High-Frequency Data Analysis for Volatility Prediction of the CSI 300 ETF: An Empirical Research Using HAR Family Models

Abstract

In this paper, four HAR family models are used to study them and five-minute high-frequency trading data of the CSI 300 ETF is the target. Descriptive statistics is firstly conducted, and it is found that all variables have obvious autocorrelation and long-term memory which are suitable for time series analysis. Then, in-sample and out-of-sample analyses are implemented to test the predictive effect of each component on the volatility of the CSI 300 ETF and to compare the predictive ability of each model. The empirical results of the in-sample data show that daily and weekly realized volatility, continuous volatility, daily and monthly jump volatility, daily and monthly realized positive semi-variance, weekly and monthly realized negative semi-variance and positive and negative signed jump variation have stronger predictive effects on volatility of the CSI 300 ETF, while weekly realized volatility, jump volatility and realized positive semi-variance, and daily realized negative semi-variance have a weaker predictive effect on the CSI 300 ETF volatility. The results of the MCS test for out-of-sample prediction show that the HAR-RV-CJ model and HAR-RV-RSV model have significantly better out-of-sample predictive ability on the CSI 300 ETF volatility than the other two HAR family models, with HAR-RV-RSV model exhibiting the highest predictive accuracy in most cases. The above results indicate that the jump factor and the asymmetric factor of positive and negative returns can effectively improve the forecasting accuracy of HAR family models, so these two factors cannot be ignored in the construction of future HAR family models.