M. Kalli, J. Griffin
Linear Vector Autoregressive (VAR) models aim to capture the self and cross autocorrelation structure of multivariate time series (yt). Kalli and Griffin (2018) introduced a Bayesian nonparametric vector autoregressive model (BayesNP-VAR) where the stationary and transition densities are modelled as infinite mixtures. This allows for nonlinearity in the conditional mean, heteroscedasticity in the conditional variance, and non-Gaussian innovations. The transition density can be viewed as a mixture of experts because the component weights depend on previous lags of yt. This allows for different transition densities to be favoured at different periods. In this paper we introduce and describe a method of identifying which time series drive the change between regimes. We extend the BayesNP-VAR by introducing nonstationarity via dependent random measures. We describe the construction of the non stationary model, consider its properties, and discuss its out-of-sample predictive performance.
Palabras clave: Bayesian nonparametrics, mixtures of experts, multivariate time series, and vector autoregression (VAR).
Programado
GT11.BAYES3 Sesión Invitada
9 de noviembre de 2023 16:50
HC4: Sala Sacristía