Partial Mutual Information Based Algorithm For Input Variable Selection For time series forecasting

Abstract

In time series forecasting, it is a crucial step to identify proper set of variables as the inputs to the model. Many input variable selection (IVS) techniques fail to perform suitably due to inherent assumption of linearity or rich redundancy between variables. The motivation behind this research is to propose an input variable selection algorithm which not only can handle nonlinear problems and redundant data, but also is straightforward and easy-to-implement. In the field of information theory, partial mutual information is a reliable measure to evaluate linear/nonlinear dependency and redundancy among variables. In this paper, we propose an IVS algorithm based on partial mutual information. The algorithm is tested on three time series with known dependence attributes. Results confirm credibility of the proposed method to capture linear/non-linear dependence and redundancy between variables.