Multimodal Demand Forecasting in Supply Chains: Integrating Large Language Models with ERP Data for Enhanced Decision Support

Accurate demand forecasting remains a critical challenge in modern supply chain management, particularly in volatile and data-intensive U.S. markets where traditional ERP-based models struggle to incorporate external contextual influences. This study proposes a multimodal demand forecasting framework that integrates structured enterprise resource planning data with unstructured textual context using large language model embeddings. Open-source retail sales data representative of U.S. ERP environments is combined with time-aligned external textual information to capture semantic signals influencing demand behavior. A hybrid deep learning architecture is developed to fuse time-series sales features with LLM-derived contextual representations through an attention-based mechanism. Experimental results demonstrate that the proposed approach significantly outperforms traditional statistical models and ERP-only deep learning baselines, achieving a reduction of over 35% in mean absolute percentage error and exhibiting superior robustness during periods of demand volatility. The findings confirm that LLM-enhanced multimodal forecasting provides substantial accuracy gains and operational value, offering a scalable and practical solution for U.S. supply chain decision-making and demand planning.