A Multi-Model Hybrid Framework for Forecasting International Crude Oil Benchmarks

Accurate forecasting of crude oil prices is essential for energy policy, investment decisions, and economic planning, yet remains challenging due to pronounced nonlinearity and volatility in oil markets. While hybrid models combining statistical and machine learning techniques are prevalent, the individual contribution of their components is seldom examined. This study presents a comprehensive comparison of standalone and hybrid machine learning models for forecasting monthly prices of three major crude oil benchmarks—Bonny Light, Brent, and West Texas Intermediate (WTI)—using monthly data from January 1998 to July 2025. The models considered include ARIMA, long short-term memory (LSTM) networks, Extreme Gradient Boosting (XGBoost), multilayer perceptrons (MLP), simple average hybrids, and multi-stage stacked hybrid architectures. Results confirm the dominance of XGBoost as the best standalone model. Simple averaging hybrids offer limited gains, whereas stacked hybrid models achieve consistent performance improvements. The proposed ARIMA–LSTM–XGBoost–MLP framework achieves the best performance, yielding RMSE values of 4.216 (Bonny), 3.881 (Brent), and 2.984 (WTI) with R² exceeding 0.95. Crucially, SHAP and sensitivity analysis deconstruct this optimal hybrid, revealing that over 90% of its predictive contribution originates from the XGBoost prediction, while ARIMA and LSTM inputs contribute minimally. This finding challenges the narrative that complex hybridization inherently benefits from model diversity. Instead, it shows that ensemble superiority can be predominantly attributed to a single, powerful algorithm. Our work emphasizes the importance of interpretability in hybrid modeling and offers practical guidance for efficient forecasting system design.