A0506
Title: Matrix autoregressive time series with reduced-rank and sparse structures
Authors: Xiaohang Wang - The Education of Hong Kong (Hong Kong)
Ling Xin - Beijing Normal-Hong Kong Baptist University (China)
Philip Yu - The Education University of Hong Kong (Hong Kong) [presenting]
Abstract: Matrix- and tensor-valued time series models have been investigated as a solution to the dimensionality challenges in high-dimensional time series analysis. These models leverage multi-classification structures within data variables to decompose large interaction networks into smaller, more manageable components. To further address dimension reduction, recent studies have explored imposing structural constraints on the individual coefficient matrices of matrix- or tensor-valued time series models. The RR-S-MAR model is introduced, a matrix autoregressive (MAR) model of order one featuring a reduced-rank structure on the left matrix and a sparse structure on the right matrix. An alternating least-squares method is developed for estimating the constrained model, while a bootstrapping approach is employed for statistical inference. Additionally, an extended Bayesian information criterion is proposed for selecting tuning parameters within the model. Simulations are conducted to evaluate the performance of the estimation algorithm and the model selection criterion in finite samples. Finally, the model is applied to economic data to illustrate real-world analysis and interpretations.
