CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: A novel computationally scalable high-dimensional vector autoregressive moving average model Authors:  Yao Zheng - University of Connecticut (United States) [presenting]
Feiqing Huang - University of Hong Kong (Hong Kong)
Guodong Li - University of Hong Kong (Hong Kong)
Kexin Lu - University of Hong Kong (Hong Kong)
Abstract: Classical VARMA models are very popular in modeling general linear processes due to their parsimony and favorable forecasting performance. Yet, the complicated identification issue and heavy computational burden hinder their practicality in the high dimensional regime. We introduce a scalable autoregressive moving average (SARMA) model that inherits the VARMA model's interpretability and the rich, dynamic structure of the VARMA model while avoiding the identification problem. Most notably, this family of multivariate linear processes contains virtually all VARMA processes and can also be easily extended to cover other forms of linear processes. In the high dimensional regime, we propose a low-Tucker-rank approach for further dimension reduction. Non-asymptotic error bounds for this model are derived, and a computationally scalable algorithm is developed. Simulation and real data analysis demonstrate the advantages of the proposed SARMA approach over existing methods.