Genesis of New Atlantic Tropical Cyclones: Which Model Should You Trust?

It's been a remarkably quiet late August in the Atlantic for hurricanes, with nary an "Invest", and precious few yellow 10% probability circles on NHC's Tropical Weather Outlook. When will this remarkable hurricane drought come to an end? The computer models we use to track hurricanes have conflicting ideas about this. Which model should we believe, and how far into the future can these models successfully predict genesis events of new tropical depressions in the Atlantic? Some answers come in a 2013 paper by a group of scientists led by Florida State's Daniel Halperin, just accepted for publication in the journal Weather and Forecasting, "An evaluation of tropical cyclone genesis forecasts from global numerical models." It turns out that two of the three most reliable models for predicting the genesis of tropical cyclones up to four days in advance are also the our top-performing models for predicting hurricane tracks: the American GFS model and the European ECMWF model. The UKMET model is also quite good at predicting tropical cyclone genesis events. The Canadian GEM model and the Navy NOGAPS model (now succeeded by the Navy NAVGEM model) do less well. When two or more models make the same genesis forecast, the odds of the event actually occurring are increased considerably. The models have improved greatly in making genesis forecasts in recent years; back in 2007, when our top three models made a 4-day genesis forecast, these verified only 17 - 28% of the time. By 2011, the hit rate had increased to 45 - 50%. However, the models still miss most genesis events. In 2011, the probability of detection of a tropical cyclone genesis event was only 8% - 23%, meaning that 77 - 92% of time, a tropical depression or tropical storm formed without the model predicting that it would form. The best model to use for looking at Atlantic tropical cyclone genesis in 2011 was the UKMET model, which combined a relatively high probability of detection rate with a low false alarm rate.


Figure 1. The probability of making a correct tropical cyclone genesis forecast for all forecast hours (06 to 96) by model, for the period 2004 - 2011. In 2011, three models--the GFS, UKMET, and European--made genesis forecasts that were correct 45 - 50% of the time. The Canadian GEM model and the Navy NOGAPS model did less well. Data taken from Halperin et al., 2013, "An evaluation of tropical cyclone genesis forecasts from global numerical models." Weather and Forecasting, doi: http://dx.doi.org/10.1175/WAF-D-13-00008.1

The paper looked at 135 Atlantic genesis events over the period 2004 - 2011, as predicted by five global computer weather forecast models: the GFS, ECMWF, UKMET, NOGAPS, and Canadian GEM. The regional GFDL and HWRF hurricane models were not considered, as they use output from the GFS model to drive them, and are not good at making genesis forecasts. The study only looked at forecasts made up to four days in advance. While there have been cases where I've seen genesis predictions made 7 - 10 days in advance come true, these are rare, and the "hit" rate of successful genesis forecasts even four days in advance is low. In the paper, a "hit" was defined as a forecast that successfully predicted genesis within 24 hours and 345 miles of the observed time and location. A "false alarm" occurred when the model predicted a storm that never developed. The other categories considered were "late genesis" events where a storm formed more than 24 hours before it was predicted to, and "early genesis", where a storm formed more than 24 hours after it was predicted to. Here are some details on each model's ability to make Atlantic tropical cyclone genesis forecasts:

European ECMWF model: The model is reluctant to predict genesis, and misses many genesis events (it had only an 8% probability of detection in 2011.) However, when it does predict genesis, it usually happens, with only a 16% false alarm percentage in 2011. The European model exhibits preferred regions of genesis, with over 60% of its genesis forecasts occurring in the MDR: 10-20°N, 60-20°W. Although the ECMWF misses many tropical cyclone genesis events in the Gulf of Mexico (as do the other models), when it does forecast genesis there, it almost always occurs. The model was highly prone to making late genesis forecasts.

American GFS model: The GFS model improved substantially in its genesis forecasts beginning in 2010, most likely due to a major model upgrade in 2010. The GFS is more aggressive at predicting genesis than the European model, and is less likely to miss a genesis event (22% probability of detection in 2011.) However, the incidence of false alarms was 32% in 2011, double what the European model had. Like the European model, the GFS exhibits preferred regions of genesis, with nearly 60% of its genesis forecasts occurring in the MDR: 10-20°N, 60-20°W.

