关键词: 随机波动, 高频交易数据, 辅助粒子滤波, 马尔科夫链蒙特卡洛

Abstract: This paper analyzed more than 170,000 high frequency Chinese stock-market transaction data for the recent 15 years, our study finds that during morning 9:30 open return is significantly negative, while five minutes before the close call auction 3:00 PM stage significantly positive returns which is named as the phenomenon of fore and after 5 minutes. As the data has significant seasonal effect (cyclicality effect), we consider the effect of season SVJt modeling and provide the two-step estimation method of this model. The high frequency stochastic volatility model has the feature of huge data and serious computational load etc., the method of estimating, fitting and prediction evaluation the model should be made corresponding improvement. We calculate the marginal likelihood by the APF and make comparison with Bayesian factor methods. In the aspect of forecasting ability, we find that the seasonal effect SVJt-s model is better than the GARCH models and basic stochastic volatility models.

Key words: Stochastic Volatility, High-frequency Transaction Data, Auxiliary Particle Filter, Markov Chain Monte Carlo.