A0503
Title: Uncover networks of R\&D activities by a two-way constrained MAR model
Authors: Xiaohang Wang - N/A (Hong Kong) [presenting]
Philip Yu - The Education University of Hong Kong (Hong Kong)
Ling Xin - BNU-HKBU United International College (Austria)
Abstract: High-dimensional time series are traditionally monitored by vector time series models that have dimensionality and interpretation issues. Recently, an inspiring idea has been proposed to formulate the high dimensional data into a matrix or tensor time series structure when the variables have multi-way classifications. The new structures lead to a substantial dimensional reduction by ignoring the networks between classifications and focusing on within-class connections. It admits explicit interpretations of the within-class networks and inspires many applications. A matrix autoregressive model is formulated with two-way constraints (MAR-2C) to monitor R\&D activities at the firm or regional level. To impose low-rank constraints on the network among different R\&D activities and impose sparsity constraints on the network among different firms/regions, the model can highlight important information by the two matrix networks. A reduced-rank shrinkage-thresholding (RR-ST) algorithm is adopted in the model estimation, and a bootstrapping approach is used to make statistical inferences. The simulation results show that the RR-ST algorithm can achieve good accuracy. In real data analysis, R\&D activities are monitored in 31 regions of China during the past 15 years by the MAR-2C model. Meaningful information like lead-lag relationships among different R\&D activities, as well as the R\&D spillovers among different regions, are uncovered from the estimated networks.
