Title: Bayesian nonparametric time varying vector autoregressive models
Authors: Jim Griffin - University of Kent (United Kingdom)
Maria Kalli - Kings College London (United Kingdom) [presenting]
Abstract: Although stationary time series models are theoretically appealing, macroeconomists consider them to be too restrictive. A popular alternative framework is time varying vector autoregression, with or without stochastic volatility (TVP-VAR or TVP-SV-VAR). Under this framework the parameters of the stationary vector autoregressive (VAR) model are allowed to change over time. This accounts for nonlinearity in the conditional mean, and heteroscedasticity in the conditional variance. We considered a Bayesian nonparametric stationary VAR (BayesNP-VAR) model and found that it outperformed the TVP-SV-VAR in terms of out-of-sample prediction for monthly macroeconomic series from the USA and Eurozone. Our aim is to extend the Bayes NP-VAR model to a time varying parameter specification, creating a nonparametric, non-stationary.