Aumentar Tamaño del texto Disminuir Tamaño del texto

Guilherme Moura, Universidade Federal de Santa Catarina


TVP-VARs: Specification, Estimation and Model Selection


The impact of model specification and of prior selection in the context of time-varying parameter vector autoregressive models (TVP-VAR) is investigated. Ignoring heteroscedasticity can exaggerate the time variation of parameters, while modeling multivariate stochastic volatility based on univariate processes leads to models that are dependent on the ordering of the observable variables. To avoid these problems, we propose the use of a multivariate stochastic volatility model based on the Wishart distribution, which generates order invariant TVP-VARs. Additionally, the large number of latent variables and parameters used in TVP-VARs increase the risk of overfitting. To alleviate this problem, we propose the use of shrinkage priors. An application based on six endogenous variables describing the U.S. economy is used to illustrate the advantages of the proposed approach.