AgriVolT: A Multi-Modal Temporal Vision Transformer for Climate-Informed Commodity Price Forecasting (Papers Track)
Sharanya Roy (Algoverse); Krisha Agarwal (Algoverse); Sahir Gupta (Algoverse); Anshul Patil (Algoverse); Ahan M R (Birla Institute of Technology and Science, Pilani, Algoverse)
Abstract
Climate extremes increasingly disrupt agriculture, driving volatility in staple food markets and threatening global food security. We present AgriVolT (Agriculture Volatility Transformer), a multi-modal temporal vision transformer that integrates climate reanalysis, satellite imagery, production data, and historical market prices to forecast commodity price fluctuations. AgriVolT employs temporal encodings for economic sequences, a price-focused prediction head, and cross-modal attention to capture non-linear interactions between climate shocks and market dynamics. In validation experiments on U.S. corn, soybean, and wheat markets, AgriVolT reduced mean absolute error (MAE) by 30–50% and mean absolute percentage error (MAPE) by up to 60% compared to AgriFM, a recent model we used to benchmark AgriVolT with. For example, it achieved a 37.1% MAPE for corn (vs. 74.8%) and 38.2% for wheat (vs. 98.7%). These results highlight AgriVolT’s potential to provide accurate, interpretable early warnings that help policymakers, humanitarian organizations, and farmers mitigate food insecurity in climate-volatile markets.