acceptodds
Under review as a conference paper at ICLR 2027

Causal time series generation via diffusion models

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 whose sampling is guided by backdoor-adjusted score functions targeting interventional and counterfactual distributions under an assumed SCM, while preserving observational fidelity. Extensive experiments on synthetic and real-world datasets show that CaTSG achieves superior fidelity, and on a known-SCM synthetic system, where the interventional target can be constructed directly, it attains the best score on every evaluated metric under multiple environment marginals. Overall, we formalize the causal TSG family and present CaTSG as a first instantiation, providing empirical evidence on synthetic and real-world settings and opening a direction for interventional and counterfactual generation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Generation, Conditional Generation, Time Series Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RrjMOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of generating (conditional) time series by incorporating Pearl's ladder of causation, explicitly allowing the generation of interventional and counterfactual samples under an intervention on a specific context. For this, the authors assume a causal structure where a context variable C impacts the observed time series X and both are (potentially) confounded. Here, the novelty stems from adjusting for the unobserved confounder, modeled as a learned bank of latent environments, when allowing interventions on the context. The method has been evaluated on simulated and real-world data in terms of the quality of the generated.

Critical strengths

The paper makes an important distinction between purely correlation-based and causally motivated generators. The biggest strength here is allowing for unobserved confounders that impact both the context variable (the one of interest for interventions) and the observed time series. The crucial difference from a 'classical' backdoor-adjustment setting is that the confounder is not observed and has to be inferred from the data, i.e., adjusting for it correctly under a given intervention on C is a non-trivial problem.

Critical weaknesses

The novelty of the work relies on the confounder being unobserved, as it is otherwise a 'typical' setting that can be addressed with standard backdoor adjustment and any conditional generative model. However, the paper does not show that the learned hidden confounder actually performs the adjustment as expected. That is, the paper does not state the assumptions under which the learned (clustered) confounder properly represents the hidden confounder (i.e., what the actual condition is that allows it).

Questions for authors

The work is generally promising and I have a few clarifying questions:

  • Since the adjustment for the hidden confounder is a critical part here, under what assumptions does the learned bank provide a valid adjustment? This is probably my biggest concern here, the lack of a clear formal assumption (and in what sense it is violated in practice).
  • Generally, I see why a diffusion model is a good tool here, but what is the theoretical novelty beyond using diffusion models for the estimation part, which could in principle be done with any other conditional generative model?
  • You are advertising counterfactuals, but Pearl's classical (unit-level) formulation first recovers the exogenous noise of the specific unit. This gives a point estimate whenever that noise is determined by the observation (e.g., in additive noise models). This noise-reconstruction step is missing here. In your case, the abduction step only infers the environment weights, and the counterfactual is then sampled from fresh Gaussian noise, even though the diffusion model could in theory allow it. How would you model point estimates of counterfactuals here, as it seems we can only get a distribution? To clarify what I mean: If we assume an additive noise model Y = f(X) + N, given (x, y) one could first reconstruct the noise via n = y − f(x) and then compute the counterfactual under do(X := x') as y' = f(x') + n. How would this work in your setup?
3 comments
AuthorsOpenReview ↗

Response to Reviewer RrjM

We thank the reviewer for the constructive comments and for recognizing the adjustment for an unobserved confounder of both context and series as the non-trivial part of the problem. We address each of your concerns point by point below.

Q1. Assumptions under which the learned bank gives a valid adjustment. We thank the reviewer for raising this central question. Under the SCM assumptions A1-A2 (Section 4.1), adjusting for the latent confounder EE gives the backdoor decomposition P(X∣do(C))=∑eP(X∣C,e)P(e)P(X\mid\mathrm{do}(C))=\sum_e P(X\mid C,e)P(e) (Eq. 2). Because EE is unobserved, as stated in Section 4.3 and Eq. 6, CaTSG replaces it by the index Z∈{1,…,K}Z\in\lbrace 1,\dots,K \rbrace of a learned environment bank and aggregates the KK environment-conditioned branches with weights produced by EnvInfer. Given the reviewer's concern about whether the learned bank properly represents the hidden confounder, this replacement preserves the decomposition under two sufficient conditions:

  • (A3) Outcome sufficiency and support. The bank realizes an index Z=g(E)Z=g(E) with P(X∣C,E)=P(X∣C,Z)P(X\mid C,E)=P(X\mid C,Z), and every queried cc lies in the support of C∣Z=kC\mid Z{=}k for every kk with P(Z=k)>0P(Z{=}k)>0.
  • (A4) Correct aggregation weights. Generation realizes the P(Z)P(Z)-weighted mixture. At the distribution level, this can be done by sampling Z∼P(Z)Z\sim P(Z). At the score level, it requires aggregating the component scores with the induced state-dependent responsibilities rk,t(xt,c)∝P(Z=k)pt(xt∣c,Z=k)r_{k,t}(x_t,c)\propto P(Z{=}k)p_t(x_t\mid c,Z{=}k).

