Bayesian Semiparametric GARCH Models

This paper aims to investigate a Bayesian sampling approach to parameter estimation in the semiparametric GARCH model with an unknown conditional error density, which we approximate by a mixture of Gaussian densities centered at individual errors and scaled by a common standard deviation. This mixture density has the form of a kernel density estimator of the errors with its bandwidth being the standard deviation. The proposed investigation is motivated by the lack of robustness in GARCH models with any parametric assumption of the error density for the purpose of error-density based inference such as value-at-risk (VaR) estimation. The contribution of the paper is to construct the likelihood and posterior of model and bandwidth parameters under the proposed mixture error density, and to forecast the one-step out-of-sample density of asset returns. The resulting VaR measure therefore would be distribution-free. Applying the semiparametric GARCH(1,1) model to daily stock-index returns in eight stock markets, we find that this semiparametric GARCH model is favored against the GARCH(1,1) model with Student t errors for five indices, and that the GARCH model underestimates VaR compared to its semiparametric counterpart. We also investigate the use and benefit of localized bandwidths in the proposed mixture density of the errors.