UKMET model: The UKMET is more aggressive at predicting genesis than the European model, and is less likely to miss a genesis event (20% probability of detection in 2011.) The incidence of false alarms was 18% in 2011, similar to what the European model had. Like the European and GFS models, the UKMET exhibits preferred regions of genesis, with more than 67% of its genesis forecasts occurring in the MDR: 10-20°N, 60-20°W. The model was prone to making late genesis forecasts.

Canadian GEM model: The Canadian model was the least likely to miss a formation event, with a 23% probability of detection in 2011. False alarms have been a major issue, though, and the Canadian model generated the second highest number of bogus genesis events of any of the five models evaluated (42% of all its genesis forecasts in 2011 were false alarms.) The Canadian model does not seem to have a preferred region of genesis—all types of genesis events occur across the entire basin. The model performs best over the main development region (MDR; 10-20°N, 60-20°W), Caribbean Sea, and Gulf of Mexico (GOM).

Navy NOGAPS model: This model was retired in 2012, but we may be able to assume that its successor, the NAVGEM model, will have some of the same characteristics. False Alarms have been a major issue, and the NOGAPS model generated the highest number of bogus genesis events of any of the five models evaluated (60% of all its genesis forecasts in 2011 were false alarms.) The model also had a low probability of detection, just 9%. NOGAPS model does not seem to have a preferred region of genesis—all types of genesis events occur across the entire basin. Despite the high levels of hurricane activity in 2004 and 2005, the NOGAPS model failed to successfully forecast any genesis events those years.

Sources of Model Data
You can view 7-day ECMWF and 16-day GFS forecasts on wunderground's wundermap with the model layer turned on.
Longer ten-day ECMWF forecasts are available from the ECMWF web site.
FSU's experimental hurricane forecast page (CMC, ECMWF, GFDL, GFS, HWRF, and NAVGEM models)
NOAA's HFIP model comparison page (GFS, ECMWF, FIM, FIM9, UKMET, and CMC models.)
Experimental HFIP models

Very Quiet in the Atlantic
A tropical wave that came off the coast of Africa on Sunday is midway between Africa and the Lesser Antilles Islands. This disturbance is moving westward at 15 mph, has a modest amount of spin, but only a limited amount of heavy thunderstorm activity. In their 8 am EDT Thursday Tropical Weather Outlook, NHC left the 5-day odds of formation of this disturbance at 20%, but increased the 2-day odds of formation to 10%. Wind shear is a moderate 10 - 20 knots over the system, but there is an area of dry air and dust from the Saharan Air Layer (SAL) to the north and west that is interfering with development. With the exception of the NAVGEM model, there is little support from the models for developing the disturbance during the next five days. The wave could spread heavy rains and gusty winds to the Lesser Antilles Islands on Sunday and Monday.

A tropical wave expected to emerge from the coast of Africa on Friday and track over the Cape Verde Islands is also showing little support for development from the models. This wave is expected to take a northwesterly track, and would likely not be able to make the long trek across the Atlantic to threaten North America or the Caribbean Islands. In their 8 am EDT Thursday Tropical Weather Outlook, NHC put the 5-day odds of development at 20%.

Wunderblogger Lee Grenci has an excellent new post, "Atlantic Hurricane Season: The Saharan Air Layer and Vertical Wind Shear".

California's Rim Fire Now 6th Largest in State History, and is 30% Controlled
California's massive Rim Fire in Yosemite National Park continues to grow, but an army of over 4,000 firefighters are making headway against the blaze, and had attained 30% containment of the fire as of 9:20 am EDT on Thursday. According to Inciweb, fire has burned over 192,700 acres. This moves the fire into 6th place for largest fire in state history, according to statistics tabulated by wunderground's weather historian, Christopher C. Burt in his latest post, "The Worst Wild Fires in U.S. History". The Rim Fire will have difficultly surpassing California's largest fire on record--the Cedar Fire in San Diego County of October 2003. That fire burned 273,246 acres (430 square miles). California has had its driest year-to-date period, so it is no surprise that the state is experiencing an unusually large fire this summer.


Video 1. Time-lapse photography shows various perspectives of the 2013 Rim Fire, as viewed from Yosemite National Park. The first part of this video is from the Crane Flat Helibase. The fire is currently burning in wilderness and is not immediately threatening visitors or employees. The second half of the video is from Glacier Point, showing Yosemite Valley, and how little the smoke from the fire has impacted the Valley.

Jeff Masters