A Commodity Demand Forecasting System Based on Dual-Phase Conditional Diffusion Model

Abstract

Accurate demand forecasting enables efficient supply chain management. However, two major challenges persist: (a) insufficient modeling of external conditional variables and poor capture of complex multimodal demand distributions and (b) conditional information that is fused only once at the system input, causing information decay and reduced responsiveness to event-driven shocks. To this end, the authors propose a dual-phase conditional diffusion model (DP-CDM) in which a reverse sliding diffusion along the temporal axis exploits temporal continuity to build an autoregressive mechanism, enhancing sequence modeling and avoiding structural misalignment. In addition, noise-degradation diffusion enriches multimodal probabilistic representations and improves robustness to external disturbances. A conditional embedding module aligns multimodal features by aggregating local histories, global trends, and SHapley Additive exPlanations (SHAP)-quantified external factors, which are injected throughout the denoising processes. A 3.7% improvement in fitting performance showed the effectiveness of this model in capturing event-driven demand dynamics.