparison of the WFO and WPC probability forecasts is done to determine if the addition of the WFO snowfall forecast added value to the probabilities originally derived from WPC. Our results show that the range between the 10th (5^th) and 90^th (95^th) is smaller for the WFO than for WPC. As a result, lower percentage of observed snowfall events are within the middle 80^th (90^th) percentile for the WFO compared to WPC. We also found that the 5^th and 10^th percentile accumulations were too high for snowfall events greater than 6 in. while the 90^th and 95^th percentile snowfall were too low for low (< 6 in. events). Brier Scores will be computed for different forecast hours and snowfall thresholds – such as probability of exceeding 6 in. – to illustrate how skill varies with forecast hour and by forecast snowfall amount. Overall, probabilistic forecasts are found to provide reliable information for IDSS.