MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process
Source
- Raw Markdown: paper_mg-tsd-2024.md
- PDF: paper_mg-tsd-2024.pdf
- Preprint: arXiv 2403.05751
- Official code: Hundredl/MG-TSD
Status And Credibility
MG-TSD was posted to arXiv on 2024-03-09, updated on 2024-03-16, and the arXiv metadata comment states International Conference on Learning Representations (ICLR) 2024. It is older than one year as of 2026-06-15, but it is a peer-reviewed ICLR source and is included in TimeCraft as a general time-series diffusion technique.
Core Claim
MG-TSD uses coarse-grained versions of time series as intermediate targets in diffusion forecasting. The intuition is that diffusion denoising and temporal smoothing both move from coarse structure toward fine details, so coarse granularity can guide the sampling path.
Key Contributions
- Defines a multi-granularity guided diffusion loss.
- Uses coarse-grained time-series targets at intermediate diffusion steps.
- Applies the method to probabilistic time-series forecasting without external data.
- Evaluates on Solar, Electricity, Traffic, Taxi, KDD-cup, and Wikipedia.
Evidence And Results
The paper reports improvements over probabilistic forecasting baselines on CRPS, NMAE, and NRMSE. Ablations and case studies support the claim that intermediate coarse targets stabilize diffusion forecasting.
Limitations
- MG-TSD is probabilistic forecasting from numeric history, not synthetic dataset generation from arbitrary conditions.
- Granularity selection uses heuristics and dataset-specific settings.
- It does not model actions, control inputs, interventions, or counterfactuals.
- Forecasting benchmark performance should not be merged with text-controlled or target-aware generation metrics.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Time-series generation and forecasting distributions | adjacent | Diffusion generates probabilistic forecast sample paths guided by coarse temporal structure. | Not a general synthetic generator or world model. |
| Dense numeric fidelity | partially closes | Coarse-to-fine guidance explicitly targets structure preservation during denoising. | Needs tests for rare events, calibration, and downstream utility. |
| Control and counterfactuals | insufficient evidence | No action or intervention channel. | Needs controllable inputs and outcome rollouts. |