The inherent instability of agricultural commodity markets, particularly for Arabica coffee futures, poses significant challenges for risk management due to the sector's vulnerability to climate shocks and supply chain disruptions. Although traditional econometric models like GARCH successfully capture stylized facts such as volatility clustering, they often struggle to account for the abrupt structural breaks and non-linear dependencies introduced by exogenous events. To address these limitations, this study proposes a hybrid framework that bridges the methodological divide between econometric rigor and deep learning flexibility. Specifically, we integrate the structural parameters of Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and its asymmetric variant, GJR-GARCH, directly into Bidirectional Long Short-Term Memory (BiLSTM) networks. By employing a residual learning strategy, this approach leverages the theoretical interpretability of econometrics while utilizing the bidirectional processing capabilities of neural networks to capture complex temporal dependencies. Empirical analysis utilizing Intercontinental Exchange (ICE) data from 2000 to 2025 reveals that the hybrid GJR-BiLSTM architecture statistically outperforms both standalone econometric models and unidirectional LSTM baselines. The results demonstrate that incorporating asymmetric shock parameters allows the model to effectively capture the distinct leverage effect and long-memory characteristics observed in coffee markets. Furthermore, the study identifies a 21-day rolling window as the optimal horizon for maximizing forecast stability. Ultimately, these findings underscore the economic value of hybrid architectures, providing actionable insights for stakeholders seeking to optimize hedging strategies and mitigate financial exposure in global coffee supply chains.
Yihuan et al. (Wed,) studied this question.