Variable Selection Methods for Multivariate Time Series Data Using Multivariate Granger Causality

We study variable selection methods for multivariate time-series data.Hmamouche et al. proposed a method that first constructs a causal graph based on Granger causality among time-series data, and then selects variables from clusters formed by clustering the vertices corresponding to each variable.However, this method only performs pairwise Granger causality tests, which may not fully capture the interactions among variables.To address this issue, we propose a variable selection method that performs multivariate Granger causality tests on all combinations of explanatory variables with respect to the target variable, selecting the combination with the strongest causality.Our method successfully constructs a predictive model with a higher accuracy compared to the previous method.