Development of a Statistical Model for Seasonal Prediction of North Atlantic Hurricane Numbers

Wednesday, 17 December 2014
Kyle Davis and Xubin Zeng, University of Arizona, Tucson, AZ, United States
Tropical cyclones cause more financial distress to insurance companies than any other natural disaster. From 1970-2002, it is estimated that hurricanes caused 44 billion dollars in damage, greater than 2.5 times the the next costliest catastrophe. Theses damages do not go without effect. A string of major catastrophes from 1991-1994 caused nine property firms to bankrupt and caused serious financial strain on others. The public was not only affected by the loss of life and property, but the increase in tax dollars for disaster relief. Providing better seasonal predictions of North Atlantic hurricane activity farther in advance will help alleviate some of the financial strains these major catastrophes put on the nation.

A statistical model was first developed by Bill Gray’s team to predict the total number of hurricanes over the North Atlantic in 1984, followed by other statistical methods, dynamic modeling, and hybrid methods in recent years. However, all these methods showed little to no skill with forecasts made by June 1 in recent years. In contrast to the relatively small year-to-year change in seasonal hurricane numbers pre-1980, there has been much greater interannual changes since, especially since the year 2000. For instance, while there were very high hurricane numbers in 2005 and 2010, 2013 was one of the lowest in history.

Recognizing these interdecadal changes in the dispersion of hurricane numbers, we have developed a new statistical model to more realistically predict (by June 1 each year) the seasonal hurricane number over the North Atlantic. It is based on the Multivariate ENSO Index (MEI) conditioned by the Atlantic Multidecadal Oscillation (AMO) index, the zonal wind stress and sea surface temperature over the Atlantic. It provides both the deterministic number and the range of hurricane numbers. The details of the model and its performance from 1950-2014 in comparison with other methods will be presented in our presentation.