differencing-diffusion

This paper proposes DiffDiff, a diffusion framework for probabilistic time series forecasting that embeds predictability asymmetry into the diffusion trajectory, outperforming six diffusion baselines on seven benchmarks across four prediction horizons.