MODELING AND FORECASTING SEASONAL PRECIPITATION IN FLORIDA - A VECTORTIME-DOMAIN APPROACH

Some major stages of vector autoregressive (AR) modelling consisting o f specification, order determination, estimation, diagnostic checking, and forecasting are described. To illustrate the utility of multivari ate AR models for climate teleconnection research, the seasonal Southe rn Oscillation Index (SOI) and a precipitation index in Florida are mo delled jointly. Two bivariate AR processes, order one and order four, are selected as candidate models to represent the joint series. To mak e a comparison of model performance, one-season-ahead forecasts of Flo rida precipitation are generated from the bivariate AR models, as well as from a simple regression model, for recent years that are independ ent from the period on which the model is constructed. Results show th at seasonal forecasts from the two-dimensional, order-one model are so mewhat better than the other two models. One new and important finding of this study is that, in addition to winter precipitation forecast s kill, there is also a forecast skill for autumn precipitation. The phy sical mechanisms by which these teleconnections might arise are briefl y discussed. Other applications of vector models are suggested.