Forecasting crude oil futures prices using BiLSTM-Attention-CNN model with Wavelet transform

Abstract

In this study, a novel hybrid model for forecasting crude oil futures price time series is proposed. The combination of Bidirectional long short-term memory network (BiLSTM), Attention mechanism and Convolution neural network (CNN) is used. Based on the principle of deep learning (DL) and wavelet transform (WT), the time series of crude oil futures price are decomposed and reconstructed into a low-frequency main sequence and several high-frequency noise sequence. Then, the subsequences obtained by decomposition are predicted by BiLSTM-Attention-CNN model in turn. Finally, with several evaluating indicators (RMSE, MAPE, MAE and R2), the forecasting errors of proposed model and other comparative models are evaluated and compared. The experimental results based on two types of crude oil futures showed the novel model outperforms other related comparative models under different training sets lengths and the modification of Diebold–Mariano test (MDM) proved that the loss functions of constructed model are statistically significant in this paper. In general, the innovative combination forecasting model proposed in this paper is a promising technology for government agencies, investors and related enterprises.