Formulation and Evaluation of a Calibrated Tornado Forecast Guidance Ensemble

Over the past few years, five different calibrated tornado guidance methods have been developed by various researchers associated with the University of Oklahoma and/or the Storm Prediction Center (SPC) of the National Weather Service. These methods are based on High Resolution Ensemble Forecast (HREF) data and are either formulated using machine learning or utilize known relationships among specific forecast environmental variables and tornado frequency. In the course of their operational duties, SPC forecasters have been considering (albeit in a purely experimental sense) the usefulness of these various methods as guidance tools. Although having a suite of guidance forecasts for a given day can provide a perspective on the range of possible outcomes, it was proposed that producing an ensemble average or percentile could be beneficial. Rather than manually attempting to ascertain trends across the suite of guidance products, ensemble statistics would computationally assimilate forecast information afforded by the set of calibrated guidance. The ensemble information also provides the degree of statistical consistency and agreement in forecast probability among the ensemble members.

Although using ensemble averages or percentiles to augment forecast guidance is not a new concept, and has been used for years for numerical weather prediction (NWP) model forecasts, such an approach has not been applied to date specifically for an ensemble of calibrated guidance products. In light of past studies that have shown that the average of an ensemble of NWP model forecasts often provides better guidance than any single member, this study investigates a similar premise that an ensemble average or percentile of the five calibrated guidance methods will provide superior forecast guidance as compared to any one calibrated product. Preliminary results from a limited set of cases has shown that the ensemble 90^th percentile statistically outperforms any one of the individual methods.