Synoptic Patterns of Atlantic Hurricanes Categorized by Track Error and Forecast Confidence

There are documented instances where several numerical tropical cyclone (TC) track forecasts agree, yet the actual track is different (e.g. Hurricane Alberto, 2000). A track forecast with high confidence, yet eventual high error such as this has high potential to hinder hurricane preparedness actions including excessive evacuation orders and increased failure to order necessary evacuations. Identifying commonalities in the synoptic set up among TCs with high confidence yet high error could potentially alert forecasters when forecast guidance may be misleading.

The NCEP Global Ensemble Forecast System (GEFS) is used to analyze 2004-2011 Atlantic TCs. Track forecast error and confidence were analyzed from forecast hours six through 120. Normalized track forecast error was calculated using the ensemble mean track and best track data. Normalized track confidence is represented by the standard deviation of track position among all ensemble members. Normalized error and confidence values were calculated first as a single value for a given TC, and later as a function of forecast hour. Error and confidence values were divided into terciles and each TC's track forecast error and confidence were defined as pairs of high, medium, or low error/confidence. Nine error/confidence combinations were then examined for common synoptic aspects, and four of the nine are focused on here: Type 1) low confidence and high error, Type 2) high confidence and low error, Type 3) low confidence and low error, and Type 4) high confidence and high error.

When analyzing the lifetime-total storm error and confidence, it was found that high confidence/low error and low confidence/high error storms are the most common situations. Although rare, highly confident yet highly erroneous track forecasts were found for 4 of 81 TCs. When observing the storm error and confidence as a function of forecast hour (rather than just lifetime mean as first done) it was found that there is a stronger statistically significant positive correlation between track forecast error and standard deviation in the early forecast hours compared to the later forecast hours. For example, when standard deviation is compared to error, R squared values range from 0.64 for forecast hour 12 to 0.31 for forecast hour 120. This implies that for early forecast hours, a forecast with a larger ensemble spread tends to have greater error than a forecast with a lower ensemble spread. This relationship becomes less apparent as forecast hour increases to the 120th hour. When forecast error was conditioned on forecast confidence, it was found that highly confident track forecasts had the smallest mean errors. The mean track forecast errors increased from high to medium to low track forecast confidence. Some of these results disagree with prior work on the topic (e.g. Hauke 2006), and the possible reasons for these differences will be discussed as well. The talk will conclude with an examination of the synoptic features that discriminate one Type of event from another and a demonstration of an experimental web page that highlights such features in real-time.