Partial mutual information for input selection of time series prediction

An important step in modeling time series is the selection of appropriate model input. Information theoretic concept of mutual information provides a general framework to evaluate the dependence between a potential model input and the output. A model-free approach, partial measure of the mutual information, is proposed in this paper, which utilizes a measure of the mutual information criterion to characterize the dependence in the case of multiple inputs and identifies the actual inputs for time series prediction. This algorithm is tested on a number of synthetic time series data sets, where the dependence attributes were known a priori. Results depict the effectiveness of the proposed method in proper input selection.