An interpretable SCA-ANN framework for WTI crude oil price forecasting

Abstract The recent conflict in the Strait of Hormuz as a critical oil transmit checkpoint combined with the market's inherent instability and the complex, nonlinear interactions among key influencing factors, makes accurate crude oil price forecasting a formidable challenge. Conventional statistical methods often struggle to capture these dynamics, underscoring the need for innovative, data-driven approaches. This study proposes a robust predictive model combining artificial neural networks (ANNs) enhanced through optimization with Sine–Cosine Algorithm (SCA), which improves the ANN's ability to identify optimal weight configurations by employing sine and cosine functions to guide the solution search process. Three model variations are explored based on the inclusion of temporal features: complete temporal data, a single time index, and the exclusion of all temporal inputs. Quantitative evaluation across training and testing datasets revealed that Approach 2 (time index) delivered the best overall predictive performance, achieving the lowest RMSE = 2.532, lowest MAE = 2.025, and highest R2 = 0.984 on the test set, outperforming Naïve Persistence, standard ANN, LSTM, and the two alternative SCA-ANN temporal-encoding. SHAP interpretability analysis applied to Approach 2 revealed that the Time Index is the single most influential feature, followed by the 21-day moving average, 14-day moving average, and 7-day moving average, providing transparent and physically consistent explanations of model behavior.