Predicting North Atlantic Hurricane Tracks Through an Ensemble Approach Learned From Existing Methods
Abstract
Since the 19th century, hurricane prediction has been a key factor in promptly warning the public and consequently reducing costs. However, forecasting methods persist with issues on prediction errors and consistency, and in response, forecasters have created ensemble approaches to better refine the current methods and eliminate those two issues. We propose a real-time ensemble approach, denoted as Real-time Integrated Threshold Ensemble (RITE), built upon existing methods to enhance prediction accuracy and stability (a “method" will refer to pre-existing technique, and an “approach" will refer to a technique built upon those methods). RITE involves averaging predicted locations from a selected number of stable and top-performing methods, determined based on their prediction errors, then combining the methods together across several consecutive years to form an ensemble approach for the following year. We determine the best number of top methods per year that optimizes an objective function based on prediction errors. We apply RITE to predict hurricane tracks using data provided by the National Hurricane Center on North Atlantic hurricanes and assess its uncertainty through bootstrapping. Our results demonstrate that RITE can provide stable and accurate predictions, in contrast to the individual methods.