Deep learning enhanced energy market prediction: A robust ARIMAX-LSTM fusion for crude oil pricing

Abstract

Crude oil is a highly strategic global resource, and price fluctuations significantly impact nearly all economic sectors. Therefore, accurate forecasting of its prices is essential for better financial stability and decision-making. This study aims to develop a robust model using monthly data from April 2004 to January 2024 to predict the price of crude oil. We propose a novel approach that blends ARIMAX and LSTM models using a weighted combination to leverage the strengths of econometric and machine learning methods. Unlike hybrid models, which are solely designed based on a decomposition-optimization structure, in our model, an explicit ensemble with weights via grid searching is used to enhance the model's flexibility and performance. As ARIMAX is more efficient in dealing with linear relationships and exogenous variables, LSTM performs much better and effectively captures nonlinear patterns and long-range dependence. Weight hyperparameter tuning and cross-validation help reduce the risk of overfitting or underfitting in the model. Our empirical results indicate that the LSTM model provides a powerful forecasting baseline. The weighted ensemble model offers a marginal improvement on the chronological test set, and the Diebold-Mariano test confirms this advantage is statistically significant. Cross-validation reveals the standalone LSTM to be highly robust, highlighting the importance of component model selection. This study contributes to a more sophisticated framework for risk assessment in energy policy by revealing the crucial trade-off between a model's period-specific accuracy and its general robustness.