A Study On Volatility Forecasting Of CSI 100 Stock Index With High Frequency Data

The stock market is a barometer of economic development,which not only concerns the stability of economic operation,but also affects the financing effect of enterprises,and its reasonable operation plays an important role in financial market order and economic development.As an important attribute of stock index,volatility can be used to measure the risk level of financial assets and has important applications in financial risk management and other fields.Therefore,it is important to select an accurate and reasonable volatility forecasting model for research value.In this paper,the theoretical part mainly starts from the current situation of volatility research at home and abroad,combs the relevant literature about this paper topic,introduces the econometric method and the theory of long and short-term memory neural network model;the empirical part uses the high-frequency realized volatility obtained from the 5-min high-frequency data of CSI 100 index for empirical analysis.Firstly,the residuals are set to different distributions to construct the benchmark Realized GARCH model;Secondly,the LSTM model is improved by Batch Normalization,and the Attention mechanism is introduced to construct LSTM and AT-BN-LSTM two models;Finally,four different hybrid models are constructed in combination with Realized GARCH models,and using four common loss functions to evaluate the criteria of different models.The study shows that:1.the realized volatility calculated based on the 5-min high-frequency trading data of CSI 100 index can better characterize the distribution of the real volatility and is a more ideal proxy variable;2.for CSI 100 index,the prediction effect of the econometric prediction method and the prediction method based on the neural network model are different,besides,the machine learning model has a certain degree of improvement on the prediction effect compared with the econometric model;3.The prediction effect of the combined model is better than that of both the single neural network model and the econometric model,and the prediction order of the combined model also has an impact on the prediction accuracy.