A1868
Title: Reduced rank time variation for vector-valued models
Authors: Luca Pedini - Polytechnic Marche University (Italy)
Alessandro Celani - Orebro University (Sweden) [presenting]
Alessandro Celani - De Nederlandsche Bank (Netherlands)
Abstract: A reduced-rank framework for time-varying parameter VAR models is proposed that tackles the curse of dimensionality in standard specifications. Conventional TVP VARs involve a number of latent states that grows quadratically with system size, making estimation infeasible in moderate dimensions. The approach preserves the matrix structure of time-varying coefficients and introduces a bilinear decomposition driven by row- and column-specific latent processes. This representation captures rich and heterogeneous dynamics across equations and variables while ensuring that the number of states grows only linearly. As a result, it delivers substantial computational gains without sacrificing flexibility. A Bayesian estimation strategy based on efficient precision sampling is developed and identification challenges arising from the multiplicative structure are addressed. Compared to existing factor-based approaches, the proposed model provides a more flexible allocation of time variation, allowing both row and column effects to evolve over time. Monte Carlo and empirical evidence show that the framework achieves accurate inference and strong forecasting performance in high-dimensional macroeconomic applications, making it well suited for settings where standard TVP-VARs are computationally prohibitive.
