Title: Filtering for stochastic volatility models with leverage by mixture approximation
Authors: Kaoru Irie - University of Tokyo (Japan) [presenting]
Yasuhiro Omori - University of Tokyo (Japan)
Naoki Awaya - University of Tokyo (Japan)
Abstract: The approximation of non-Gaussian distributions by finite mixture of normals is commonly used in Bayesian analysis of macroeconomic and financial time series models that are typically state space models with non-Gaussian observations such as stochastic volatility models. We apply this approximation to the SV models to realize its sequential analysis by the customized version of the existing sequential Monte Carlo methods of particle filtering/learning. The leverage parameter, which is the correlation of volatility and observation, can be sequentially sampled with the other parameters, utilizing the conditional normality of posteriors available under the mixture approximation. The approximation bias is also sequentially corrected through particle filtering. The proposed computational method is illustrated by the sequential analysis of the univariate and multivariate SV models with simulation and real data.