Professor Hui and Professor Zheng propose a general framework for multi-level matrix factor models.

In the era of artificial intelligence, the capabilities for collecting and processing information have rapidly improved, enabling more efficient handling of complex data streams. Large-scale matrix-type time series have become increasingly common across diverse domains such as economics, finance, and industry, reflecting the growing integration of advanced computational methods.

Matrix factor models for modeling such time series have recently emerged as a focal point for econometricians, driven by their potential to uncover underlying patterns in high-dimensional data. The current basic assumption of these models is that only a common factor process influences all variables within the system. Yet, real-world data is often collected from disparate groups, and overlooking potential group structures can result in inadequate modeling and inefficient estimation. Consequently, incorporating a multi-level structure into matrix factor models is essential for robust analysis.

To address these challenges, a research team led by Associate Professor Hui Yongchang from the School of Economics and Finance at Xi'an Jiaotong University (XJTU), in collaboration with Professor Zheng Shurong from Northeast Normal University, has proposed a general framework for multi-level matrix factor models.

They provided detailed discussions on a two-layer factor model, comprising global factors and local factors, to better capture hierarchical dependencies. Both layers consist of unobservable matrix time series, with the global factor universally affecting all matrix time series and the local factor influencing only specific group-level series.

The team extended prior matrix white noise assumptions to accommodate error processes with weak linear correlations or nonlinear weak dependencies, enhancing model flexibility. New parameter estimation methods were introduced for this specification, and their effectiveness was rigorously validated through proofs and extensive simulation experiments.

These findings were recently published under the title Multilevel matrix factor model in the Journal of Econometrics, a premier journal in the field of economics.