StaTS: Spectral Trajectory Schedule Learning for Adaptive Time Series Forecasting with Frequency Guided Denoiser

Researchers introduce StaTS, a new diffusion model for time series forecasting that learns adaptive noise schedules and uses frequency-guided denoising. The model addresses limitations of fixed noise schedules in existing diffusion models by incorporating spectral regularization and data-adaptive scheduling for improved structural preservation.

• →StaTS introduces adaptive noise scheduling for diffusion models in time series forecasting, moving beyond fixed schedules.
• →The Spectral Trajectory Scheduler (STS) uses spectral regularization to improve structural preservation and invertibility.
• →Frequency Guided Denoiser (FGD) estimates spectral distortion to modulate denoising strength across different steps.
• →Experiments show consistent performance gains on real-world benchmarks while maintaining efficiency with fewer sampling steps.
• →The two-stage training procedure stabilizes the coupling between schedule learning and denoiser optimization.