Under A1-A4, ∑kP(X∣C,Z=k)P(Z=k)=P(X∣do(C))\sum_k P(X\mid C,Z{=}k)P(Z{=}k)=P(X\mid\mathrm{do}(C)). Regarding the reviewer's question of in what sense these conditions are violated in practice, the current training objective enforces neither condition, and each can fail. A3 is violated when the finite bank merges values of EE that have different outcome mechanisms P(X∣C,E)P(X\mid C,E), for example because KK is too small. A4 is violated when EnvInfer's weights differ from the required ones, since swapped prediction does not constrain EnvInfer to recover them. We therefore evaluate the resulting generator end to end against a directly constructed interventional target: on Harmonic-VM, across five training runs, CaTSG is closer to this target than the same-backbone w/o Env in all 16 metric-marginal comparisons (Table 8). To make these conditions explicit, we now state them formally in Appendix H.3 and summarize them in Section 4.1.

Q2. Novelty beyond using diffusion for estimation. We thank the reviewer for this question. As stated in Sections 1 and 4, our contribution beyond using diffusion as an estimator is the formulation of causal time series generation as a task family spanning observational, interventional and counterfactual generation, together with a time-series-specific construction of these targets under latent confounding. Section 4 introduces backdoor adjustment and abduction-action-prediction as two classical tools and adopts diffusion as the generative backbone, so diffusion is the instantiation, not the source of causal identification. Our methodological contribution is to translate these causal targets into a unified conditional time series generator with a learned finite environment representation and environment-conditioned score branches. Appendix H.6 distinguishes this formulation and construction from prior causal generative models.

Q3. Unit-level counterfactuals and noise abduction. We thank the reviewer for this question. As stated in Section 4.1 and the limitations, our current CF-TSG target is deliberately distributional: it preserves the inferred factual environment information while resampling residual generative uncertainty. We therefore do not claim unit-level identification of the exogenous noise or a unique counterfactual point estimate. A point-estimate extension could use a deterministic diffusion inversion: map the factual x0x_0 to a sample-specific latent under (c,w)(c,w) and decode the same latent under (c′,w)(c',w), the diffusion analogue of n=y−f(x)n=y-f(x) followed by y′=f(x′)+ny'=f(x')+n. However, identifying that latent with the exogenous noise of the SCM would require additional invertibility and structural assumptions that our formulation does not impose. Nevertheless, we agree with the reviewer that unit-level noise abduction is a valuable direction for future work.

Thank you again for the valuable comments. We hope our responses address your concerns, and we would be grateful if you could consider raising your score. Please feel free to let us know if you have any further questions; we are happy to discuss them.

I want to thank the authors for their response! Regarding my first question, I was wondering more about whether there are sufficient conditions on the data-generating process, rather than on the learned bank, under which the learned environments provide a valid adjustment.

Regarding the counterfactuals in Pearl's framework, since only the environment is abducted and the remaining noise is redrawn, it seems to coincide with a rung-3 counterfactual only if the environment captures all exogenous variation (otherwise it is closer to a rung-2 interventional distribution conditioned on the abducted environment). It would help to state this assumption and formally distinguish the two rungs (it is not necessarily weakening the work, just clarifying its scope).

AuthorsOpenReview ↗

Response to Reviewer's Follow-Up

Thank you for the kind follow-up and for clarifying these points!

(1) On the data-generating process, a sufficient set of conditions under A1-A2 is: (B1) XX depends on EE only through a finite-valued g(E)g(E) with K∗≤KK^\ast\le K values, (B2) every queried cc lies in the support of C∣g(E)=kC\mid g(E){=}k for every kk, and (B3) the joint distributions P(C,X∣g(E)=k)P(C,X\mid g(E){=}k) belong to a family whose finite mixtures are identifiable. Under B3, the observational distribution P(C,X)=∑kP(g(E)=k)P(C,X∣g(E)=k)P(C,X)=\sum_k P(g(E){=}k)P(C,X\mid g(E){=}k) determines the weights and the components up to relabeling, and B2 gives P(X∣C,g(E)=k)P(X\mid C,g(E){=}k) at every queried cc. Because g(E)g(E) blocks the backdoor path C←E→XC\leftarrow E\rightarrow X, P(X∣do(C))P(X\mid\mathrm{do}(C)) is then identified from observational data, and a bank that reproduces P(C,X)P(C,X) with these components realizes A3 up to relabeling. These conditions establish identifiability, and recovery by CaTSG's training is not guaranteed.

(2) On counterfactuals, thank you for the helpful suggestion. Formally, CF-TSG targets ∑eP(e∣x,c) P(X∣do(C=c′),e)\sum_e P(e\mid x,c)\,P(X\mid\mathrm{do}(C{=}c'),e), an interventional distribution averaged over the abducted environment posterior. It coincides with Pearl's rung-3 counterfactual when XX is determined by (C,E)(C,E).

We have incorporated both points into Section 4.1 and Appendix H.3 of the updated PDF, and we hope this makes the assumptions and scope clearer. Thank you again for the helpful discussion!

