Evaluating Tropical Cyclogenesis Forecasts from Four Global Numerical Models

Tropical cyclone (TC) forecasts rely heavily on output from numerical models. Each model in the suite of models used by forecasters has its own strengths and weaknesses. Some research has investigated the skill of the various models with respect to track, with the assumption that a TC already exists. However, little research has considered how well (or poorly) global models forecast TC genesis. Some studies have considered the Western North Pacific Basin, but there have been numerous upgrades to the numerical models since then. A few studies examined the North Atlantic Basin, but they analyzed a relatively small sample of storms. This paper will analyze TC genesis forecasts in four global models (CMC, GFS, NOGAPS, UKMET) over eight seasons (2004-2011) in the North Atlantic Basin.

All model indicated TCs will be counted and classified as a hit, false alarm, or an incorrect timing event. We will also consider miss events. The method of finding TCs in the model environment is based on a mixture of methods used previously in the literature. Hits are defined as when a model predicts genesis within 24 hours of the National Hurricane Center best track genesis time and within 5° latitude and longitude of the best track genesis location. A false alarm is defined as a model indicated TC that never develops. An incorrect timing event is defined as when the model is predicting genesis at a location where a TC already exists in the best track, but the timing of genesis is off (i.e., more than 24 hours from the best track genesis time).

Results will show which model best predicted TC genesis (with the acknowledgement that the “best” model can change from year to year) and whether or not model performance has improved over time. Basic statistics will be conducted on the results. The results will be subdivided into geographical regions and analyzed spatially and temporally. This may provide insight regarding regions where a model performs best (or worst) and whether forecast skill decreases with increasing forecast hour, as one would expect.

Future research will investigate the “whys” behind the results presented here. Synoptic patterns will be analyzed, and commonalities among each category (hits, false alarms, misses, and incorrect timing events) will be investigated.