A new hybrid model for multi-step WTI futures price forecasting based on self-attention mechanism and spatial–temporal graph neural network

Abstract

WTI futures prices are impacted by supply, demand and a variety of financial factors, including U.S. dollar exchange rates, interest rates, market sentiment and related market linkages. The frequent changes in these factors cause WTI futures prices to fluctuate dramatically and complicate the trading decisions of investors and the policy-making of governments; consequently, accurate forecasting of WTI futures prices has become a topic of intense interest in the field of energy research. To thoroughly investigate the impact of various factors on crude oil prices, this paper introduces the self-attention mechanism and the spatial–temporal graph neural network Graph WaveNet (GWNet) to predict crude oil prices. The self-attention mechanism is employed to learn time-varying interactions between variables to tackle a problem where the graph structure is unknown. The graph convolution and the dilated causal convolution in GWNet capture the spatial and temporal dependencies, respectively. The empirical findings demonstrate that the proposed Graph WaveNet with Self-Attention (GWNet-Attn) robustly and significantly outperforms all baseline models in various prediction horizons and that the dollar index (USDX), LIBOR, and VIX have surpassed supply and demand as the most influential predictors of WTI futures prices.