A Financial Time Series Forecasting Method Based On Multiple Decomposition

Finance,which is closely related to personal life and social economic development,can allocate funds crossing time and region.Predicting the development of the financial industry accurately can not only maintain and increase the value of personal assets,but also promote the smooth operation of social economy.On the basis of historical data,financial time series prediction makes effective reasoning for its future development trend.However,with the influence of various internal and external factors,financial time series has the character of nonlinear and nonstationary,which makes it difficult to predict accurately.It is of great practical and theoretical significance to build an effective prediction model for financial time series forecasting.Numerous scholars have taken efforts to predict financial time series.The extant main prediction models include traditional statistical models and recent artificial intelligence models.These models contribute to the development of financial time series prediction.However,due to the complexity of financial time series,these models are difficult to adapt to the current needs for accurate prediction.Therefore,new data processing methods and prediction models need to be further developed.In this thesis,based on the “decomposition and integration” framework,a new prediction model is proposed.This model,which integrates the multiple improved complete ensemble empirical mode decompositions with adaptive noise(MICEEMDAN),sine cosine algorithm(SCA)and random vector functional link(RVFL)neural networks,is called MICEEMDAN-SCA-RVFL.Firstly,multiple ICEEMDANs with random parameters are employed to separate the original time series into groups of subseries.Secondly,RVFL networks are selected to individually forecast each subseries and get a series of prediction values.During this process,in order to improve the prediction performance and search efficiency,SCA optimization is chose to optimize the parameters of RVFL networks.Thirdly,prediction values of subseries are summed to produce a group of prediction results of single decomposition.Finally,this model integrates these prediction results of all single decomposition with respective weights and obtains final prediction result.In order to prove effectiveness of the proposed model,four financial time series datasets are selected for empirical research,including Shanghai Stock Exchange Composite Index(SSEC),Shenzhen Stock Exchange Composite Index(SCI),West Texas Intermediate(WTI)crude oil prices and USD/RMB exchange rate(USD/CNY).During the experiments,as benchmark models,three single prediction models and their respective integrated models are chosen to compare with the proposed hybrid prediction model.Experiment results show that the proposed MICEEMDAN-SCA-RVFL model based on multiple decomposition can significantly improve the prediction accuracy of the financial time series.On this basis,this thesis further verifies the effectiveness of the proposed multiple decomposition model.In contrast to single decomposition,the prediction results show that the proposed MICEEMDAN-SCA-RVFL model can get better prediction results and is more suitable to forecast financial time series.In addition,the SCA algorithm can effectively help RVFL networks to search appropriate parameters and further improve the prediction performance.Moreover,a proper integration ratio should be selected according to different time series.In summary,our proposed model is able to effectively improve the forecasting performance of financial time series,which can not only better help investors allocate their assets,but also provide decision support for the related institutions or organizations.