Multi-year prediction model of North Atlantic hurricane activity

An empirical prediction algorithm is developed to assess the potential of u seful. multi-season forecasts of North Atlantic hurricane activity. The alg orithm is based on combining separate univariate autoregressive moving aver age (ARMA) models for each of three dominant components of hurricane activi ty. A Bayesian criterion is used to select the order of each model. In a si ngle retroactive hindcast experiment, the algorithm is found to make better hindcasts than an ARMA model of the detrended series. A real-time forecast of hurricane activity for the 1997 North Atlantic hurricane season proves to be more accurate than two competitive single-season forecast models. It is expected that the routine use of the forecast algorithm in an operationa l setting will result in only marginal skill against climatology; it could however offer considerable forecast Value as realized by benefits to decisi on makers in the reinsurance industry.