Research And Application Of Time Series Model Based On Machine Learning

Time series analysis is an important way for people to understand the objective world and natural phenomena.In recent twenty years,the time series analysis methods have been developing rapidly,and have been widely applied in many practical fields.With their related technology,time series prediction has become a hot research field.With the development of time series analysis,time series prediction methods have been developing rapidly,and time series prediction plays an important role in many fields.Prediction is the basis of decision-making,and decision is the continuation of prediction.Therefore,accurate prediction is crucial to making a correct decision.In order to improve the prediction effect and application value of time series,many researchers have been devoted to the research of this problem.Especially in the rapid development of machine learning and deep learning methods,the research of time series analysis and prediction methods are developing rapidly,but their effect can not meet the high requirements of practical application in many practical fields,and there are still many problems need to be solved.Taking time series analysis and prediction as the research background,this thesis mainly studied the application of time series crosscorrelation analysis,time series multi-scale analysis and time series prediction.This thesis mainly carried out the following three aspects of research.In order to understand whether there are correlations among different phenomena,what kind of relationship exists between them and how intensity of the correlation between them,This thesis studied the interaction and common changes in financial market and the correlation analysis method of time series comprehensively and systematically.In order to reveal the dynamic behavior of financial time series,based on multi-scale,multi-attribute analysis idea and the MMA method,combined with Hurst index histogram distribution,this thesis proposed the DH-MXA method.Through the distribution histogram of the Hurst exponential surface of local correlation of two time series,the DH-MXA method can distinguish the nuances of local correlation between two time series.The DH-MXA method also can analyze the correlation between any two time series,and can give the type of correlation and the intensity of correlation between them.In order to predict the future change of the phenomenon effectively and accurately,it is necessary to understand the multi-scale characteristics and the internal fluctuation of the phenomenon,and observe and measure the fluctuation and multi-scale effect of the phenomenon at different scales.This thesis studied various methods for analyzing the multifractal characteristics of time series systematically.Particularly,this thesis studied the analysis methods for complex phenomena.Combined with the Rényi entropy and the MMA analysis method,this thesis proposed an analysis method of multi-attribute,multiscale and multi fractal for time series based on the Rényi entropy.The proposed method is referred to as the REM3 PA method.The REM3 PA method inherits the advantage of the Rényi entropy and the MMA analysis method.With the advantage of the MMA analysis method in the analysis of the multiscale of time series and the advantage of entropy in dealing with the problem of heavy tails of time series analysis,the REM3 PA method has the ability to analyze the multi-attribute,multi-scale and multi fractal of time series,can reveal the characteristics of short-term and long-term characteristics of financial time series,especially can reveal the relationship of volatility and importance between the attributes of time series.In order to improve the prediction effect of time series,after systematically studied various prediction methods of time series based on statistical technology and machine learning,the thesis compared the performance of a variety of prediction methods.At the same time,especially in the selection of experimental data,a lot of research has been done.The thesis try to select several representative time series data to test all kinds of prediction methods,since only good data can correctly check the validity and practicability of the prediction method.With the dividing idea of treating complex problem and the gradually optimized strategy of machine learning,this thesis proposed the LSTM-TFE,LR-TFE and BR-TFE methods based on the EEMD,LSTM neural network,LR and BR methods.These proposed methods used the EEMD method to decompose the complex time series into a number of relatively simple sub sequence with more regular and stable change.Then they used the LSTM neural network,LR and BR to train the predict model for each sub sequence by machine learning and to predict the future values for each sub sequence with the trained model.Finally,they added the prediction results of all sub sequences and formed the prediction results of the original complex time series.In order to verify the proposed method effectively,four representative time series data were selected for testing.From the experimental results,it was found that the proposed methods had good prediction effect and had the prospect of practical application.But there are also some defects,and the effect of time series prediction will be not good in the some time series with severe change.The limitation of the proposed method was pointed out,which pointed out the direction for the future application of the proposed method in the practical field.The experimental results of this thesis provide an important reference for the research and application of the time series prediction model.