| With the frequency of transaction data collection in the financial market has gradually increased,and the high-frequency data with non-linearity,non-stationarity and high-noise have attracted the attention of many researchers,and various studies have also focused on the short-term prediction of high-frequency data recently.Traditional time series models rely on linear regression to explain the relationship between variables and then forecast the financial series,but can’t discover the characteristics of high-frequency data well.The unsupervised learning process of neural network is a better way to analysis the high-frequency data with non-linearity.As the innovation of computer technology,the structure of neural network has been continuously deepened,and the deep neural network(DNN)with deeper network has been widely used.Empirical mode decomposition(EMD)is widely used in the analysis of nonlinear and non-stationary high-frequency data in time domain and frequency domain.Therefore,this paper uses two empirical mode decomposition methods in Hilbert-Huang transformation and combines deep neural network model to make short-term prediction of high-frequency data.In the first part of this paper,the DNN model is used to make short-term prediction for 5-minute data of Shenzhen Component Index.In order to obtain more accurate results,we introduced the empirical mode decomposition and ensemble empirical mode decomposition(EEMD),and built the EMD-DNN and EEMD-DNN models.The sample data were decomposed to obtain the intrinsic mode functions(IMFs)at different scales.According to fluctuation frequency,the IMFs were reconstructed.Finally,we trained the reconstructed data and predicted.In the second part,the high-frequency IMFs were de-noised based on the multi-resolution analysis method in the wavelet analysis.We replaced the high-frequency components in the IMFs with the de-noised one,and combined with the low-frequency components to construct a new sample,then predicted by applying DNN.The results show that the EEMD-DNN model with wavelet de-noising is more accurate than the EMD-DNN model with wavelet de-noising.An example shows that the EEMD-DNN model with wavelet de-noising is more accurate than the model without wavelet de-noising and the EMD-DNN model with wavelet de-noising in the high-frequency data forecasting. |
