DiffDiff: Embedding Predictability Asymmetry in Diffusion Trajectories

DiffDiff is a diffusion framework for time series forecasting that incorporates the asymmetry between observed history and uncertain future components. By embedding this into the diffusion trajectory, the model better accounts for high-frequency dynamics versus slow-varying content.