High-Frequency Financial Time Series Prediction Based On Lmproved EMD-LSTM

Today’s international financial markets are often changing rapidly.As my country’s economy becomes more and more internationally connected,the relationship between my country’s economy and the global economy is getting closer,and high-frequency financial data in the financial market continues to emerge and become a record of instantaneous changes important data resources.High-frequency financial data is a microscope of the financial market.Compared with traditional financial data,financial high-frequency data itself has the characteristics of small fluctuations and rapid trend repetition.It has important research significance for understanding the structure of financial micro-markets and can better reflect stock indexes.Changing trends and assessing risks have positive guiding significance for investment decision-making.However,the research of financial high-frequency data faces the following problems:(1)There are a lot of noise trading in Chinese stock market.Because Chinese market is currently in a transitional stage from the era of retail investment to the era of institutional investment.Investors are still keen to obtain various gossips in the market,lack long-term investment ideas,and formulate investment strategies based on gossip that has nothing to do with company fundamentals and wrong subjective ideas.(2)High-frequency financial data generally do not obey the assumption of white noise and stationarity.However,traditional time series methods mostly set the white noise assumption by default,and require high data stability.However,neural network models that are highly versatile and deviate from model assumptions mostly do not consider data timing issues.How to choose a suitable forecasting model is the second problem of financial high-frequency data.(3)The large volume of high-frequency financial data will make the model construction time significantly longer.This problem will become more serious when combined with complex models.The greatest significance of high-frequency financial data is timeliness.Excessive training time will Greatly shorten the research significance.In order to solve a large number of common noise trading problems in high-frequency financial data.In this paper,based on the Empirical Mode Decomposition(EMD)method,the threshold empirical mode decomposition(TEMD)method improved by the change-point technology is used to denoise the 1-minute data of the SSE 50 ETF.deal with.And use this method to compare the traditional wavelet domain de-noising(WDD),ensemble empirical mode decomposition method(EEMD)and standard EMD.The results show that the TEMD method is advantageous to other models in terms of greater signal-to-noise ratio,and the noise reduction effect is better than the three comparison models.Aiming at the problem of non-stationary financial high-frequency data and inconsistent with the white noise assumption,this paper uses the Long Short Time Memory(LSTM)method of neural networks for timing problems.This method successfully solves the problem of ignoring data timing factors in feedforward networks,and at the same time solves the problems of gradient disappearance and gradient explosion in recurrent networks,and it breaks away from the data assumptions of white noise and stationarity.Finally,considering the large volume of financial high-frequency data,and the fact that it needs to adjust a large number of hyperparameters after being combined with a complex LSTM network,the calculation efficiency is low.This paper proposes a sub-grid search(SGS).Method to optimize the model parameters,thereby improving the efficiency,so that the optimized network can be updated with the daily intraday suspension time.For high-frequency financial data with frequent trend changes,this advantage has high practical significance.Through the combination of the noise reduction method TEMD,the hyperparameter optimization method SGS and the LSTM method,the prediction research on the SSE50 ETF data found that this method is compared with similar methods in terms of root mean square error,trend prediction accuracy,timeliness and other indicators Both have certain advantages.It shows that the improved LSTM model using these two methods has a significant performance improvement in the research of financial high-frequency data.Except for theoretical review and method introduction,the research of this article can be summarized into four parts.The first part is to select the 1-minute SSE 50 ETF data from July 22 to 26,2019 as the experimental object and analyze the statistical indicators,stationarity,skewness,kurtosis,Hurst index and other indicators of the data.It is found that the experimental data does not conform to the characteristics of stationarity and normality,and is not suitable for traditional time series modeling.Finally,the Hurst index judges that the data has long memory and can be used for LSTM network modeling.In the second part of this article,on the basis of determining that the experimental data can be used for LSTM model modeling,the data set is divided into a training set,a validation set and a test set,which are used together with SGS to optimize network hyperparameters,validation is further introduce to improve the robustness of the results,and finally the construction of the LSTM network is completed.The third part is to decompose the experimental data by EMD to obtain the Intrinsic Mode Function(IMF)components.The energy function of IMF component is constructed to judge the change point of IMF component,and threshold noise reduction is performed on the IMF component of the noisy part,thereby constructing the TEMD noise reduction method,and compare the noise reduction performance with WDD,EMD,EEMD.The fourth part is after completing the construction of LSTM model and TEMD,WDD,EMD,EEMD methods.The TEMD-LSTM,WDD-LSTM,EMD-LSTM,EEMD-LSTM,standard LSTM and BP network are used together to carry out a modeling comparison study using experimental data,and to compare and analyze the comprehensive performance of the model.