第6回数理科学研究センター談話会
『第6回数理科学研究センター談話会』D. B. Nugroho氏を招き， Parameter Estimation in Log-normal Stochastic Volatility Models: Griddy Gibbs versus Metropolis-Hastings と言う題目でご講演いただきます．
第6回 数理科学研究センター談話会
4月25日（水）16:00-17:00
題目
Parameter Estimation in Log-normal Stochastic Volatility Models: Griddy Gibbs versus Metropolis-Hastings
講演者
Didit B. Nugroho
アブストラクト等は以下の通り：
The 6th Meeting of the Research Center of Mathematical Sciences, KGU
Title: Parameter Estimation in Log-normal Stochastic Volatility Models: Griddy Gibbs versus Metropolis-Hastings
Author: Didit B. Nugroho
Supervisor: Takayuki Morimoto, Ph.D.
Abstract: We compare the performance of two Markov Chain Monte Carlo (MCMC) samplers, Griddy Gibbs (GG) and Metropolis-Hastings (MH), to estimate parameters and latent variables in four log-normal stochastic volatility (LNSV) models. The daily stock indices we use are TOPIX and three stocks of the TOPIX Core 30: Hitachi Ltd., Nissan Motor Co. Ltd. and Panasonic Corp. , from January 2004 to December 2011. It is shown that the volatility by MH sampler is more persistent and less variable than those by GG sampler. Using the daily Realized Volatility as true volatility, it was found that the GG sampler is superior according to all six loss functions for TOPIX in the LNSV model with fat-tails and leverage effect, while the MH sampler is superior according to five loss functions for Hitachi in the SV model with leverage effect and for Nissan in the SV model with fat-tails and leverage effect and three loss functions for Panasonic in the SV model with fat-tails. Furthermore, GMLE (Gaussian quasi-maximum likelihood function) loss function is only minimized on all stocks in the same model and sampler: LNSV model with fat-tails and leverage effect by GG sampler. Keywords: Log-normal stochastic volatility, Markov Chain Monte Carlo, Griddy Gibbs sampler, Metropolis-Hastings algorithm, TOPIX