A Research On Realized Volatility Forecast Based On EEMD-XGBoost Machine Learning Method

As the "barometer" of a country's economy,the stock market can realize the rational allocation and full utilization of resources,and investors can obtain considerable income in the stock market.However,risks and returns coexist,and how to achieve good management of stock market risks is a problem that cannot be ignored.In addition,the stock market is affected by many factors,and the influence of external variables on stock market volatility cannot be ignored either.So far,the heterogeneous autoregressive(HAR)model is one of the most classic models in volatility modeling research based on high-frequency data.Many scholars have proposed many improvements based on this model,such as simply decomposing it into continuous paths Changes and discontinuous jump path changes are studied.However,the realized volatility is a very complex nonlinear time series.This simple decomposition is difficult to understand its changing law,which is not conducive to our analysis and modeling.The ensemble empirical mode decomposition method(EEMD)can not only adaptively select fluctuation characteristics from the sequence,but also has robustness to the decomposition of nonlinear time series.In addition,the HAR model is essentially a linear model.It uses a linear model to model the realized volatility,ignoring its nonlinear characteristics.The machine learning method of XGBoost has excellent nonlinear fitting capabilities,and because of its The advantages of excellent learning effects have attracted wide attention.Based on the above background,this article elaborates on the research background,research significance,relevant research status and theoretical basis,and proposes a prediction model based on EEMD-XGBoost for the short-term,medium-term and long-term prediction problems of realized volatility.In the empirical study,the EEMD decomposition of the realized volatility series of the Shanghai Stock Exchange Index from January 5,2015 to December 25,2020 is carried out,and the decomposition series are reconstructed according to the zero-mean significance test result,and then the decomposed series is Realize the short-term,medium-term and long-term prediction problems of volatility,take different reconstructed sequences and the original sequence of realized volatility as input variables,respectively fit the XGBoost model,and compare the prediction effect with the EEMD-HAR model.In addition,the Economic Policy Uncertainty Index(EPU)and Volatility Index(VIX)are used as external variables to study their impact on the realized volatility.The main conclusions are as follows:(1)The sequence is decomposed by EEMD and reconstructed with zero mean significance test.The low-frequency sequence has a clear trend,which is roughly the same as the original sequence.This indicates that the low-frequency sequence should contain more predictions.Information;the trend item reflects the overall trend of the realized volatility,which proves the effectiveness of the EEMD decomposition method.(2)Fitting the XGBoost prediction model with the sequence decomposed and reconstructed by EEMD and the original sequence of the realized volatility as input variables,and comparing its prediction effect with the EEMD-HAR model.The forecast effect of monthly volatility has been achieved,and the EEMD-XGBoost model is significantly better than the EEMD-HAR model.(3)Regarding the impact of external variables on the realized volatility,the results show that the economic policy uncertainty index(EPU)and the Volatility index(VIX)do not significantly improve the forecasting effect of the daily and weekly realized volatility,but for The forecasting effect of the realized volatility on a monthly basis has been significantly improved,which indicates that the EPU index and the VIX index tend to have a longer-term impact on the realized volatility.