A Study on Crude Oil Price Forecasting Model Integrating Ceemdan-Vmd Multiscale Decomposition with Cnn-Bilstm

Abstract

As a key global energy benchmark, WTI crude oil prices exhibit significant volatility driven by a complex interplay of economic, geopolitical, and financial factors. This results in highly nonlinear and non-stationary price behavior, posing substantial challenges to accurate forecasting. This study proposes a hybrid forecasting model (CEEMDAN-VMD-CNN-BiLSTM) to improve the accuracy of WTI crude oil price predictions, addressing its volatility driven by complex economic, geopolitical, and financial factors. The model integrates Complementary Ensemble Empirical Mode Decomposition (CEEMDAN), Variational Mode Decomposition (VMD), Convolutional Neural Networks (CNN), and Bidirectional Long Short-Term Memory (BiLSTM). Pearson correlation analysis selects the top four indicators from nine potential features, including LBMA gold price, Brent crude oil price, USD/CNY exchange rate, and Federal Funds rate. The raw price data and features undergo dual decomposition, with CNN extracting local temporal features and BiLSTM capturing bidirectional dependencies. The model, trained on data from 1997-2017 and validated on 2017-2025, achieves a MAPE of 3.66% and an R2 of 95.94%, outperforming BiLSTM and CNN-BiLSTM. Comparative analysis with eight other models, including LSTM, SVM, and Transformer, confirms the superior performance of the proposed model, with a MAPE of 5.57% and an R2 of 93.87%. This approach enhances forecasting accuracy and provides insights for optimizing crude oil futures markets and risk management strategies.

Keywords: WTI crude oil price, Pearson correlation analysis, CEEMDAN-VMD dual decomposition algorithm, CNN-LSTM algorithm, Ensemble learning