Financial Time Series Prediction Based On Multiscale Decomposition And Long Short-Term Memory Networks

The financial time series is non-linear,non-stationary,multi-scale and long-memory,which affects the accuracy of the financial time series prediction to some extent.Based on its different characteristics,many scholars have developed different solutions to improve the prediction accuracy.Considering the nonlinear and non-stationary characteristics,neural network has gradually become a mainstream prediction model.Considering the characteristics of multi-scale,the multi-scale decomposition method can solve the problem of data mesoscale aliasing.Considering the characteristics of long-memory,LSTM network with a special memory structure provides a new method to predict.In order to improve the accuracy of the prediction model,this study proposes a financial time series prediction model based on multi-scale decomposition and LSTM network.This study includes the following four main sections.Section one reviews that,the multiscale decomposition method can solve the phenomenon of data mode aliasing,of which Empirical Mode Decomposition and Ensemble Empirical Mode Decomposition are two main methods.Furthermore,the principle and process of those two methods are emphatically introduced,and the advantages of introducing EEMD decomposition method into financial time series analysis are well analyzed.Section two introduces the method of neural network,analyzes the model principle and existing problems of Recurrent Neural Network,and introduces the model structure and parameter design of Long Short-Term Memory,a neural model that improve itself based on RNN.Section three introduces the combined prediction model based on the idea of "decomposition,reconstruction and integration",and analyzes the data processing method,loss function selection and training optimization method of the model.Section four applies the EEMD-LSTM model to analyze the CSI 300 Index empirically,analyzes the multi-scale characteristics of the sequence according to the scale decomposition results,and evaluates the effectiveness and universality of the model.The main conclusions of this study are as follows.Firstly,EEMD decomposition algorithm is a highly efficient adaptive decomposition algorithm that is very suitable for financial time series.It can decompose the wave characteristics of the sequences at different scales,so as to improve the accuracy of model prediction.Secondly,the memory structure of LSTM makes the model more accurate in the prediction of memorized financial time series.Thirdly,the high frequency sequence,low frequency sequence and trend sequence obtained from the scale decomposition of the financial time series have certain economic significance,which respectively represent the short-term impact caused by normal fluctuations as well as irregular events,the long-term impact caused by major events and the long-term development trend.Last but not least,EEMD-LSTM model fully combines the advantages of EEMD decomposition algorithm and LSTM network in time series prediction,which can effectively improve the accuracy of financial time series prediction.