Decoding Market Signals with Self-adaptive Temporal Modeling in Multi-period Decision-Making

The volatility of agricultural commodity prices presents substantial challenges for resource allocation and strategic planning in multi-period decision-making within supply chains. However, the complex and non-stationary, non-linear nature of agricultural price dynamics renders these fluctuations difficult to model and predict effectively. To address this challenge, we introduce an AI-driven predictive framework that integrates attention mechanisms with random forests (AM-RF). Leveraging a decade-long observation period (from July 26, 2013, to July 25, 2023), we integrate approximately 21,000 news headlines with macroeconomic indicators, while innovatively incorporating the Chain-of-Thought (CoT) strategy to automatically extract sentiment signals from news content using large language models (LLMs). Our results demonstrate that our AM-RF model significantly outperforms benchmark models in multi-step forecasting across various time horizons, short-term (1 day, 5 days), mediu mterm (60 days), and long-term (180 days) with superior accuracy and robustness. Furthermore, our model captures differences in the time-series contribution of external environmental factors. Specifically, market sentiment emerges as the most influential factor in short- and medium-term price forecasting, while the economic environment assumes a more prominent role in long term predictions. In contrast, market fundamentals and price costs exhibit relatively marginal contributions, especially in longer forecasting windows. Our study highlights the potential of AI and LLM in addressing the complexities of dynamic market environments and provides robust evidence to support data driven decision-making frameworks within agricultural supply chains.