A1529
Title: Modewise additive factor model for matrix time series
Authors: Yuefeng Han - University of Notre Dame (United States) [presenting]
Abstract: Matrix-valued time series arise in domains like finance, retail analytics and environmental science, where capturing structured dynamics along both rows and columns is essential. We propose a modewise additive factor model for matrix-valued time series that separates latent dynamics along rows and columns, enabling, for example, store-level and product-level trend analysis. Under mild conditions, the loading spaces for both modes are identifiable up to orthogonal rotation, without requiring restrictive covariance or independence assumptions. To recover these spaces, we develop two estimation algorithms, Modewise INner-product Eigendecomposition (MINE) and COMplement-Projected Alternating Subspace Estimation (COMPAS), and establish their near optimal convergence rates under standard mixing and tail probability conditions. Through simulations and a retail sales case study, we demonstrate the methods accuracy, efficiency and interpretability, offering a versatile tool for structured temporal data analysis.
