acceptodds
Under review as a conference paper at ICLR 2027

Learning Multimodal Predictive Regions with Diffusion: Conformal Prediction for Multivariate Time Series

13 views

Abstract

Multivariate time series forecasting informs decision making in domains such as traffic and energy systems. In these domains, effective decisions require understanding the uncertainty in future outcomes and the dependencies between them. To quantify this uncertainty, conformal prediction offers a model-agnostic way to calibrate multivariate prediction regions. However, it typically assumes exchangeability, which is generally violated by temporally dependent data. In addition, conformal prediction relies on connected regions, which is inefficient for systems like traffic or energy models when the predictive distribution is multimodal or non-connected. We address both issues using a conditional diffusion model within a conformal framework. Rather than running the reverse diffusion process, we use denoising errors to assess how well a possible future outcome fits the learned conditional distribution. These errors are averaged into a scalar nonconformity score and calibrated using split conformal prediction. For finite samples from strictly stationary -mixing time series, we establish bounds on marginal and conditional coverage, with a small penalty for temporal dependence. Experiments on non-Gaussian multimodal VAR processes and the Traffic and Solar datasets show competitive coverage/volume tradeoffs, with at least 7% smaller prediction regions in strongly multimodal settings.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
machine learning, time series, diffusion model, conformal prediction, multimodal distribution, artificial intelligence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2LzAOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2(−1)
Presentation
2
Soundness
2(−1)
Read review
Paper summary

This paper proposes a diffusion-based conformal prediction method that uses denoising errors as nonconformity scores to construct flexible prediction regions for multivariate time series. The authors establish finite-sample approximate marginal and conditional coverage guarantees under stationary β\beta-mixing assumptions.

Critical strengths
  1. The proposed method is novel, well-motivated and supported by a theoretically grounded framework.

  2. The experimental results are promising, particularly in demonstrating the potential of the proposed method to construct flexible prediction regions.

Critical weaknesses
  1. The paper lacks sufficient validation and discussion of its theoretical assumptions, including the β\beta-mixing assumption for time-series data and the assumptions concerning diffusion model estimation (see Questions for Authors).

  2. Several important methodological and experimental details are insufficiently explained, limiting clarity (see Questions for Authors).

  3. The experimental evaluation is somewhat limited in terms of baseline selection and response dimensions dyd_y, making it difficult to show broader advantages of the proposed method.

Questions for authors
  1. Assumption 2 introduces an independent-block learning condition with a deterministic approximation radius rm,ar_{m,a}. However, the authors may need to discuss whether this assumption holds for the diffusion model used in the experiments, or to provide an explicit convergence assumption for it. Although Remark 4.1 shows that the dependence penalty under geometric (\beta)-mixing vanishes as the training sample size increases, this does not guarantee that rm,a→0r_{m,a}\to 0. Consequently, Corollary 4.2 does not establish convergence of the proposed nonconformity score to the oracle score, which still depends on rr. It would strengthen the theoretical analysis if the authors could provide sufficient conditions under which rm,ar_{m,a} vanishes, along with an explicit convergence rate or a justification of Assumption 2 for their neural diffusion estimator.

  2. Theorem 4.4 establishes conditional coverage over neighborhoods AA satisfying P(X∈A)≥γP(X\in A)\geq\gamma, rather than the standard conditional coverage guarantee conditioned on X=xX=x. While this distinction is acknowledged in the theoretical formulation, the resulting guarantee is substantially weaker, and the coverage error may become large for small γ\gamma. Furthermore, it looks like the theorem relies on neighborhood-specific calibration, which differs from the global calibration procedure in Algorithm 1. Additionally, it is unclear how these neighborhoods are selected in practice, and the experiments do not appear to evaluate the proposed localized conditional coverage. The authors should clarify the practical implications of Theorem 4.4 and provide empirical evidence supporting the conditional coverage claims.

  3. The authors should provide further motivation and explanation for the use of antithetic sampling and common random numbers (CRN). Their roles in the proposed conformal prediction framework are not immediately clear from the current presentation. In particular, it would be helpful to clarify how these are beneficial.

  4. It would be helpful to visually distinguish methods that achieve the target coverage level in Table 1. In its current form, it is difficult to identify which methods meet the target coverage, particularly when comparing prediction-region volumes.

