- Updated: March 11, 2026
- 7 min read
StaTS: Spectral Trajectory Schedule Learning for Adaptive Time Series Forecasting with Frequency Guided Denoiser
Direct Answer
StaTS introduces a diffusion‑based time‑series forecasting framework that learns its own noise schedule and denoising strategy through a Spectral Trajectory Scheduler (STS) and a Frequency Guided Denoiser (FGD). By adapting the diffusion process to the spectral characteristics of the data, StaTS delivers more accurate probabilistic forecasts while requiring fewer sampling steps, making it both more reliable and computationally efficient.
Background: Why This Problem Is Hard
Probabilistic forecasting of multivariate time series is a cornerstone of many enterprise applications—demand planning, energy load balancing, financial risk assessment, and autonomous system monitoring. Traditional autoregressive or transformer‑based models excel at point predictions but often struggle to capture uncertainty, especially when the underlying signal is noisy or exhibits complex seasonal patterns.
Diffusion models have emerged as a promising alternative because they treat forecasting as a generative denoising problem, naturally producing full predictive distributions. However, two practical bottlenecks have limited their adoption:
- Fixed noise schedules. Most diffusion forecasters pre‑define a linear or cosine schedule that adds Gaussian noise in a predetermined way. This rigidity creates intermediate latent states that are difficult to invert, leading to artifacts and a final “terminal” state that deviates from the theoretical near‑noise assumption.
- Lack of spectral awareness. Existing approaches condition only on the time domain, ignoring how noise injection distorts the frequency spectrum of the series. Without modeling this degradation, the denoiser cannot reliably restore high‑frequency components (e.g., sudden spikes) that are crucial for accurate forecasting.
These issues manifest as slower convergence, higher sampling costs, and degraded performance on real‑world datasets where spectral content varies across variables and time horizons. A solution that adapts the diffusion trajectory to the data’s spectral structure is therefore a pressing research need.
What the Researchers Propose
The authors present StaTS (Spectral Trajectory Schedule Learning), a diffusion framework that jointly learns a data‑adaptive noise schedule and a denoiser that is explicitly guided by frequency‑domain information. The architecture consists of two tightly coupled components:
- Spectral Trajectory Scheduler (STS). Instead of a hand‑crafted schedule, STS predicts a per‑step noise magnitude that respects the spectral composition of the input series. A spectral regularizer penalizes schedule choices that would excessively attenuate important frequency bands, encouraging stepwise invertibility.
- Frequency Guided Denoiser (FGD). FGD estimates the spectral distortion introduced by the current schedule and modulates its denoising strength accordingly. This results in heterogeneous restoration—stronger correction for high‑frequency components that are more vulnerable to noise, and gentler treatment for low‑frequency trends.
Training proceeds in two stages. First, the STS is optimized while keeping the denoiser fixed, allowing the schedule to settle into a spectrally coherent trajectory. In the second stage, the FGD is fine‑tuned to the learned schedule, stabilizing the coupling between schedule and denoiser. This alternating scheme prevents the classic “chasing” problem where both modules continuously adapt to each other without converging.
How It Works in Practice
The practical workflow of StaTS can be broken down into three phases: preprocessing, diffusion‑forward, and diffusion‑reverse.
1. Preprocessing
- Raw multivariate series are normalized and optionally transformed into a frequency representation (e.g., via short‑time Fourier transform) to expose spectral features.
- The normalized series serve as the conditioning input for both STS and FGD.
2. Diffusion‑Forward (Noise Injection)
- At each diffusion step t, STS predicts a noise level σₜ that is tailored to the current spectral profile.
- Gaussian noise with variance σₜ² is added to the series, producing a latent state xₜ. Because σₜ respects spectral regularization, the resulting latent retains a recognizable structure rather than becoming a uniform blur.
3. Diffusion‑Reverse (Denoising)
- FGD receives the noisy latent xₜ and the schedule‑induced spectral distortion estimate.
- It applies a frequency‑aware denoising filter that varies across dimensions and time steps, effectively “undoing” the specific spectral damage caused by σₜ.
- The process iterates backward from the final step T to 0, yielding a forecast distribution that reflects both learned uncertainty and preserved spectral fidelity.
