A0419
Title: Reordering variables in VARs with stochastic volatility: Implications for forecasting and structural analysis
Authors: Gergely Ganics - Banco de Espana (Spain) [presenting]
Florens Odendahl - Banque de France (France)
Abstract: Although it is known that the widely used lower triangular decomposition of the covariance matrix for Bayesian vector autoregressions (BVARs) with stochastic volatility is not invariant to the variable ordering, this issue has received little attention in applied work. It is documented that the ordering empirically matters in a reduced form forecasting exercise as well as for structural estimations, for both U.S. and euro area data. In particular, it is found that the ordering affects the quality of point and density forecasts and the shape of the impulse response function. To avoid the variable ordering problem, using an ordering-invariant autoregressive inverse Wishart (ARIW) process is proposed to model stochastic volatility. In the empirical results, the ARIW specification provides a competitive alternative to the specification using the lower triangular decomposition.