Reviewer_jK4dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper looks at conditional time series generation from a causal point of view. Besides the usual observational task P(X | C), it defines an interventional task P(X | do(C)) and a counterfactual task P(X' | X, C, C'), according to Pearl's causal ladder. The authors propose CaTSG, which is a diffusion model in which the latent confounder E is approximated by a small learned bank of K environment embeddings. A module (EnvInfer) trained with a SwAV-style swapped-prediction loss, assigns soft weights over these embeddings, and the K environment-conditioned predictions of the denoiser are mixed with these weights under classifier-free guidance. For interventional sampling the weights are recomputed at every step from (xt,c)(x_t, c). For counterfactual sampling they are computed once from the factual pair (x0,c)(x_0, c) and kept fixed while c is replaced by c'. The method is evaluated on two synthetic damped-oscillator datasets with known structural equations and on two real-world datasets. It is compared against GAN and TimeWeaver baselines.

Critical strengths
  • The task formalization. Splitting conditional time series generation into observational, interventional and counterfactual generation gives the field a clear vocabulary. I see this as the main contribution of the paper and the reason I think it is worth improving.
  • A benchmark where the causal claims can actually be checked. The oscillator datasets come with known structural equations and paired factual/counterfactual trajectories. The authors are also open about the limits of their approach (E is not identified, the branches are not identified, the formulation is sequence-level only). Most work in this area has no way to test whether the generated samples follow the causal target, so this setup is useful beyond this paper.
Critical weaknesses

1. Causal part of the method and causal target.

  • You need weights re,t∝pt(xt∣c,e) p(e)r_{e,t} \propto p_t(x_t \mid c, e)\,p(e) (Sec. 4.2) for the backdoor adjustment, but EnvInfer is trained with the SwAV-style clustering loss (Sec. 4.3). Neither loss was designed to make qϕq_\phi match re,tr_{e,t}. Since qϕq_\phi also sees c, I would assume it to pick up exactly the C–E dependence the adjustment should remove. I am therefore not convinced Int-TSG samples from P(X | do(C)) (and not just from some mixture version of P(X | C)).
  • Because only the weighted sum of the K branches is supervised (Eq. 7), the model is in the end a mixture of K denoisers. Could the improvements not come from the extra capacity?
  • In Table 2 CaTSG targets P(X | do(C)) but is scored against observational test data from a different environment. Why should this help? This needs an explanation.

2. The evaluation does not show where the gains come from.

  • The w/o Env variant (same backbone, no environment bank) is the most important baseline, but it is not in Table 2 and only shows up for two datasets in Fig. 5c.
  • Table 2 uses ω = 0.2, while the controlled interventional comparison in Sec. 5.2 uses ω = 0. How was ω chosen?
  • The counterfactual results only compare against ablations. On the synthetic data the true X' is known for every pair, so why not report the error to it instead of only MDD and KL?
  • Many Table 2 entries have ±0.000 over five training runs, also for GANs (e.g. WaveGAN, Air Quality, KL 16.513 ± 0.000). This seems unlikely and also affects the claim about non-overlapping intervals.

Also a minor point: privacy is named as a motivation but never evaluated. Since the counterfactual mode conditions on a real individual's series, a simple memorization check could be useful.

I am happy to raise my score if the authors add w/o Env to Table 2 for all datasets and show on the synthetic data what qϕq_\phi actually learns compared to re,tr_{e,t}.

Questions for authors
  1. On Harmonic-VM, E is known. How do the weights of qϕq_\phi compare with the true responsibilities re,tr_{e,t} and with the observational posterior p(e∣x,c)p(e \mid x, c)?
  2. How does w/o Env at ω = 0.2 perform on all four datasets in Table 2?
  3. On which split were ω, K and H selected?
  4. Do the five seeds in Table 2 change the training run or only the generation? How do the ±0.000 standard deviations arise?
4 comments
AuthorsOpenReview ↗

Response to Reviewer jK4d (1/2)

We thank the reviewer for the constructive comments and for recognizing the task formalization as the main contribution and the oscillator benchmark as useful beyond this paper. We have completed both analyses the reviewer identified: w/o Env on all four datasets of Table 2 (2a below), and the comparison of EnvInfer's weights with the true responsibilities and the observational posterior on Harmonic-VM (1a below). We address each of your concerns point by point below.

1a / Q1. What EnvInfer learns compared with rr and p(e∣x,c)p(e\mid x,c). We thank the reviewer for raising this concern. Below, we address in two parts the reviewer's concerns by distinguishing the weight-level diagnostic from the target-level evaluation.

