Title: The Hurricane Track Fit Consensus Model for Improving Hurricane Forecasting
Abstract: We present a new method for creating a model consensus to improve real-time hurricane track prediction. The method is based on the statistical fitting of historic numerical model track forecasts to the observed storm positions and learning from their historical errors and biases. Our method is closest to the HFIP Corrected Consensus Approach (HCCA) methodology while using an alternative model formulation. Our method creates a separate consensus model for each forecast hour making it possible to independently correct the bias of each input model for that specific hour. This approach, which we call the Hurricane Track Fit (HFIT) model, is computationally efficient and scalable to additional numerical models as input, and it produces interpretable coefficients weighing model contributions. The new method is evaluated for the 2014-2021 hurricane seasons in the Atlantic basin using the input from the best-performing operational track forecast guidance at the National Hurricane Center (NHC): the U.S. National Weather Service Global Forecast System deterministic and ensemble mean models, European Centre for Medium-Range Weather Forecasts deterministic model, the NWS Hurricane Weather Research and Forecasting model and the NHC equally weighted numerical model track consensus (TVCA). The results of the cross-validation for the 2014-2021 hurricane track dataset show that the HFIT consensus model consistently reduces the track forecast errors compared to those from the input models and the official NHC forecasts (OFCL). For example, at 24h the HFIT track forecast errors are smaller by 18.5% and 15.6% than those in AVNI and EMXI respectively, and 23% and 15% smaller at 72h. The HFIT forecasts show a reduction of errors compared to OFCL by 8.1% at 24h and 7.5% at 72h. We also discuss the successful real-time operational performance of HFIT during the 2022 hurricane season.