Characteristics of ensemble forecasting systems for southeastern snowfall events

Abstract

The use of ensemble forecasting has burgeoned with the advent of greater technological resources. Forecasters now have available to them a range of possible forecast outcomes that can be utilized to understand forecast biases and to convey the most likely or worst case scenarios. However, ensemble forecasting systems must be used responsibly. If an ensemble is under-dispersive or heavily-biased, the observations may fall outside the ensemble, in which case it provides little useful information. Examining the behavior of the European Center’s Ensemble Prediction System (EC EPS), these characteristics are analyzed in terms of snow water liquid equivalent forecasts for the 2010-2015 DJF period. Selected for the advantages of a large member count, a linear regression model is created using bins populated by ensemble variances and error variances of the EC EPS, ordered by ascending ensemble variance, to calibrate the ensembles to provide a 1:1 prediction of the range of errors from the ensemble mean forecast. Training statistics indicate the model based on ensemble variance explains more than 90% of the error variance at all lead times. However, in-depth analysis of two out-of-sample, disparate winter weather events demonstrate how the linear regression model relies too heavily on the ensemble spread and can over-forecast or under-forecast depending on how large the ensemble spread is relative to the 2010-2015 DJF period. All things considered, the application of a linear model relating ensemble variance to error variance provides a promising means to adjust an ensemble system’s bias in estimating the range of possible outcomes for snowfall in the Southeastern US (SE US).