Title: On the conjugate multivariate stochastic volatility processes
Authors: Kaoru Irie - University of Tokyo (Japan) [presenting]
Victor Pena - Baruch College, City University of New York (United States)
Abstract: Some of the multivariate stochastic volatility models for a dynamic covariance matrix are known to be conjugate and appealing to the sequential analysis of steaming data. There are two established conjugate models: the inverse Wishart-matrix beta process and multiple univariate inverse gamma-beta processes combined by Bartlett decomposition. The two models are closely related but not equivalent. While they can provide identical predictive distributions, the former model has more parameters than the latter one, being conservative in retrospective uncertainty quantification. In practice, the difference between the two models cannot be assessed by their marginal likelihoods but by other model comparison measures, including the deviance information criterion and mixture estimation model. We illustrate these points by the analysis of daily returns from currency exchange rates.