A0879
Title: Reduced sources of error in time-varying parameter models
Authors: Rodney Strachan - The University of Queensland (Australia) [presenting]
Eric Eisenstat - The University of Queensland (Australia)
Joshua Chan - Purdue University (United States)
Abstract: There have been several advances in the estimation of large vector autoregressive models (VARs). Increasing the dimension of time-varying parameter VARs (TVP-VARs) remains a challenge. We propose an approach to increasing the number of variables we can model in a TVP-VAR that takes advantage of the strong correlations among the states. It has been shown, using principal component analysis, that the states can be modelled with a few factors. We specify a TVP-VAR with a reduced rank covariance matrix for the states such that we are able to significantly reduce the dimension of the states without reducing the dimension of the VAR. The specification of the reduced rank model induces manifolds as supports for the parameters. Using a judicious selection of parameter expansions and priors for the expanding parameters, we develop a specification that is fast, efficient and easy to compute. The original model specification and mapping to the expanded model uses homeomorphic transformations such that estimation and inference is invariant to the ordering of the states in the model.