What sets StaTS apart is the closed‑loop feedback between schedule and denoiser: the schedule is not static, and the denoiser is not a generic neural net but a spectral‑sensitive operator that adapts its strength based on the schedule’s decisions. This synergy reduces the number of reverse diffusion steps needed to achieve high‑quality forecasts, cutting inference time by up to 40 % in the authors’ experiments.
Evaluation & Results
The authors benchmarked StaTS on several widely used time‑series datasets, including electricity consumption (Electricity), traffic flow (Traffic), and financial market indices (Finance). Each dataset features distinct seasonalities, noise levels, and dimensionalities, providing a rigorous testbed for adaptive diffusion.
Experimental Setup
- Baselines: traditional diffusion forecasters with fixed schedules (e.g., DiffWave‑TS), transformer‑based probabilistic models (e.g., ProbTransformer), and classic statistical methods (e.g., Prophet).
- Metrics: Continuous Ranked Probability Score (CRPS) for probabilistic accuracy, Mean Absolute Scaled Error (MASE) for point‑forecast quality, and sampling latency measured in milliseconds per forecast.
- Sampling steps: StaTS was evaluated with 20, 30, and 50 reverse steps to demonstrate efficiency gains.
Key Findings
- Improved probabilistic accuracy. Across all benchmarks, StaTS reduced CRPS by 7‑12 % relative to the strongest diffusion baseline, indicating tighter predictive intervals.
- Better point forecasts. MASE improvements of 4‑9 % were observed, showing that the spectral guidance also benefits deterministic predictions.
- Fewer sampling steps. With only 30 reverse steps, StaTS matched or outperformed baselines that required 50‑100 steps, translating to a 35‑45 % reduction in inference latency.
- Robustness to heterogeneous variables. In multivariate settings where some series exhibit high‑frequency spikes while others are smooth, the Frequency Guided Denoiser allocated denoising resources appropriately, preserving spikes without over‑smoothing the smoother series.
These results collectively demonstrate that learning a spectral‑aware schedule is not a theoretical nicety—it yields concrete gains in both forecast quality and computational efficiency, which are critical for production‑grade AI systems.
Why This Matters for AI Systems and Agents
For practitioners building AI‑driven agents that rely on accurate future predictions—such as supply‑chain optimizers, autonomous fleet managers, or financial trading bots—the benefits of StaTS are threefold:
- Higher decision confidence. Tighter predictive intervals reduce the risk of over‑ or under‑reacting to forecast uncertainty, leading to more stable control loops.
- Scalable inference. The ability to achieve strong performance with fewer diffusion steps lowers GPU memory footprints and latency, enabling real‑time forecasting in edge or cloud‑native environments.
- Adaptability to diverse data streams. By automatically tailoring the noise schedule to each variable’s spectral profile, StaTS can be deployed across heterogeneous sensor networks without hand‑tuning per‑domain hyperparameters.
These advantages align with the needs of modern AI orchestration platforms that must balance accuracy, speed, and resource utilization. Teams looking to integrate probabilistic forecasting into their agent pipelines can leverage StaTS as a drop‑in replacement for existing diffusion models, gaining immediate performance lifts.
For more on integrating advanced forecasting models into production pipelines, see our AI platform overview.
What Comes Next
While StaTS marks a significant step forward, several open challenges remain:
- Extending to irregular time grids. Current experiments assume regularly spaced timestamps. Adapting the spectral scheduler to handle missing or unevenly spaced observations would broaden applicability.
- Joint learning with exogenous inputs. Many forecasting scenarios incorporate external covariates (e.g., weather, macro‑economic indicators). Integrating these signals into the spectral regularization framework is an open research direction.
- Theoretical guarantees. Formal analysis of the invertibility properties induced by the learned schedule could provide stronger convergence assurances.
- Hardware‑aware scheduling. Future work could co‑optimize the schedule for specific accelerator constraints, further reducing latency on edge devices.
From an industry perspective, these avenues open the door to domain‑specific adaptations—think predictive maintenance for industrial IoT, where sensor data is irregular and heavily noisy, or real‑time demand forecasting for e‑commerce platforms that must ingest promotional calendars as exogenous features.
Developers interested in experimenting with StaTS can start with the open‑source repository and explore custom extensions. Our solution hub provides templates for deploying diffusion‑based forecasters in containerized environments, complete with monitoring and auto‑scaling hooks.
References
For a complete technical description, see the original arXiv paper. The authors have also released code at https://github.com/zjt-gpu/StaTS/ for reproducibility.