Financial Time Series Prediction Based On Multiple Empirical Mode Decomposition Generative Adversarial Networks

Financial time series prediction is a time series analysis technology that makes reasonable speculations about the future data development based on the historical laws and changing trends of financial data.It has important significance for government departments,investment institutions,and investors.Due to the characteristics of non-linearity,non-stability,and many input features of financial time series,the establishment of a high-precision financial time series prediction model has always been a research hotspot in the financial and computer fields.This thesis presents three financial time series prediction models based on generative adversarial networks,and uses the CSI 300 Index from January 2002 to March 2020 to verify their forecasting performance.The main work of this article is as follows:1.This thesis presents a financial time series prediction model based on Empirical Mode Decomposition Wasserstein Generative Adversarial Networks(EMD-WGAN).In view of the non-linear and unstable characteristics of financial time series,EMD and the generator of WGAN are combined as the generator of the model;Aiming at the unstable characteristics of WGAN model,the loss function of WGAN-GP generator is combined with the mean square error of real data and generated data as the objective function of this model generator.The prediction performance of the model is verified by the CSI 300 Index,and its mean square error is 0.0019.The experimental result shows that the model can better analyze and process non-linear and non-stationary financial time series and achieve good prediction results.2.This thesis presents a financial time series prediction model based on Dual-Stage Attention-Based Wasserstein Generative Adversarial Networks(DAWGAN).In view of the problem that there are many input features of financial time series and it is difficult to adaptively select,an input attention mechanism is introduced into the generator to adaptively select input features;Aiming at the problem of long-term dependence of financial time series that is difficult to capture,a time attention mechanism is introduced into the generator to capture the long-term dependence of financial time series.The prediction performance of the model is verified by the CSI 300 Index,and its mean square error is 0.0014.Experimental results show that the model can adaptively select input features and capture the long-term dependence of financial time series and reduce the prediction error of the model.3.The above two financial time series prediction models based on WGAN are fused to obtain Multivariate Empirical Mode Decomposition Wasserstein Generative Adversarial Networks(MEMD-WGAN).The model is improved on the basis of EMD-WGAN,using Multivariate Empirical Mode Decomposition(MEMD)instead of EMD.Besides,it introduces input attention mechanism and time attention mechanism in the generator,and its discriminator adopts CNN.Experimental results show that the prediction model has higher prediction accuracy than LSTM,WGAN-gp,EMD-WGAN and DA-WGAN.The prediction performance of the model is verified by the CSI 300 Index,and its mean square error is 0.0004.Experimental results show that the model can not only analyze and process the non-linear and non-stationary financial time series well,but also adaptively select input features and capture the long-term dependence of financial time series,which improves the prediction accuracy of the model.