Integrating Reconstructive Decomposition And Deep Learning Methods For Crude Oil Price Chaotic System Forecasting

Based on the existing literature,the crude oil market is a complex nonlinear chaotic system,and the crude oil price is easily affected by various factors such as military affairs,politics,economy and diplomacy,which makes the crude oil price change have the characteristics of randomness,mutation and chaos.As the country with the largest crude oil import volume in the world,China’s continuously changing international crude oil price has a huge impact on China’s economic development.Based on this,it is of great significance to accurately grasp the development trend of crude oil price.Chaotic economic time series exists in all aspects of social life.Since chaotic data can not be directly modeled with nonlinear technology,and at the same time,according to the law of initial sensitivity,chaotic systems are unpredictable in the long run,so it should be analyzed first in the study of predicting crude oil price series.Whether the nonlinear system is chaotic.Based on the chaotic characteristics of crude oil price series,this thesis proposes a PSRVMD-CNN-BILSTM model combining phase space reconstruction technique(PSR),variational modal decomposition algorithm(VMD),convolutional neural network(CNN),and bidirectional long and short-term memory network(BILSTM)to predict crude oil prices.First,the crude oil price sequence is reconstructed using PSR,and the two parameters time delay(τ)and embedding dimension(m) of the phase space are confirmed by the autocorrelation coefficient method and the false nearest neighbor point method,and the maximum Lyapunov exponent is calculated to be 0.0525,which verifies the chaos of the sequence.Next,use the VMD algorithm to perform high-frequency denoising processing on the data.Finally,the reconstructed and denoised phase space is fed into the CNN-BILSTM model for multi-step prediction of one day,a week,half a month,and 19 days,respectively.And compared with other baseline models,non-reconstructed models,non-denoised models and processed models.This thesis uses Brent crude oil price data for 5638 days from January 4,2002 to December31,2021 for empirical analysis,and uses MSE,RMSE,MAE and MAPE as evaluation indicators to evaluate the model.The empirical results show that the proposed PSR-VMDCNN-BILSTM model has significantly improved prediction accuracy compared with other control models,and has a better prediction effect at each time step.At the same time,the results of the DM test fully illustrate the model superiority.This model not only provides a new idea for the field of crude oil price prediction,but also can be extended to other economic time series with chaotic characteristics.