|
Abstract:
|
In recent studies on financial time series, univariate and multivariate stochastic volatility (SV) models have been widely explored, but they are difficult for making inference. We consider a class of multivariate SV models. This class is based on the assumption that observed time series of length N is m-dimensional, and latent process is k-dimensional. Also assume we can observe more series and increase m, but not k. For k=1, the class reduces to the multiplicative factor model. For inference, we construct the likelihood by integrating out latent variables, using the saddlepoint approximation. Therefore, computational expense is largely reduced. By increasing m, we can enjoy the asymptotic property of the approximation. With a similar setting to that in Jacquier et. al. (1994), our simulation results show that estimates are improved by increasing either m or N.
|
