An explainable deep learning framework for marine gas oil price forecasting: integrating LSTM with SHAP-based feature importance

Abstract Bunker fuel prices constitute a major component of maritime transport costs, representing 30–70% of total operating expenses and critically influencing competitiveness, planning, and regulatory compliance in global shipping. Their strong linkage to crude oil dynamics, coupled with additional volatility arising from supply–demand imbalances and regional market shocks, makes accurate forecasting both challenging and essential. This study proposes a novel hybrid framework that integrates explainable artificial intelligence through Shapley additive explanations–based feature engineering with a recurrent neural network long short-term memory (LSTM) model to forecast weekly Fujairah marine gas oil bunker prices 8 weeks ahead considering its non-linearity, structural breaks, and exogenous shocks. Based on both multivariate and univariate analyses, the proposed model LSTM–SHAP, consistently outperforms autoregressive integrated moving average (ARIMA), ARIMA with exogenous variables (ARIMAX), and conventional LSTM configurations. Empirical results demonstrate significant improvements in predictive accuracy, with the mean absolute percentage error decreasing from 11.78% for the ARIMAX(1,1,1) model to 5.52% for the proposed model, while the coefficient of determination increased from 57.77% to 92.51%. Moreover, the Diebold–Mariano and Harvey–Leybourne–Newbold tests confirm that these improvements are statistically significant across all model comparisons. The results underscore the robustness, interpretability, and operational relevance of the proposed framework for enhancing decision-making in volatile maritime fuel markets.