Seasonal Prediction of Tropical Cyclone Activity Near Taiwan Using the Bayesian Multivariate Regression Method

Abstract

Abstract A Poisson generalized linear regression model cast within a Bayesian framework is applied to forecast the seasonal tropical cyclone (TC) counts in the vicinity of Taiwan. The TC season considered is June–November and the data period used for model development is 1979–2007. A stepwise regression procedure is applied for predictor selection. Three large-scale climate variables, namely, relative vorticity at 850 hPa (Vor850), vertical wind shear, and sea level pressure over the western and central North Pacific from the antecedent May, are selected as predictors. Leave-one-out cross validation is performed and forecast skill is thoroughly evaluated. The skill level of the Bayesian regression model is better than what can be achieved by climatology and persistence methods. Most importantly, the Bayesian probabilistic inference can provide an uncertainty expression in the parameter estimation. Among the three predictors, Vor850 is found to be the most important because it reflects the variation of the ridge position of the westward extension of the western Pacific subtropical high. The model shows negative bias during the years with successive TCs, which are generated by easterly waves before approaching Taiwan. Recommendations for real-time operational forecast and future development are discussed.