  5. The authors specify the deterministic predictors used by the baseline methods (LSTM and LOO bootstrap), but do not clearly identify the deterministic predictor used in the proposed residual-based method (Algorithm 2) for the main experiments in Table 1. Could the authors clarify which predictor is used and provide its implementation details?

  6. Could the authors report the computational cost of each method, such as wall-clock training and inference times? If detailed runtime measurements are not currently available, approximate estimates would still be helpful for understanding the computational tradeoffs among the methods.

  7. In Section 5, the authors state that "all volumes are measured using the common importance-sampling procedure." Could the authors confirm whether this procedure is applied consistently to all methods, including CONTRA and FlowCP? Additionally, could the authors explain the rationale for choosing importance sampling over uniform Monte Carlo or quasi-Monte Carlo approaches used in prior evaluations of these methods?

  8. In Appendix B.4, the authors state that 16,384 samples are used for volume estimation when dy=8d_y=8. However, to my understanding, the experiments only consider dy∈2,4d_y\in{2,4}. Could the authors clarify whether any experiments with dy=8d_y=8 were conducted? Furthermore, I believe the experimental evaluation would be strengthened by considering higher response dimensions, as explored in existing studies [1,2,3]. Such experiments would provide additional evidence of the proposed method's effectiveness and scalability.

  9. Another active research direction for constructing flexible prediction regions involves optimal transport [4,5]. While I am not necessarily requesting additional experiments with these methods, I believe the current baselines represent a relatively limited subset of existing approaches. A broader discussion of related methods, particularly those based on optimal transport, would help better contextualize the contributions and advantages of the proposed method.

  10. In prior studies [2,3], Traffic and Solar have primarily been used as multivariate forecasting benchmarks dy>1d_y>1, rather than datasets with explicitly demonstrated multimodal conditional distributions. Could the authors clarify whether these datasets exhibit multimodality or, how the datasets were processed to have multimodality Additionally, visualizations of the prediction regions for dy=2d_y=2 on these real-world datasets would be valuable for understanding the proposed regions and comparisons with the baseline methods.

  11. Based on Algorithm 2, the authors used training set to train deterministic predictors and diffusion model. Can authors clarify how data is split for each baseline method?

[1] CONTRA: Conformal prediction region via normalizing flow transformation

[2] Flow-based Conformal Prediction for Multi-dimensional Time Series

[3] Conformal prediction for multi-dimensional time series by ellipsoidal sets

[4] Multivariate Conformal Prediction using Optimal Transport

[5] Optimal Transport-based Conformal Prediction

7 comments

Thank you for the sufficiently detailed review! We have clarified the learning assumption and localized coverage, moved Algorithm 2 into the main text, updated the coverage-aware tables, and added training-time estimates and optimal-transport comparisons. The 95% comparison in Appendix B.6 now includes all five Traffic subsets: Algorithm 1 undercovers on Traffic, whereas Algorithm 2 attains above-target mean coverage. We have also generated the requested real-data visualizations. Below, we distinguish completed revisions from remaining work. We address every point one by one below.

  1. Assumption 2 and convergence of the neural estimator

We agree that a vanishing mixing penalty does not imply that the denoiser approximation radius vanishes. We have updated Assumption 2 in Section 4.1, highlighted in blue, to separate the finite-sample learning condition from an explicit asymptotic assumption requiring rm,a→0r_{m,a} \to 0. The revised remark gives a convergence rate conditional on an assumed rate for $r_{m,a} $, and clarifies the additional learning condition needed for oracle-score convergence. We do not establish this learning condition for the neural estimator used in our experiments.

  1. Localized VS pointwise conditional coverage

We have clarified the scope and implementation of Theorem 4.4 in Section 4.3, highlighted in blue. It provides neighborhood-level coverage conditional on X∈AX\in A, where P(X∈A)≥γP(X \in A) \ge \gamma, using a calibration quantile computed from observations within A. This is not pointwise coverage conditional on X=xX=x. Neighborhoods can be defined from training data, subject to the theorem’s assumptions. Smaller neighborhoods provide finer localization at the cost of a larger finite-sample error bound. The current experiments evaluate global calibration. Evaluating neighborhood-level coverage requires a separate experiment using the localized calibration procedure.

  1. Antithetic sampling and common random numbers

These techniques control different aspects of Monte Carlo variability.

