Panel vector auto-regressive (VAR) models are effective tools for capturing temporal relationships between a set of variables (e.g., macroeconomic indicators of an economy) while accounting for interdependencies between a set of entities (e.g., market sectors, or whole economies). For modeling macroeconomic data, this challenge is often further accentuated by the presence of variables observed at different frequencies. Existing Bayesian approaches that link entity-specific VAR models often impose strict fusion of VAR coefficients to a common value across entities. This paper develops a balanced and less stringent Bayesian approach for mixed frequency panel VAR models that employs group shrinkage prior distributions to borrow strength across entities, while allowing for entity-specific idiosyncrasies. A key novel feature is the ability to incorporate and learn the interdependence structure between entities through an inter- entity covariance (matrix) parameter. The proposed methodology is evaluated both on synthetic data and on two economic applications: employment indices across neighboring US states and macroeconomic indicators of tightly integrated European economies. Finally, we establish the theoretical properties of the proposed approach.
Co-author (s): Dr. Kshitij Khare, Professor in the Department of Statistics, University of Florida, Dr. George Michailidis, Professor in the Department of Statistics & Data Science, University of California, Los Angeles
Journal: Annals of Applied Statistics
