A Bayesian Regression Approach for Predicting Seasonal Tropical Cyclone Activity over the Central North Pacific

Abstract

In this study, a Poisson generalized linear regression model cast in the Bayesian framework is applied to forecast the tropical cyclone (TC) activity in the central North Pacific (CNP) in the peak hurricane season (July?September) using large-scale environmental variables available up to the antecedent May and June. Specifically, five predictor variables are considered: sea surface temperatures, sea level pressures, vertical wind shear, relative vorticity, and precipitable water. The Pearson correlation between the seasonal TC frequency and each of the five potential predictors over the eastern and central North Pacific is computed. The critical region for which the local correlation is statistically significant at the 99% confidence level is determined. To keep the predictor selection process robust, a simple average of the predictor variable over the critical region is then computed. With a noninformative prior assumption for the model parameters, a Bayesian inference for this model is derived in detail. A Gibbs sampler based on the Markov chain Monte Carlo (MCMC) method is designed to integrate the desired posterior predictive distribution. The proposed hierarchical model is physically based and yields a probabilistic prediction for seasonal TC frequency, which would better facilitate decision making. A cross-validation procedure was applied to predict the seasonal TC counts within the period of 1966?2003 and satisfactory results were obtained.