An efficient branch-and-bound strategy for subset vector autoregressive model selection

Abstract

A computationally efficient branch-and-bound strategy for finding the subsets of the most statistically significant variables of a vector autoregressive (VAR) model from a given search subspace is proposed. Specifically, the candidate submodels are obtained by deleting columns from the coefficient matrices of the full-specified VAR process. The strategy is based on a regression tree and derives the best-subset VAR models without computing the whole tree. The branch-and-bound cutting test is based on monotone statistical selection criteria which are functions of the determinant of the estimated residual covariance matrix. Experimental results confirm the computational efficiency of the proposed algorithm.