Direct comparison. The Harmonic-VM diagnostic shows that qϕq_\phi does not recover the observational posterior posited in the concern. Its mutual information with the context's α\alpha interval is only 0.079-0.126 nats, compared with 0.639 for the observational posterior, and its entropy remains close to the uniform value across noise levels. Its noise-level profile also differs from the interventional responsibilities rtr_t. These results support interpreting qϕq_\phi in CaTSG as learned aggregation weights rather than as a direct estimator of either reference quantity, consistent with objectives that supervise the aggregate branch prediction and augmentation-consistent assignments (Appendix C.2). The comparison over the five training runs of Table 8 is given below, with both references computed in closed form from the known components (MI: mutual information in nats between the argmax and the α\alpha interval of the context):

ttEntropy of qϕq_\phiEntropy of rtr_tMI, qϕq_\phiMI, rtr_t
101.3521±0.00051.3521 \pm 0.00050.00280.1257±0.03660.1257 \pm 0.03660.6313
501.3521±0.00051.3521 \pm 0.00050.06230.1211±0.04050.1211 \pm 0.04050.5048
1001.3521±0.00051.3521 \pm 0.00050.18360.1072±0.04670.1072 \pm 0.04670.3558
1501.3521±0.00061.3521 \pm 0.00060.35960.0879±0.05190.0879 \pm 0.05190.1556
1901.3521±0.00061.3521 \pm 0.00060.58390.0794±0.05970.0794 \pm 0.05970.0001

Target-level evidence. At the target level, the result is consistent across all comparisons: over five training runs on Harmonic-VM, CaTSG is closer than the same-backbone w/o Env model to the directly constructed P(X∣do(C))P(X\mid\mathrm{do}(C)) target in all 16 metric-marginal comparisons (Table 8). Thus, the weight diagnostic sharpens the interpretation of EnvInfer, while Table 8 directly supports the generated distribution as an empirical approximation to the interventional target.

1b. Extra capacity. As stated in Section 4.2, the KK predictions share all denoiser parameters: they are evaluations of the same backbone under different environment embeddings, rather than KK independently parameterized denoisers. The closest parameter- and computation-matched control is w/o SW, which retains the same environment bank, EnvInfer and KK-branch evaluations but removes the swapped-prediction loss and its warm-up stage (Table 7). On Harmonic-VM, it performs substantially worse than CaTSG (KL 0.114 vs 0.017, MMD 0.049 vs 0.012; Figure 5c). Thus, the KK-branch architecture and computation alone do not reproduce the gain, and the complete environment-training procedure contributes beyond the shared architecture. The w/o Env ablation, which keeps the same denoiser backbone and sampler (Section 5.5), separately rules out a larger denoiser backbone. We have also revised the discussion of the Table 3 counterfactual results in Section 5.3.

1c. Why an interventional target should help on observational test data. Regarding why targeting P(X∣do(C))P(X\mid\mathrm{do}(C)) can help on observational test data from a different environment, under the SCM the training observational target, the test distribution and the interventional target weight the same components P(X∣C,e)P(X\mid C,e) by Ptrain(e∣C)P_{\text{train}}(e\mid C), Ptest(e∣C)P_{\text{test}}(e\mid C) and P(e)P(e), respectively. The interventional target can therefore help only when its mixture is closer to the test mixture than the training observational mixture; the SCM alone does not guarantee this ordering. As stated in Appendix G.1, Table 2 compares generated samples with held-out observational data, whereas Table 8 compares them with P(X∣do(C))P(X\mid\mathrm{do}(C)) itself. We now state when the interventional target can help in Section 5.2.

AuthorsOpenReview ↗

Response to Reviewer jK4d (2/2)

2a / Q2. w/o Env in Table 2. Following the reviewer's suggestion, we have added w/o Env at ω=0.2\omega=0.2 on all four datasets to Table 2. CaTSG improves on w/o Env in all 16 dataset-metric comparisons, for example in J-FTSD:

J-FTSDHarmonic-VMHarmonic-VPAir QualityTraffic
CaTSG0.548±0.0850.548 \pm 0.08512.677±2.43112.677 \pm 2.4312.929±0.4972.929 \pm 0.4970.723±0.0550.723 \pm 0.055
w/o Env2.274±0.1902.274 \pm 0.19013.923±1.26313.923 \pm 1.2636.843±0.8276.843 \pm 0.8273.256±0.4863.256 \pm 0.486

Since w/o Env shares the denoiser and the sampler with CaTSG, the gains come from the environment-conditioned components rather than from the shared backbone.

2b / Q3. Choice of ω\omega, KK and HH. As stated in Section 5, KK and HH were selected per dataset on the validation split; Figure 5d reports test-split scores after selection, which we now state in the figure description. Regarding how ω\omega was chosen, as stated in Appendix G.5, we use ω=0.2\omega=0.2 as the common default for all datasets; we fixed it in advance, considering that small guidance weights give the best distribution-level fidelity (FID) in the original classifier-free guidance study [1], and we now state this reason in the same appendix. The controlled interventional comparison uses ω=0\omega=0 because that is the unguided setting of the score identity it tests (Eq. 4), as stated in Appendix G.1; Table 9 reports the same comparison at ω=0.2\omega=0.2.

