Improving Seasonal Hurricane Forecasting for the Atlantic Basin Using Redundancy Analysis

Seasonal forecasting of hurricanes is a topic of great interest to the public, government and private sectors. To improve understanding of the dynamics controlling the predictability of hurricanes, and ultimately the accuracy of forecasts, multiple studies have related hurricane tracks and counts to empirically-defined indices such as the MJO and the NAO. We note that these indices were not developed to predict hurricanes but rather to summarize other aspects atmospheric variability.

In this study we use a statistical approach (Redundancy Analysis, RA) that allows us to use hurricane observations in the definition of the optimal predictive indices. To illustrate the approach we predict hurricane statistics for the North Atlantic derived from the Hurdat2 database for the period 1948 to 2016. We focus on all named storms with a start date between August and October. Storm counts were binned using an equal area grid that covered the North Atlantic. Seasonal mean sea level pressure (MSLP) was used to define a small number of indices that could be used to optimally predict spatial patterns of gridded hurricane counts and their time-varying amplitudes. The MSLP data were obtained from the NCEP reanalysis for the same season and years as the hurricane counts. Cross validation was used extensively to guard against overfitting.

The best MSLP predictor of the hurricane counts can be physically interpreted as the position and strength of the subtropical ridge which is known to be a major steering factor for hurricanes. Implications of this study for seasonal hurricane forecasting will be discussed.