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

A research team led by Jintao Zhang has introduced StaTS (Spectral Trajectory Schedule Learning), a novel diffusion model architecture that addresses fundamental limitations in probabilistic time series forecasting. Traditional diffusion models use fixed noise schedules that often produce intermediate states that are hard to invert and terminal states that deviate from near-noise assumptions, while existing methods rely primarily on time domain conditioning without adequately modeling schedule-induced spectral degradation. StaTS overcomes these challenges through a dual-component system: the Spectral Trajectory Scheduler learns data-adaptive noise schedules with spectral regularization to improve structural preservation and stepwise invertibility, while the Frequency Guided Denoiser estimates schedule-induced spectral distortion and uses it to modulate denoising strength for heterogeneous restoration across diffusion steps and variables.

The technical innovation lies in StaTS's two-stage training procedure that stabilizes the coupling between schedule learning and denoiser optimization, allowing the model to dynamically adjust its noise schedule based on the specific characteristics of the time series data. This adaptive approach enables the model to maintain strong forecasting performance while using significantly fewer sampling steps—a critical advantage for real-world applications where computational efficiency matters. The researchers validated StaTS on multiple real-world benchmarks, demonstrating consistent performance gains over existing methods while maintaining robustness across different forecasting scenarios. The model's ability to preserve structural information across noise levels represents a significant advancement in diffusion-based time series forecasting, potentially enabling more accurate predictions in domains like finance, energy demand forecasting, and industrial process monitoring.