SFdiff : Diffusion Model with Self-Generation for Probabilistic Forecasting

ICLR 2026 Conference Submission22560 Authors

Keywords: Forecasting, Diffusion models

Abstract: Diffusion models have emerged as an effective approach for time-series probabilistic forecasting, aiming to generate future observations based on historical data through a denoising process. In this paper, we introduce self-generation technique designed to enhance the performance of conditional generation in time-series forecasting. Self-generation involves synthesizing not only future observations, but also historical data itself conditioned on the given historical context. While noise is often introduced during the observation process, our method can reduce the amount of noise in observed historical data, thereby enhancing forecasting accuracy. Additionally, to further boost forecasting performance, we incorporate classifier-free generation methods into conditional generation for time-series forecasting. In the experiment, we demonstrate that our method outperforms other condition generation methods.

Primary Area: learning on time series and dynamical systems

Submission Number: 22560