Title: A Bayesian semiparametric stochastic volatility model with Markovian mixtures
Authors: Qiao Yang - ShanghaiTech University (China) [presenting]
John Maheu - McMaster University (Canada)
Chenxing Li - McMaster University (Canada)
Abstract: A previous Bayesian semiparametric stochastic volatility (SV-DPM) model is extended. Instead of using a Dirichlet process mixture (DPM) for return innovations that capture a constant unknown density, we use an infinite hidden Markov model (IHMM). This allows for time variation in the return density beyond that attributed to latent volatility. The new model (SV-IHMM) also nests the SV-DPM as a special case and greatly improves the density forecast from the SV model with Student-t innovations (SVt) compared to the SV-DPM. The model is applied to several applications, and a comparison is made with the Dirichlet process version. The results show that SV-IHMM generally requires fewer states than the SV-DPM on average and predicts better out-of-sample. Furthermore, predictive densities from the SV-IHMM exhibit clear distributional shifts over time. These results are robust to different hyperparameter prior values.