A High-frequency Volatility Forecasting Study Based On VPIN Model

With the advancement of technology, the high-frequency finance brought in two effects:the high-frequency data for research work to bring datanoise and lead to low-frequency model failure; high frequency trading to new research presented higher timeliness requirements. This article focused on high-frequency volatility analysis and forecasting, comparing the various proxies and related variables predictive models of high-frequency volatility. In empirical analysis, we used CSI 300 stock index futures as the underlying tick data. Also, we conducted a comparative analysis of the variable proxies, chose five minutes realized volatility as a proxy variable fit forecast. Taking into account the high frequency realized volatility weakened by significant microstructure noise, we use the latest findings on microstructure, VPIN, to reduce the impact ofmicrostructure noise. Using HAR-VPIN models to examine predictive validity of VPIN solved the shortage of VPIN’s prediction efficiency problem. Finally, combined withhighfrequency trader risk management practices, this paper presents a scenario of high frequency volatility forecasting model to analyze the application of high-frequency volatility.The main conclusions are:1. VPIN parameters sensitivity analysis of this article marking a different algorithm parameters such as the number of baskets, the starting point, and trading direction. The main question the study found that playing the wrong application of the wrong standard algorithms and time frame. VPIN computing robustness is controllable in high-frequency prediction research. HAR-VPIN regression model showed that, VPIN factor can fully explain the volume information, VPIN is the main driving factor of volatility.2. We Compare several high-frequency volatility proxy variables. The absolutecan not guarantee accuracy of measurement. High and low values suffers sample window problem. Using high-frequency data to calculate the realized volatility is a better approach. We recommend the use of five minutes frame as the volatility forecasting sample interval, both to avoid the microstructure noise at higher frequencies, and can prevent the sampling frequency is too low which cause the timeliness of the information is not useful.3. We compares various forecasting models, HAR-VPIN model has out performed under the majority of the loss function. Only under certain metrics its predictive ability is weaker than GARCH model. We believe including the " external information" that the other models do not have causing the prediction error can be significantly reduced.