2c. Error to the true counterfactual. Table 3 compares CaTSG with its ablations because the baselines of Table 2 do not support the factual-pair-conditioned C→C′C\to C' task (Section 5.3). The ground-truth counterfactual set is the reference for MDD and KL. The reviewer is correct that these marginal metrics do not assess factual–counterfactual pairing, because they are invariant to permuting generated samples across factual units. As stated in Section 4.1, our sampler preserves the inferred factual environment while resampling residual uncertainty, so it targets a counterfactual distribution rather than recovery of the particular residual realization paired with each factual sample. Pairwise error therefore evaluates the stronger unit-level target, and we interpret Table 3 only as evidence of marginal counterfactual fidelity.

2d / Q4. The $\pm$0.000 entries. Thank you for catching this. Admittedly, each method in Table 2 was trained once per dataset, and the reported mean ±\pm std summarizes five sample sets generated independently from that trained model. We have revised Section 5.2 and Appendix F.3 accordingly and removed the interval-overlap analysis. The point estimates remain a controlled comparison under the same sampling protocol, while the ±\pm values quantify sampling rather than training variability.

Minor. Privacy. We would like to clarify that privacy appears only in the first sentence of the introduction, as a general application of time series generation; CaTSG makes no privacy claim.

Thank you again for the constructive comments and for clearly stating what would change your assessment; we would be grateful if you could consider raising your score. Please feel free to let us know if you have any further questions. Thank you!

References

[1] Classifier-Free Diffusion Guidance. arXiv:2207.12598, 2022.

Thank you for the response and the additional experiments.

I am keeping my score because of the qϕq_\phi analysis. Across noise levels, qϕq_\phi gives nearly uniform weights and recovers neither re,tr_{e,t} nor the observational posterior p(e∣x,c)p(e \mid x, c). By your own condition A4, the backdoor adjustment requires aggregating the branches with these responsibilities, so the model does not perform the adjustment described in Section 4. The revised paper still presents qϕq_\phi as a proxy for re,tr_{e,t} and does not include this analysis.

We thank the reviewer for the careful follow-up and address the two remaining points below.

Description of qϕq_\phi and the diagnostic. We have removed the remaining proxy descriptions of qϕq_\phi (Section 4.3, Appendices C.3 and D), which now describe it as aggregation weights that are not trained to recover re,tr_{e,t}. We did not include the weight diagnostic at first because the paper evaluates its interventional claim on the generated distribution (Table 8), and Appendix G.1 already notes that this does not test whether qϕq_\phi recovers re,tr_{e,t}. As it addresses the reviewer's question, we have now added it to Appendix G.1 (Table 10).

Whether the model performs the adjustment. Regarding A4, re,tr_{e,t} refers to the true environment components, whereas qϕq_\phi weights learned branches that are not constrained to recover them, so the mismatch limits the component-level interpretation of qϕq_\phi. As stated in Section 4.1 and Appendix H.3, which we added in response to Reviewers MSCZ and RrjM, A3-A4 are sufficient conditions for exact adjustment that the current training does not enforce, and we therefore assess the generated distribution against the target directly: against P(X∣do(C))P(X\mid\mathrm{do}(C)) constructed from the known SCM, CaTSG is closer than the same-backbone w/o Env in all 16 metric-marginal comparisons across five training runs (Table 8).

We sincerely thank the reviewer again for the careful review. We hope the clarified scope, together with these additions and the direct interventional evaluation in Table 8, addresses your concerns, and we would be grateful if you could take them into account in your final assessment. Thank you very much!

Reviewer_MSCZOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper distinguishes observational, interventional, and counterfactual time-series generation under a sequence-level structural causal model with a latent environment affecting both conditions and outcomes. It introduces CaTSG, a diffusion framework combining environment-specific denoising branches, an environment inference network, swapped-prediction training, and environment-code regularization. Interventional sampling recomputes environment weights along the diffusion trajectory, whereas counterfactual sampling fixes weights inferred from the factual sequence while changing the conditioning trajectory. These are whole-trajectory interventions rather than sequential interventions with feedback. Evaluation covers two oscillator systems, air quality, and traffic. The reported results show strong distributional fidelity, including a direct known-SCM intervention experiment. However, the central issue is whether the learned sampler realizes the stated causal distributions, rather than merely producing accurate observational or environment-conditioned samples.

