MGDF: an embedded multi-graph deep learning method for crude oil price forecasting

Crude oil plays a critical role in the global energy system, and fluctuations in its price have far-reaching implications for economic stability and energy policy. This study develops a novel Multi-Graph Deep Forecasting Model (MGDF) framework to enhance the accuracy of crude oil price forecasting. The proposed model integrates thirteen influential variables across six dimensions: macroeconomic policy, market sentiment, geopolitical risk, supply and demand, cross-market influence, and economic activity as embedded features. A central innovation of MGDF is the construction of multi-layer graphs that capture both quantitative and semantic dependencies: (i) mutual information graphs characterize evolving linear and nonlinear interrelations among predictors, while (ii) LLM-based text graphs extract semantic linkages from unstructured news data using large language models. These graph embeddings are combined with Temporal Convolutional Networks (TCNs) to capture time-series patterns and integrated with a Long Short-Term Memory (LSTM) architecture for sequential forecasting. Empirical results demonstrate that MGDF consistently outperforms benchmark models across multiple evaluation metrics, including MSE, MAE, RMSE, and R-squared. Robustness is further confirmed through Model Confidence Set (MCS) and Diebold-Mariano (DM) tests, underscoring the model’s statistical reliability. The findings provide both a methodological contribution to the energy forecasting literature and practical insights for policymakers and market participants in mitigating risks associated with oil price volatility. JEL classification : C22; C53; Q43