Consensus of Dynamical Tropical Cyclone Track Forecasts—Errors versus Spread

Abstract

The relationships between consensus spread of five dynamical model tracks and the consensus mean error is explored for a western North Pacific tropical cyclone database of 381 cases. Whereas a small spread of the five tracks is often indicative of a small consensus track error, some cases with large errors also are found even though the consensus spread is small. Some of the success of the dynamical model consensus approach arises because a substantial number (21%) of the cases with a large consensus spread have reduced errors after the consensus averaging. In nearly all the cases in this sample, the best of the five models has a 72-h track error of less than 300 n mi, but no tools are available to allow the forecaster to always select this best model. It is demonstrated that the forecaster can also add value by forming a selective consensus after first eliminating one or more likely erroneous track(s) and averaging the remaining tracks. Conceptual models and symptoms in the predicted fields to assist the forecaster in this error detection have been separately described by the authors, and their successful application would result in more accurate selective consensus forecasts than nonselective consensus forecasts.