From Time-Series Prediction to System Modeling: A Dual-Attention Framework for Multi-Source Interaction in Soybean Futures Markets

Futures price forecasting is often treated as a time-series prediction task. However, agricultural futures markets function as complex information systems in which prices emerge from the interaction of heterogeneous sources, including trading behavior and news-driven sentiment. Ignoring such cross-domain interactions limits the ability of traditional models to capture systemic price dynamics. This study reconceptualizes soybean futures forecasting as a multi-source information interaction problem and proposes a dual-attention LSTM framework to model cross-system coupling effects. A RoBERTa-based sentiment classifier is first developed to quantify market sentiment from news headlines. The extracted sentiment features are then integrated with historical trading variables and fed into an LSTM network equipped with temporal and feature-level attention mechanisms to capture dynamic evolution patterns and heterogeneous factor interactions. Empirical results show that the proposed system consistently outperforms conventional models. With a sliding window of 30 and a forecasting horizon of 7 days, the R2 improves from 0.922 to 0.9797, demonstrating enhanced capability in modeling medium-term price dynamics. The findings highlight that futures forecasting should be approached as a system-level information integration task rather than a purely statistical extrapolation problem.