TEDM: Time Series Forecasting with Elucidated Diffusion Models
Published in International Conference on Learning Representations (ICLR 2026), 2026
TEDM is an autoregressive diffusion framework for multivariate time-series forecasting. The central idea is to collapse diffusion time and physical time into the same axis, reducing sampling complexity from O(SH) to O(H) for horizon H and S diffusion steps.
On standard long-sequence benchmarks (H=96), TEDM reports the best diffusion-model scores on ETTh2, ETTm2, and Exchange, with competitive results on ETTm1 and Weather. I especially like this paper because it turns a practical deployment bottleneck (slow iterative sampling) into a principled modeling choice while keeping strong forecast quality.