OpenNovelty.org

Causal Time Series Generation via Diffusion Models

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.3Download Report PDF
Time Series GenerationConditional GenerationTime Series Analysis
Abstract:

Time series generation (TSG) synthesizes realistic sequences and has achieved remarkable success. Among TSG, conditional models generate sequences given observed covariates, however, such models learn observational correlations without considering unobserved confounding. In this work, we propose a causal perspective on conditional TSG and introduce causal time series generation as a new TSG task family, formalized within Pearl’s causal ladder, extending beyond observational generation to include interventional and counterfactual settings. To instantiate these tasks, we develop CaTSG, a unified diffusion-based framework with backdoor-adjusted guidance that causally steers sampling toward desired interventions and individual counterfactuals while preserving observational fidelity. Specifically, our method derives causal score functions via backdoor adjustment and the abduction–action–prediction procedure, thus enabling principled support for all three levels of TSG. Extensive experiments on both synthetic and real-world datasets show that CaTSG achieves superior fidelity and also supporting interventional and counterfactual generation that existing baselines cannot handle. Overall, we propose the causal TSG family and instantiate it with CaTSG, providing an initial proof-of-concept and opening a promising direction toward more reliable simulation under interventions and counterfactual generation.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
24
Contribution Candidate Papers Compared
4
Refutable Paper

Research Landscape Overview

Core task: causal time series generation with interventions and counterfactuals. The field organizes itself around several complementary branches. Causal Inference and Counterfactual Reasoning Frameworks provide foundational methods for identifying causal structures and reasoning about what-if scenarios in temporal data, often drawing on structural causal models and potential outcomes frameworks. Causal Generative Models for Time Series focus on learning data-generating processes that respect causal dependencies, employing architectures such as variational autoencoders, diffusion models, and flow-based approaches to produce synthetic sequences under hypothetical interventions. Causal Forecasting and Prediction emphasize forward-looking tasks where causal knowledge improves predictive accuracy or robustness, for instance by incorporating discovered causal graphs or leveraging large language models with temporal reasoning capabilities (e.g., Causal LLM Forecasting[5], Causal-TSF[6]). Synthetic Control and Counterfactual Estimation methods construct comparison units from observed data to estimate treatment effects in quasi-experimental settings, with classical techniques like Bayesian Structural Time Series[22] and newer refinements such as Conformal Synthetic Controls[10]. Counterfactual Explanation for Time Series aims to interpret model decisions by generating minimal perturbations or alternative histories, while Specialized Applications and Domain-Specific Methods tailor these ideas to domains like healthcare, recommendation systems, and autonomous driving. A particularly active line of work explores how modern generative architectures can be adapted for causal reasoning. Diffusion-based and flow-based models are being extended to handle interventions and counterfactual queries, balancing sample quality with adherence to learned causal structures. Causal Diffusion Generation[0] exemplifies this direction, situating itself within the Diffusion-Based Causal Generation cluster alongside approaches that use variational inference (e.g., Counterfactual Variational Inference[31]) or flow models (DoFlow[37]). Compared to works like Counterfactual VAR Interventions[1] or Counterfactual Behavior Forecasting[3], which often rely on vector autoregressive or trajectory-prediction frameworks, Causal Diffusion Generation[0] emphasizes the flexibility and expressiveness of diffusion processes for capturing complex temporal dependencies under intervention. A central open question across these branches is how to ensure that generated counterfactuals remain both realistic and causally consistent, especially when ground-truth causal graphs are partially known or must be inferred from observational data.

Claimed Contributions

Causal time series generation task family

The authors formalize causal time series generation as an extension of conditional TSG along Pearl's causal ladder, introducing two new tasks beyond association: interventional TSG (Int-TSG) and counterfactual TSG (CF-TSG). This framework enables generation under interventions and counterfactual queries rather than purely observational correlations.

6 retrieved papers
CaTSG unified diffusion framework with backdoor-adjusted guidance

The authors introduce CaTSG, a diffusion-based generative framework that derives causal score functions via backdoor adjustment and the abduction–action–prediction procedure. The framework uses backdoor-adjusted guidance and a learnable latent environment bank to support observational, interventional, and counterfactual generation within a single unified model.

8 retrieved papers
Can Refute
Causal score functions for diffusion-based TSG

The authors derive interventional and counterfactual score functions by applying backdoor adjustment and the abduction–action–prediction procedure within the diffusion framework. These causal score functions replace standard conditional scores and enable principled generation across all three causal levels without requiring ground-truth interventional labels.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Causal time series generation task family

The authors formalize causal time series generation as an extension of conditional TSG along Pearl's causal ladder, introducing two new tasks beyond association: interventional TSG (Int-TSG) and counterfactual TSG (CF-TSG). This framework enables generation under interventions and counterfactual queries rather than purely observational correlations.

Contribution

CaTSG unified diffusion framework with backdoor-adjusted guidance

The authors introduce CaTSG, a diffusion-based generative framework that derives causal score functions via backdoor adjustment and the abduction–action–prediction procedure. The framework uses backdoor-adjusted guidance and a learnable latent environment bank to support observational, interventional, and counterfactual generation within a single unified model.

Contribution

Causal score functions for diffusion-based TSG

The authors derive interventional and counterfactual score functions by applying backdoor adjustment and the abduction–action–prediction procedure within the diffusion framework. These causal score functions replace standard conditional scores and enable principled generation across all three causal levels without requiring ground-truth interventional labels.

Table of Contents