Forecasting global agricultural trade imbalances using a hybrid deep learning and gradient boosting framework

Abstract Global agricultural trade imbalances, influenced by policy asymmetries, climate variability, and supply chain disruptions, have become a major challenge to economic stability and food security. Although previous studies have examined agricultural trade dynamics, relatively little attention has been devoted to forecasting future agricultural trade imbalances using data-driven approaches that can simultaneously capturing local patterns, temporal dependencies, and nonlinear relationships. To address this gap, this study develops a hybrid CNN-BiLSTM-Attention-XGBoost framework that integrates convolutional neural networks for feature extraction, bidirectional long short-term memory networks for temporal modeling, a self-attention mechanism for adaptive feature weighting, and XGBoost for nonlinear prediction. The framework is evaluated using agricultural trade data from 20 countries covering the period 2000–2023. Experimental results demonstrate strong forecasting performance, achieving an MAE of 1.36, RMSE of 1.61, R² of 0.99, and MAPE of 1.28%, while consistently outperforming benchmark models. Ablation and robustness analyses further confirm the complementary contributions of the integrated components and the framework’s stability under various perturbation settings. The results suggest that the proposed framework provides an effective tool for forecasting agricultural trade imbalances and may support trade monitoring, risk assessment, and evidence-based policy analysis in complex global agricultural markets.