DP-CDM: A Dual-Phase Conditional Diffusion Model for Demand Forecasting in Digital Supply Chains

Accurate demand forecasting is vital for digital supply chains, enabling efficient inventory, planning, and logistics. Existing models suffer from two key limitations: (i) they fail to adequately model the influence of external conditional variables, and exhibit limited ability to capture complex multi-modal distributions inherent in real-world demand data; and (ii) conditional information is concatenated with historical inputs only once at the model entrance, which often leads to information fading during deep propagation and reduces sensitivity to event-driven demand shocks. To address these challenges, we propose DP-CDM, a dual-phase conditional diffusion model for demand forecasting. First phase, a reverse sliding diffusion is applied along the temporal axis, which exploits temporal continuity to construct an autoregressive learning mechanism, thereby strengthening sequence modeling and avoiding structural misalignment. Second phase, a noise-degradation diffusion enriches multimodal probabilistic representations while improving robustness against exogenous disturbances. Moreover, we design a conditional embedding module with multi-modal feature alignment, which aggregates local historical windows, global trends, and SHAP-quantified external factors into multimodal embeddings. They are injected consistently throughout the dual denoising process to guide the final forecasts. Extensive experiments demonstrate DP-CDM reduces MAPE by 1.5 percentage points and improves R² by 4.4%, highlighting it effectiveness in capturing event-driven dynamics.