  • Antithetic sampling pairs Gaussian perturbations ϵ\epsilon and −ϵ-\epsilon at the same diffusion time. At equal evaluation cost, this can reduce score estimation variance when the paired losses have negative covariance.
  • Common random numbers reuse diffusion times and perturbations across calibration observations and candidate responses. This gives a fixed score function and repeatable membership decisions. Sharing draws can reduce the variance of score differences when the resulting estimates are positively correlated.

Neither technique automatically improves coverage. Our validity argument treats the bank as frozen and uses the stated temporal-dependence bounds. We have updated the real-data ablations in Section 5, highlighted in blue, with details in Appendix C.

  1. Identifying methods that attain target coverage

We have updated Table 1 in Section 5, highlighted in blue, to report the 90% results, with entries below nominal mean coverage grayed out and excluded from the highlighted volume ranking. The corresponding 95% results, including all five Traffic location subsets, are reported in Appendix B.6, also highlighted in blue.

  1. Deterministic predictor used by Algorithm 2

The real-data implementation uses ordinary least squares with an intercept and no ridge penalty. It’s fitted on a flattened 50-step histories from the proper training block. The predictor is then frozen, and an unconditional diffusion model is trained on its residuals. We have moved the general residual formulation into Section 3.3, highlighted in blue, and will add these implementation details to Appendix B.1 to distinguish the evaluated specialization from the general conditional formulation.

  1. Computational cost

We have added approximate wall-clock training-time bounds, in hours, to Appendix B.5, highlighted in blue.

  1. Consistency and motivation of volume estimation

The implementation uses importance sampling to estimate real-data prediction-region volumes for DISCO, FlowCP, CONTRA, MultiTransport, OpTransport, and the global Mahalanobis baseline. MultiDimSPCI is an exception: its ellipsoidal-shell volume is computed analytically.

Our prediction regions are evaluated through membership tests and may contain disconnected components. Uniform Monte Carlo or Sobol QMC over a finite box estimates only the volume inside that box. A small box may omit components, while a large box may waste evaluations outside the region. For real-data experiments, we therefore use importance sampling with a Gaussian kernel-density proposal and a heavy-tailed Student-t component. This provides full support and concentrates evaluations in relevant areas, although numerical integration uncertainty remains. Appendix B.2 already describes randomized QMC volume estimation for the synthetic experiments, and Appendix B.4 describes importance sampling for the real-data experiments.

  1. Higher response dimensions

We evaluated dy=2,4. Traffic and Solar contain only 15 and 9 locations, respectively, so eight-location subsets would overlap substantially across trials. We therefore limited this evaluation to smaller dimensions and acknowledge that higher-dimensional scalability remains untested. Appendix B.4 now lists the correct sampling budgets for the reported dimensions.

  1. Optimal transport methods

Thank you for the interesting direction! These are relevant alternatives, and we have included MultiTransport and OpTransport in the 90% comparison in Section 5 and the 95% comparison in Appendix B.6, highlighted in blue. Appendix B.1 describes their implementation sources, and Appendix B.5 now includes their training-time bounds.

  1. Multimodality and real-data visualizations

We have added a figure for 90% two-dimensional prediction regions on Traffic and Solar in Section 5, with the addition highlighted in blue. These visualizations illustrate region geometry and individual membership decisions. We have not established conditional multimodality in either dataset, and our preprocessing does not introduce it. The synthetic mixtures provide the controlled evaluation of multimodal geometry.

  1. Data splitting for each method

The revised experiments use dataset-specific chronological splits: 80% pre-calibration / 10% calibration / 10% test for Traffic, and 60% pre-calibration / 20% calibration / 20% test for Solar. For DisCo, 20% of the pre-calibration block is reserved for validation, giving approximately 64/16/10/10 for Traffic and 48/12/20/20 for Solar, after constructing and aligning histories.

Algorithm 2 fits its predictor and residual denoiser on the fitting block and computes calibration residuals using the frozen predictor on the separate calibration block. The main comparison aligns methods at target index 54 and uses the same test interval within each dataset. Baselines have method-specific internal fitting and calibration allocations, which we will document explicitly in Appendix B.1.

The revised 90% results are included in Section 5 and the 95% results in Appendix B.6, highlighted in blue. These show competitive volumes at nominal-or-higher mean coverage in several settings. However, in the main 95% comparison in Appendix B.6, Algorithm 1 has below-nominal mean coverage on Traffic, so we do not interpret its smaller volumes there as evidence of superior efficiency.

