Title: Large time varying parameter VARs for macroeconomic forecasting
Authors: Gianni Amisano - Federal Reserve Board (United States) [presenting]
Domenico Giannone - FED New York (United States)
Michele Lenza - European Central Bank (Germany)
Abstract: Bayesian time varying parameter models have been widely used for forecasting purposes and as exploratory devices. In spite of their obvious overparameterization, these models are capable of reproducing salient features of the data. The main challenge is to use them with many endogenous variables. The problem is that VAR models are over-parameterized and each coefficient is endowed with its own drift term. Therefore the state equation shock covariance matrix becomes huge. In addition, in order to contain the effects of random walk dynamics on coefficients it is important to specify a prior that greatly limits the amount of time variation and to provide a sensible initialization. We propose to use an informative Bayesian VAR on the pre-sample to obtain a sensible initialization and calibrate the prior in a conservative way. In addition, we use the VAR Kronecker structure to reduce the number of free parameters in the state equation. We also augment the model with a suitable stochastic volatility specification. We conduct two experiments, the first one in VARs with 7 and 20 variables. We show that our specification works very well in prediction, both in terms of point and density forecasts, and yields substantial gains especially in periods of recessions and economic turbulence.