B0297
Title: Latent variable model for graph-estimation in multivariate stationary time series
Authors: Arkaprava Roy - University of Florida (United States) [presenting]
Abstract: Vector-valued multivariate time series data are routinely collected in many application areas. Although stationarity, causality and invertibility are very useful modeling assumptions for any time series data, methodological developments are limited under these assumptions for multivariate time series. Under some assumptions on the autocovariance matrices, we achieve those properties for a new class of Gaussian multivariate time series. In this proposed class, the normalized multivariate time series is assumed to be some orthogonal rotation of a set of independent latent univariate time series. To capture the graphical dependence structure among the variables, we also propose to sparsely estimate the marginal precision matrix and develop related computational methodologies. An efficient Markov Chain Monte Carlo (MCMC) algorithm is developed for posterior computation. We also study theoretical consistency properties. We show excellent performance in simulations and GDP data applications.