A Bayesian regression approach for predicting seasonal tropical cyclone activity

In this study, a Poisson generalized linear regression model is applied to predict the tropical cyclone (TC) counts series occurring in the Central North Pacific (CNP) in the peak hurricane season (JAS) from 1966-2003. The SLP, PW, SST, SOI, and CLIPER are chosen as the predictors. With a non-informative prior assumption for the model parameters, a Bayesian inference for this model is derived in detail. A Gibbs sampler based on Monte Carlo Markov Chain (MCMC) method is designed to integrate the desired posterior and predictive distribution. The proposed hierarchical model is physically based and also gives out a probabilistic prediction of seasonal TC frequency, which would better facilitate decision making. A cross-validation procedure is applied to the seasonal TC series in the CNP and rather satisfactory results are achieved.