Prediction of North Atlantic tropical cyclone activity via Bayesian variable selection and Bayesian model averaging

Seasonal forecasting of tropical cyclones (TCs) activity for the North Atlantic region has gained popularity because it can provide basic information towards improved preparation against the socio-economic consequences of the weather and climate disasters. Sea surface temperatures (SSTs) during the hurricane season have been shown to be important factors and can predict TC activity well. But predictions need to be made before the beginning of the hurricane season, when the predictors are not yet observed. Several climate models issue forecasts of the SSTs. The existing literatures noted that not all climate models have similar predictive power, and considered various kinds of weighting schemes to combine the forecasts of SSTs from different climate models. Such models use the observed SSTs for estimation and use the forecasts of these SSTs for prediction. The main goals of this dissertation are: a) develop a fully Bayesian regression model for the frequency of tropical storms (response variable) that makes a distinction between the true SSTs and their forecasts, both of which (explanatory variables) are included in the model; b) simultaneously deal with missing predictors and perform variable selection, as all climate models do not have equally good predictive power; and c) build Bayesian multivariate regression models that can simultaneously analyze and predict the count of tropical storms, the count of hurricanes, and other measures of TC activity and severity which can be derived from the North Atlantic tropical storm dataset.