Critical strengths
  1. A consequential problem formulation with some direct causal-target evaluation. Separating observational, interventional, and counterfactual generation identifies a real limitation of standard conditional generation. Appendix G.1 constructs an interventional target directly from a known SCM, and CaTSG leads all 16 reported combinations of environment marginals and metrics. This is substantially more informative than relying solely on plausible-looking real-world examples.
  2. Strong generative results and useful transparency about scope. Table 2 reports the best point estimate in 14 of 16 dataset–metric combinations, with five training seeds in the main quality evaluation and a same-backbone environment-free comparison. The appendix also acknowledges non-identifiability, sequence-level interventions, and residual resampling. These disclosures make the work more assessable and should be retained in a narrower interpretation of the contribution.
Critical weaknesses
  1. The learned procedure is not sufficiently connected to the causal distributions that define the contribution. The backdoor identity requires valid component mechanisms and adjustment weights, but the loss supervises their combined denoising prediction rather than identifying those components. Observational responsibilities satisfy p(e|x,c) ∝ p(x|c,e)p(c|e)p(e), whereas interventional responsibilities omit p(c|e). It is unclear how the same inference network learns both roles without a correction or additional identifying information. Separately, for factual weights w_e, the score of the counterfactual mixture Σ_e w_e p_t(x'_t|c',e) requires responsibilities that depend on the current noisy sample. Fixing w_e and averaging component scores is generally not equivalent; even a two-component Gaussian example can exhibit different modal behavior. The acknowledged approximation therefore concerns the target itself, not just numerical accuracy. These gaps prevent the empirical fidelity results from establishing the central causal claim. See Sections 4.1–4.3, Appendix C.2, and Algorithm 3.
  2. The experimental evidence does not sufficiently isolate causal correctness or generalization. In Traffic, length-96 windows with stride one are constructed before splitting, allowing raw hourly observations to recur across splits, and the target scaler is fitted before splitting (Appendix E.2). This limits the interpretation as generalization to unseen observations. In the reintegrated oscillator tasks, acceleration can be estimated directly from the supplied velocity or position trajectories, making derivative-based baselines essential to interpreting the apparent benefit of latent-environment inference. This derivative objection does not apply in the same way to the separate fixed-trajectory intervention experiment, but that experiment uses a single training checkpoint. Finally, marginal counterfactual distances do not establish factual–counterfactual pairing accuracy, and residual resampling limits individual-level counterfactual semantics. Together, these issues leave the strongest contribution insufficiently validated.
Questions for authors
  1. Clarification of the causal target: What assumptions or training constraints ensure that EnvInfer produces interventional responsibilities rather than observational posteriors? If no such guarantee is available, which causal-target claims do you retain as empirical approximations?
  2. Assessment-changing methodological check: How does the current fixed-weight counterfactual sampler compare with sampling an environment once from the factual posterior and generating conditionally on that environment? A known multimodal SCM would directly reveal whether the approximation changes the intended distribution.
  3. Assessment-changing reevaluation: Do the Traffic gains persist when raw time blocks are split before windowing, an appropriate purge gap is used, and all normalization statistics are fitted on training data only?
  4. Assessment-changing controls: On the reintegrated oscillator tasks, how competitive are finite-difference or smoothed-derivative baselines and a capacity-matched mixture model? For the direct intervention experiment, do the gains persist across training seeds, and how is the requested target environment marginal supplied to the sampler?
3 comments
AuthorsOpenReview ↗

Response to Reviewer MSCZ (1/3)

We thank the reviewer for the constructive comments and for recognizing the direct evaluation against a constructed interventional target and the transparency about the method's scope. We address each of your concerns point by point below.

W1 / Q1. How EnvInfer can play the interventional role. We thank the reviewer for this important question. Below, we address in turn the role of EnvInfer, the conditions for a valid adjustment, and the claim we retain.

(1) Role of EnvInfer. EnvInfer is not trained as a supervised estimator of either the observational posterior p(e∣x,c)p(e\mid x,c) or the interventional responsibility re,tr_{e,t}, which differ by the factor p(c∣e)p(c\mid e) (Appendix C.2). Rather, it learns aggregation weights jointly with the environment-conditioned branches. As stated in Section 4.3 (Eq. 7), the denoising objective supervises only their weighted aggregate, so the branches and the weights are not separately identified, which we now state in Section 4.2. Adjusting with a learned latent-environment representation follows a number of prior works, e.g., [1-3]. As in these works, its causal interpretation rests on stated assumptions rather than on recovering the latent confounder itself, which we make explicit next.

(2) Conditions for a valid adjustment. Specifically, in addition to the SCM assumptions A1-A2 stated in Section 4.1, a learned index Z=g(E)Z=g(E) preserves the backdoor decomposition under two sufficient conditions: (A3) P(X∣C,E)=P(X∣C,Z)P(X\mid C,E)=P(X\mid C,Z), with every queried cc in the support of C∣Z=kC\mid Z{=}k for every kk with P(Z=k)>0P(Z{=}k)>0, and (A4) generation uses P(Z)P(Z) as the mixture weights, either by sampling Z∼P(Z)Z\sim P(Z) or through the state-dependent responsibilities that these weights induce. Under A1-A4, ∑kP(X∣C,Z=k) P(Z=k)=P(X∣do(C))\sum_k P(X\mid C,Z{=}k)\,P(Z{=}k)=P(X\mid\mathrm{do}(C)). The current training objective enforces neither condition, so we do not interpret qϕq_\phi itself as an identified causal responsibility. We now state these sufficient conditions in Appendix H.3 and summarize them in Section 4.1.

