VAR and GSTAR-BASED feature selection in support vector regression for multivariate spatio-temporal forecasting

Abstract

Multivariate time series modeling is quite challenging particularly var(--highlight-yellow); color: inherit;">in term of diagnostic checking for assumptions required by the underlying model. For that reason, nonparametric approach is rapidly developed to overcome that problem. But, var(--highlight-yellow); color: inherit;">feature var(--highlight-yellow); color: inherit;">selection to choose relevant input becomes new issue var(--highlight-yellow); color: inherit;">in nonparametric approach. Moreover, if the multiple time series data are observed from different sites, then the location possibly play the role and make the modeling become more complicated. This work employs Support Vector Regression (SVR) to model the multivariate time series data observed from three different locations. The var(--highlight-yellow); color: inherit;">feature var(--highlight-yellow); color: inherit;">selection is done based on Vector Autoregressive (VAR) model that ignore the spatial dependencies as well as based on Generalized Spatio-Temporal Autoregressive (GSTAR) model that involves spatial information into the model. The proposed approach is applied for modeling and forecasting rainfall.