Verification of Multi-Model Ensemble Forecasts for North Atlantic Tropical Cyclones

Within the last decade, operational ensembles have allowed for probabilistic forecasts of TC track and intensity. However, verification of these ensembles has been limited, leaving unanswered questions. Do the ensembles show greater deterministic skill than deterministic models, and does the skill improve when combining multiple ensemble systems? Do the ensembles have systematic biases or dispersion issues in a particular direction relative to the observed track? How does the probabilistic skill of the ensembles compare against the NHC’s climatology-based cone of uncertainty?

This study verifies the medium-range (day 3-5) track and intensity forecasts of three global ensembles (51 ECMWF members, 23 UKMET members, and 21 GEFS members) and their combination (Multi-Model Global: MMG) for the 2008-2016 North Atlantic seasons. The results are compared with several other models, such as the deterministic ECMWF (ECdet), HWRF, GFDL, and the NHC’s official forecast (OFCL). The probabilistic skill of the ensembles is quantified using Brier Skill Scores, with the ECdet and OFCL cone as references. The NHC’s best-track data is used as the verifying analysis.

The mean total track errors of the MMG, ECdet, and OFCL are among the lowest of the models studied. All models have a significant negative along-track bias by day 5, while there is little bias in the cross-track direction. Much of this along-track bias is attributed to TCs that underwent extratropical transition. All EPSs are underdispersed in the along-track direction, while the ECMWF is slightly overdispersed in the cross-track direction. The MMG and ECMWF track forecasts have more probabilistic skill than the ECdet and comparable skill to the OFCL cone forecast. TC intensity errors for the HWRF and GFDL are lower than the coarser models within the first 24 h, but are comparable to the ECdet at longer lead times. The ECMWF and MMG have comparable or better probabilistic intensity forecasts than the ECdet, while the GEFS’s weak bias limits its skill.