Diffusion Models: Early Stopping Improves Time Series Forecasts

Diffusion models are increasingly used for time series forecasting, particularly for their ability to model uncertainty. However, this research challenges the common assumption that the iterative sampling process of these models always leads to better forecasts. The study found that while general temporal structures are recovered early in the denoising process, continued refinement at very low noise levels can introduce statistical drift, ultimately harming the final forecast quality. This observation helps explain why previous diffusion model designs for time series often favored narrower architectures and specific schedules. Building on this insight, the researchers propose a novel, label-free global stopping criterion. This criterion can automatically detect the optimal point to terminate the reverse diffusion process, preventing the degradation caused by excessive denoising. By stopping early, the method not only improves predictive accuracy but also significantly speeds up inference. Furthermore, to ensure effective training when early stopping is applied, a Bernoulli timestep sampler is introduced. This sampler concentrates training on the high-noise regions where early stopping occurs, while still covering the full diffusion process. Extensive experiments across various real-world datasets confirm the superior performance of this new approach.