The School of Economics invites you to an Econometrics seminar presented by Guillaume Chevillon (ESSEC Business School).

Long-memory prone priors for large systems within a finite-order VAR

Co-authors: 

Luc Bauwens (Université Catholique de Louvain) and Sébastien Laurent (Aix-Marseille University)

Abstract

We propose a class of prior distributions adapted to the modeling and forecasting of large dimensional systems within a vector autoregression (VAR) of finite-order. These priors are derived from the conditions stated in Chevillon, Hecq and Laurent (2018) who show that as the number of variables within a VAR(1) becomes very large, the ensuing dependence across units can generate fractional long memory in each variable individually. We extend and generalize their results and use these as means of a prior distribution for estimation in systems that are too large for efficient unconstrained multivariate inference. In applications to stock return realized volatilities and sectoral consumer price inflation, our suggested multivariate priors may yield substantial improvements in forecasting performance over standard techniques such as VAR, HAR or ARFIMA models.