Research On Realized Volatility Prediction Of High Frequency Data Based On LSTM Model In Machine Learning

In recent years,with the outbreak of the economic crisis and the occurrence of emergencies such as the new coronal epidemic,the financial market has become turbulent,the stock market volatility has become more frequent,and the healthy development of financial markets in various countries has been seriously threatened.Therefore,volatility,as a measure of financial market risk,has become one of the current research hotspots.There are many factors that cause fluctuations in the financial market,such as economic environment,technological factors,government policies,etc.,which can have an impact on the financial market,leading to sudden and significant changes in stock prices,namely jumping behavior.Jumping behavior usually affects the accurate prediction of stock price volatility in financial markets,and is a factor that cannot be ignored.Effectively predicting and estimating the volatility of financial stock markets can help financial practitioners and regulators implement asset pricing,risk management,option pricing,and macro decision-making smoothly to a certain extent.At present,although there are a large number of literature predicting and analyzing realized volatility,it is not difficult to find that there is still room for improvement in the impact of accurate characterization of jumps on realized volatility based on empirical evidence.Therefore,this study is based on Chinese stock market data and constructs a new prediction model under the framework of realized volatility estimation to investigate the prediction of realized volatility with different jump variables.The specific content is as follows:Firstly,based on the theoretical framework,the realized volatility and HAR type volatility models were introduced in detail,and different types of jump variables were characterized,namely double power jump,significance jump,sign jump(where sign jump is also divided into positive sign jump and negative sign jump),median jump,volume jump,and real jump.Furthermore,the jump phenomenon in Chinese stock data was described from multiple perspectives.Secondly,the jump variables measured under different conditions are added to the HAR,HARQ,and HARQ-F models,and the HAR-Jump,HARQ-Jump,and HARQ-FJump realized volatility prediction models are constructed for comparative analysis to explore the predictive effects of different models.Finally,considering the normality assumption,nonlinear constraints,and leverage effects of traditional volatility prediction models,machine learning methods that can characterize nonlinear relationships are considered for analyzing stock data.At the same time,combining the advantages of traditional models and neural networks,a hybrid model is constructed to predict,and the advantages and disadvantages of the model are judged through loss function and MCS test.The results indicate that:(1)there is heterogeneity in Chinese individual stock data;(2)Jumping variables are a factor that cannot be ignored,and the prediction performance of the jumping model is better than that of the original model;(3)The hybrid model composed of neural network models and traditional models generally outperforms traditional models in predicting performance;(4)The MCS results indicate that HARQ-SVJ-LSTM has the highest prediction accuracy among the nine models,followed by HARQ-F-SVJ-LSTM.