Intelligent dual-stage attention-based deep networks for energy market predictions

Abstract

The modeling and prediction of oil prices are very important tasks. However, the predictions of oil price by traditional models are not very effective. The challenge comes from the nonlinear and non-stationary dynamics of oil prices, and they are also heavily correlated with global economic condition and financial fluctuations. This study employs a dual-stage attention-based recurrent neural network (DA-RNN) for oil price forecasting. The DA-RNN architecture includes both an encoder and a decoder. The encoder features an input attention mechanism designed to adaptively identify and choose significant and pertinent driving series. The decoder incorporates a temporal attention mechanism to obtain long-term dependencies of the encoded inputs. The dual-stage attention mechanism of DA-RNN enables both input selection and temporal focus, allowing the model to adaptively choose important and relevant driving series while capturing long-term temporal dependencies. Empirical results indicate that DA-RNN achieved lowest prediction errors, for example, RMSE values of approximately 2.2 for WTI, 2.4 for Dubai, and 2.3 for Brent crude oil prices, which reduces about 30 % error compared to other models. These findings clearly demonstrated that the DA-RNN model outperforms traditional econometric methods and machine learning models, highlighting its potential as a powerful tool for energy market predictions.