(3) Claim supported by the evidence. The empirical claim we retain concerns the generated distribution as a whole. To evaluate this claim across training runs, we have retrained CaTSG and the same-backbone w/o Env five times each and compared their samples on Harmonic-VM with the interventional target constructed directly from the known SCM (Appendix G.1):

P(E)P(E)MethodMDDKLMMDCond-MSE
80-20CaTSG0.0061±0.00080.0061 \pm 0.00080.0120±0.00140.0120 \pm 0.00140.0059±0.00280.0059 \pm 0.00280.1382±0.02440.1382 \pm 0.0244
w/o Env0.0186±0.00160.0186 \pm 0.00160.3055±0.04570.3055 \pm 0.04570.1245±0.01030.1245 \pm 0.01030.6158±0.03610.6158 \pm 0.0361
Full rangeCaTSG0.0124±0.00120.0124 \pm 0.00120.0640±0.01230.0640 \pm 0.01230.0179±0.00270.0179 \pm 0.00270.2963±0.02210.2963 \pm 0.0221
w/o Env0.0250±0.00130.0250 \pm 0.00130.6065±0.08500.6065 \pm 0.08500.1860±0.01370.1860 \pm 0.01370.9955±0.05340.9955 \pm 0.0534
Test intervalCaTSG0.1736±0.02540.1736 \pm 0.02540.0248±0.01190.0248 \pm 0.01190.0116±0.00590.0116 \pm 0.00590.1113±0.02750.1113 \pm 0.0275
w/o Env0.2541±0.04950.2541 \pm 0.04950.1215±0.02950.1215 \pm 0.02950.0972±0.00820.0972 \pm 0.00820.4363±0.02350.4363 \pm 0.0235
Equal mixtureCaTSG0.0151±0.00080.0151 \pm 0.00080.1402±0.02670.1402 \pm 0.02670.0330±0.00500.0330 \pm 0.00500.5058±0.02120.5058 \pm 0.0212
w/o Env0.0280±0.00150.0280 \pm 0.00150.9678±0.10630.9678 \pm 0.10630.2246±0.01530.2246 \pm 0.01531.3410±0.06651.3410 \pm 0.0665

CaTSG is closer to the target than w/o Env in all 16 metric-marginal comparisons, which supports the generator as an end-to-end approximation to the interventional target without requiring the individual components or weights to be uniquely identified. We have added these five-run results to Table 8 and revised the description of EnvInfer's weights in Section 4.2 accordingly.

AuthorsOpenReview ↗

Response to Reviewer MSCZ (2/3)

W1. Fixed-weight counterfactual rule. We agree with the reviewer that the default fixed-weight rule generally changes the sampled distribution, rather than merely introducing numerical approximation error. Specifically, ∑ewe∇log⁡pt(xt′∣c′,e)=∇log⁡∏ept(xt′∣c′,e)we\sum_e w_e\nabla\log p_t(x'_t\mid c',e)=\nabla\log\prod_e p_t(x'_t\mid c',e)^{w_e}, so the rule gives the score of a normalized weighted geometric mean rather than that of the AAP mixture. We have revised this characterization in Section 4.2. This revision concerns the default sampling rule; the AAP counterfactual estimand defined in Section 4.1 remains unchanged.

Q2. Comparison with sampling one environment from the factual weights. To address the reviewer's question, we have compared the two rules on the trained models and on a known multimodal SCM.

Trained models. We have added the single-draw sampler, which draws k∼wk\sim w once per factual unit and runs the branch εθ(⋅,t,c′,ek)\varepsilon_\theta(\cdot,t,c',e_k) throughout generation, and compared both rules against the constructed counterfactual target over five training runs:

Counterfactual ruleMDDKLPE
Fixed weights (default)0.2120 ±\pm 0.05690.1185 ±\pm 0.12181.0454 ±\pm 0.3917
Single draw0.2120 ±\pm 0.05690.1189 ±\pm 0.12261.0461 ±\pm 0.3910

Within every training run, the differences are far smaller than the run-to-run variation, so the choice of rule does not materially affect the reported Harmonic-VM comparison.

Known multimodal SCM. We replaced the trained model by the exact scores of a known mixture: for each Harmonic-VM counterfactual context, the environments α∈{0.1,0.4,0.8}\alpha\in\lbrace 0.1,0.4,0.8 \rbrace give three distinct counterfactual trajectories, and we sampled with equal weights:

RuleMDDKL
Exact mixture0.0178 ±\pm 0.00220.0011 ±\pm 0.0007
Fixed weights0.1455 ±\pm 0.00110.4937 ±\pm 0.0171
Single draw0.0181 ±\pm 0.00130.0007 ±\pm 0.0005

With balanced weights, the fixed-weight rule changes the distribution, as characterized in W1 above: its samples lie between the components rather than on one of them, while the single-draw rule matches the exact mixture. With weights as concentrated as the learned factual weights qϕ(e∣x0,c)q_\phi(e\mid x_0,c) (mean largest weight 0.935), the departure is smaller (MDD 0.0276±0.00110.0276\pm0.0011 vs 0.0074±0.00050.0074\pm0.0005 for the exact mixture).

W2 / Q3. Traffic splitting and normalization. We thank the reviewer for raising this point. As stated in Appendix E.2, the original Traffic experiment evaluates window-level, temperature-stratified conditional generation rather than chronological extrapolation over disjoint raw observations. This window-first protocol follows prior conditional TSG evaluation on the same dataset [4], and all methods in Table 2 are compared under it.

