Causal Time Series Generation via Diffusion Models
Source
- Raw Markdown: paper_catsg-2025.md
- PDF: paper_catsg-2025.pdf
- Preprint: arXiv 2509.20846
- Official code: microsoft/TimeCraft/CaTSG
Status And Credibility
CaTSG was first posted to arXiv on 2025-09-25 and the inspected arXiv metadata reports version 3, updated 2026-05-14. It is part of the official Microsoft TimeCraft repository. No peer-reviewed venue is listed in the arXiv metadata snapshot.
Core Claim
CaTSG reframes conditional time-series generation through Pearl’s causal ladder: observational, interventional, and counterfactual generation. It instantiates this with a diffusion framework that uses backdoor-adjusted guidance, latent environment embeddings, and abduction-action-prediction for counterfactual sampling.
Key Contributions
- Defines causal time-series generation as a task family beyond observational conditional generation.
- Derives causal score functions through backdoor adjustment.
- Introduces a latent environment bank and EnvInfer module to model unobserved confounding.
- Evaluates synthetic harmonic-oscillator datasets with ground-truth counterfactuals and real Air Quality and Traffic datasets.
Evidence And Results
The paper reports better observational fidelity and competitive interventional behavior across four datasets. It can quantitatively evaluate counterfactuals on synthetic datasets because those datasets expose ground-truth counterfactuals. On real datasets, counterfactual evidence is mostly case-study and plausibility based.
Limitations
The paper explicitly notes that conditions are restricted to time-series inputs, the SCM structure is predefined, and real-world counterfactual evaluation remains difficult because ground-truth counterfactuals are usually unavailable. CaTSG is best interpreted as an initial proof-of-concept for intervention-aware generation, not as settled evidence that diffusion generators can recover real causal mechanisms from observational data.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Causal structure, counterfactuals, and control | partially closes | Defines observational, interventional, and counterfactual time-series generation and implements backdoor-adjusted diffusion guidance. | Predefined SCMs, synthetic counterfactual labels, and limited real-world validation. |
| Time-series generation and editing | partially closes | Generates samples under causal conditions and latent environments. | No general editing interface and no broad multimodal conditions. |
| Benchmark hygiene | warning | Real-world counterfactual evaluation lacks ground truth. | Needs intervention datasets with observed outcomes or carefully designed simulators. |
Links Into The Wiki
- CaTSG
- TimeCraft
- Time-Series Generation
- Causal Time Series
- Time-Series Benchmark Hygiene
- Foundation Time-Series Model Research Agenda
Open Questions
- Can CaTSG work when the causal graph is partially unknown or changes by domain?
- Which real-world benchmarks, such as Causal Chambers or the Action-Conditioned Time-Series Datasets tier map, can validate intervention and counterfactual generation without relying only on synthetic labels?