Decoding agricultural supply chain trends with a news-driven temporal modelling framework

While the influence of news on agricultural prices is widely recognised, how news information shapes price expectations remains poorly understood. Using daily U.S. soybean futures prices and news texts from a specialised agricultural platform, we construct sentiment and topic factors and develop a Linear–Residual Attention Transformer (LRAT) model, which outperforms both statistical and deep-learning benchmarks across 1-, 5-, 10-, and 30-day horizons. The results reveal a temporal mechanism in news signals: short-term price movements are primarily driven by sentiment, medium-term dynamics are anchored by event-related content such as policy, supply, and trade, and the joint use of sentiment and topics yields the strongest gains at intermediate horizons. These patterns remain robust across trend prediction, alternative topic specifications, additional controls, and trading backtests. Interpreted through the lenses of news-driven expectation updating and narrative economics, our findings show that real-time agricultural news provides incremental information for price risk management and supports supply chain decision making.