Climatological and Synoptic Patterns of Atlantic Tropical Cyclones Categorized by GFS Track Ensemble Spread and Error

The advent of ensembles permits forecasters to have an implied level of confidence based upon the level of (dis)agreement among those ensembles. However, there are occasionally situations where the ensemble members may agree but be in large error. Such events not only mislead forecasters but also may undermine public confidence in the forecast when they occur more than rarely, or even during a single impacting event. Accordingly, the purpose of this research is, first, to identify and quantify any relationships between NCEP (National Center for Environmental Prediction) GFS (Global Forecast System) ensemble track spread and error for tropical cyclones (TCs). Second, it seeks to determine factors that can lead to unique combinations of ensemble spread and error. Of particular interest in this study are the aforementioned cases for which there is low spread among the ensembles' track forecasts, yet high error results. The GFS was used to analyze 2004-2011 Atlantic TCs. Forecast track ensemble spread and error were analyzed through forecast hour 120. Normalized error and spread values were calculated first as a single lifetime value for each TC; second, as a function of forecast hour for each TC; and third for each six-hourly forecast segment for each storm. For each of the three analyses, terciles (high, medium, and low) of both spread and error were determined, giving nine error/spread combinations. Climatological, synoptic, and physical characteristics are examined for four of the nine combinations: high spread/high error, low spread/low error, high spread/low error, and low spread/high error. A statistically significant relationship was observed between GFS ensemble spread and resulting track error when analyzing the TC's lifetime-total spread and error (r=0.78; p<0.01). Track forecasts with low spread among ensemble members, yet high resulting error were rare, however (three of 81 TCs). When observing the storm spread and error as a function of forecast hour it was found that there is a statistically significant relationship between track forecast error and standard deviation among all forecast hours (r ≈ 0.54 – 0.79, p < 0.01). Expectedly, this relationship is stronger for early forecast hours compared to later ones. In the third analysis, where forecasts of the same forecast hour were not averaged, error was conditioned on ensemble spread. The error distributions of each spread group (low, medium, and high) for each forecast hour (12, 24, 36, 48, 72, 96, and 120) were analyzed. It was found that mean track forecast error increased from low to medium to high ensemble track forecast spread groups for all forecast hours. These error distributions were fit to a gamma distribution and randomly sampled to test for significance in the differences among high, medium, and low ensemble spread groups. Differences were statistically significant for all comparisons among forecast hours through 48, but not for all comparisons among forecast hours 72 – 120. These results suggest that (known) ensemble spread can be a useful predictor for (as yet unknown) ensemble mean forecast error in short to medium term forecasts, although the direction of that error cannot be known. A low spread/high error forecast was observed at least once in 61.8% of all storms analyzed. There were three regions where this combination occurred more frequently: 1) western Gulf of Mexico, 2) western Caribbean, and 3) western Atlantic near the Bahamas. Noteworthy differences existed in the mean 300 hPa height and wind fields among certain spread/error groups when analyzing certain regions. For example, the mean 300 hPa trough position distinguished low spread/low error forecasts from low/spread high error forecasts in Region 3. However, there was little distinction in the mean 300 hPa synoptic setup among spread/error groups in other regions, such as along the United States E. coast above 35°N. Physical factors such as topography and interaction among multiple TCs also may play a role in the resulting spread/error combination.