To evaluate the setting raised by the reviewer, we additionally split raw time blocks before windowing, removed the windows within a purge gap equal to the window length (96 hourly steps), and fitted all normalization statistics on the training split only:

MethodMDDKLMMDJ-FTSD
TimeGAN0.439 ±\pm 0.00010.484 ±\pm 0.0000.788 ±\pm 0.00018.261 ±\pm 1.361
Pulse2Pulse0.283 ±\pm 0.0014.873 ±\pm 0.0370.441 ±\pm 0.00128.199 ±\pm 11.366
TW-SSSD0.160 ±\pm 0.0000.268 ±\pm 0.0020.076 ±\pm 0.0005.512 ±\pm 0.497
w/o Env0.106 ±\pm 0.0010.390 ±\pm 0.0250.035 ±\pm 0.0023.986 ±\pm 0.317
CaTSG0.101 ±\pm 0.0030.272 ±\pm 0.0900.028 ±\pm 0.0043.752 ±\pm 0.155

Under this split, CaTSG remains better than the same-backbone w/o Env model on all four metrics and attains the lowest MDD, MMD and J-FTSD. Its mean KL is close to TW-SSSD (0.272 versus 0.268). Because the test set contains only 354 windows from six blocks, we treat this result as a sensitivity analysis rather than a replacement for the original window-level comparison.

AuthorsOpenReview ↗

Response to Reviewer MSCZ (3/3)

W2 / Q4. Derivative baselines. To address the reviewer's question, we have added finite-difference and smoothed-derivative baselines:

DatasetMethodMDDKLMMD
Harmonic-VMCentral difference of vv0.00380.00520.0039
Smoothed derivative of vv0.00130.00090.0005
CaTSG0.0980.0170.012

These baselines explicitly encode the derivative relation, whereas CaTSG and the other learned methods are not given this operator in their training objectives; on Harmonic-VM they are closer to the test distribution than CaTSG. Therefore, the Harmonic-VM results of Table 2 alone cannot attribute the improvement over the learned baselines solely to latent-environment inference, and we have revised Section 5 accordingly. The shortcut does not apply to the direct interventional evaluation of Table 8, because differentiating the factual velocity trajectory cannot construct its interventional target distribution, which reweights the environments.

Q4. Capacity-matched mixture model. Regarding the capacity-matched mixture model, the closest such control among our ablations is w/o SW (Figure 5c). It retains the same shared denoiser, environment bank, EnvInfer and KK environment-conditioned evaluations per sampling step as CaTSG, and therefore the same parameters and computation, but it omits the swapped-prediction loss and its warm-up stage, so EnvInfer receives no training gradient (Table 7). On Harmonic-VM, CaTSG obtains KL 0.017 and MMD 0.012, compared with 0.114 and 0.049 for w/o SW. Thus, the parameter- and computation-matched architecture without the proposed assignment training is substantially less competitive.

Q4. Training seeds for the direct intervention experiment. Following the reviewer's suggestion, we have retrained CaTSG and w/o Env five times each and now report their mean ±\pm std over these five training runs in Table 8 (shown in our reply to W1/Q1 above); the gains persist across training runs: the CaTSG mean is lower than that of w/o Env in all 16 metric-marginal combinations, with an across-run standard deviation of CaTSG below a third of this gap in every combination, and on the primary 80-20 marginal the worst of the five CaTSG runs has MDD 0.0072, against 0.0186±0.00160.0186 \pm 0.0016 for the five w/o Env runs.

Q4. How the target environment marginal reaches the sampler. We would like to clarify that the sampler does not receive the target environment marginal: CaTSG uses weights learned from the training data, and the same generated samples are evaluated against all four marginals. We now state this explicitly in Appendix G.1.

W2. Pairing accuracy and individual-level semantics. As stated in Section 4.1 and the limitations, our CF-TSG formulation is distributional: it preserves the inferred factual environment information while resampling residual generative uncertainty. Accordingly, Table 3 evaluates marginal counterfactual fidelity rather than unit-level factual–counterfactual pairing. Indeed, marginal distances are invariant to a permutation of the generated samples and therefore cannot establish pairwise recovery. We interpret these results only as evidence of marginal counterfactual fidelity, not as a claim of unit-level counterfactual identification.

Thank you again for the detailed comments, which have helped us improve the paper. We hope the additional analyses and revisions address your concerns, and we would be grateful if you could take them into account in your final assessment. Please feel free to let us know if you have any further questions; we are happy to discuss them.

References

[1] Time Series Deconfounder: Estimating Treatment Effects over Time in the Presence of Hidden Confounders. International Conference on Machine Learning, 2020.

[2] Towards Out-of-Distribution Sequential Event Prediction: A Causal Treatment. Advances in Neural Information Processing Systems, 2022.

[3] Deciphering Spatio-Temporal Graph Forecasting: A Causal Lens and Treatment. Advances in Neural Information Processing Systems, 2023.

[4] Time Weaver: A Conditional Time Series Generation Model. International Conference on Machine Learning, 2024.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.