Comparison Of The Models Based On High Frequency Volatility

In this paper, based on high-frequency data and high-frequency volatility theory, we establish five high-frequency volatility models, Realized Volatility, Adjustment Realized Volatility, Realized Range Volatility, Realized Bipower Volatility,Two Scales Realized Volatility. On these five high-frequency Volatility estimators, we set up autoregressive fractionally integrated average model, and then carried out sample forecasts. Return series is assumed to obey normal distribution, student T distribution and the Generalized Error Distribution, we select 90%,95%,97% and 99% confidence level, and then calculate the value at risk,using the back testing we select the optimal value at risk models in different distribution and different confidence level. The article also established a GARCH model based on low frequency data, and calculate the Value at Risk under the model, and then compared the models based on high-frequency volatility.The main conclusions can be summarized as follows:First, in the 10% level of significance, based on the generalized error distribution the Realized Bipower Volatility model was the best. At the 5% significance level, from the perspective of testing, based on the distribution of T students Realized Volatility and Realized Range Volatility based on the generalized error distribution have been achieved the best result. However, from the views of the absolute value number of days between the actual failure and the expected failure, the Adjustment Realized Volatility based on student T distribution is best. In the 3% level of significance, from the perspective of the testing, Realized Range Volatility based on students T distribution achieved the best result. However, from the views of the absolute value number of days between the actual failure and the expected failure, Realized Volatility based on Student T distribution and Realized Range Volatility based on generalized error distribution performance the best. In the 1%significance level, Realized Range Volatility based on the students T distribution is best. Second, in the significance of different levels, we select value at risk models including Realized Volatility, Adjust Realized Volatility and Realized Range Volatility, which fully shows that the value of established risk based on high-frequency volatility model is better than based on the GARCH model in low-frequency data set.Third, on the whole, the generalized error distribution and student T distribution are used for a high degree of confidence testing, and normal distribution for the lower confidence test.