Thank you for the response. Some of my concerns and questions have addressed and answered. I still have some remaining questions and concerns:

  1. Localized VS pointwise conditional coverage

I see the authors clarified this point more prominent in the revised manuscript. But, the experiment results only report marginal average coverage and no empirical evidence for this claim. Please correct me if I overlook something.

  1. Deterministic predictor used by Algorithm 2

I need further clarification of deterministic predictor. In author's response, The real-data implementation uses ordinary least squares with an intercept and no ridge penalty. Then how does LSTM, LOO estimator used for FlowCP, MultiDimSPCI, and and CONTRA? Did they use different predictor from the proposed methods in Table 1?

  1. Higher response dimensions

I believe experiments with dy=2,4d_y = 2,4 are too low dimension to claim argument about varying dimensions.

  1. Multimodality and real-data visualizations

Thank you for updating the figures. While figures are clear in visualizing the prediction sets, I am not convinced if real-data experiments support the proposed method's claim. Figure 3 on synthetic data show multimodality, but real data do not look like they have multimodality, which also align with other studies that use the same datasets.

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

This paper proposes DISCO, a diffusion-based conformal prediction framework for multivariate time-series forecasting. Instead of running reverse diffusion to evaluate candidate future outcomes, the method aggregates denoising errors across multiple noise levels into a scalar nonconformity score. The score is calibrated using chronologically split conformal prediction, with a fixed common-randomness bank to ensure consistent score evaluation. The paper considers a direct formulation (Algorithm 1) and a residual-based formulation (Algorithm 2), allowing potentially non-convex and disconnected prediction regions.

The authors provide finite-sample approximate marginal and localized conditional coverage bounds under strict stationarity and β-mixing assumptions. Experiments on Gaussian and multimodal VAR processes, as well as the Traffic and Solar datasets, compare the proposed approach against several conformal prediction baselines and include ablations of the score-construction components. The synthetic multimodal results are promising, while the real-data results reveal a notable trade-off between prediction-region volume and empirical coverage.

Critical strengths

1. Meaningful modeling of multimodal predictive uncertainty.

The use of diffusion denoising errors to define flexible prediction regions is an appealing approach to multivariate uncertainty quantification. In particular, allowing disconnected regions is well motivated for multimodal conditional distributions. The synthetic experiments, including the separated Gaussian-mixture setting and the visualizations in Figure 1, provide encouraging evidence that the method can represent geometry that is difficult to capture with conventional connected regions.

2. Useful theoretical and empirical scope.

The paper explicitly addresses temporal dependence through approximate finite-sample marginal and localized conditional coverage bounds under stationary β-mixing assumptions. The comparison with multiple relevant baselines and the ablations of stratification, antithetic sampling, and common random numbers are also useful for understanding the proposed construction.

Critical weaknesses

1. The real-data volume improvements are difficult to interpret without controlling for achieved coverage.

In Table 1, Algorithm 1 often produces relatively small prediction regions but falls below the nominal coverage level. For example, on Traffic with response dimension 2 at the 95% nominal level, Algorithm 1 with SMLD achieves 92.4% mean coverage and a volume of 0.455, whereas Algorithm 2 with OU achieves 96.8% coverage and a volume of 0.865. These operating points do not establish a like-for-like efficiency comparison. Some Algorithm 2 results are more encouraging at comparable coverage, but the evidence would be clearer if prediction-region volume were systematically assessed against achieved coverage. Furthermore, the temporal-dependence sensitivity experiment in Figure 5 reports region volumes without the corresponding empirical coverage, leaving it unclear whether the behavior under stronger dependence reflects changes in region efficiency, coverage, or both.

2. The specific methodological contribution relative to DiffConf is not sufficiently differentiated.

The related-work discussion identifies different application goals and assumptions for DiffConf and DISCO, but it does not clearly explain the technical difference between their diffusion-based nonconformity-score constructions. A more explicit distinction between the proposed multi-noise-level denoising-error score and the score used in DiffConf is important for assessing the methodological contribution beyond applying diffusion-based conformal inference to a different problem setting.

Questions for authors

1. Coverage–volume trade-off on real data. Can the authors provide coverage–volume comparisons for the main DISCO variants and strong baselines in representative Traffic and Solar settings? Evaluating several predefined nominal coverage levels, without selecting operating points using test outcomes, would help establish whether the reported volume advantages persist at comparable achieved coverage.

2. Distinction from DiffConf. How does the proposed nonconformity score, which aggregates denoising errors across noise levels, differ technically from DiffConf's score construction? Which aspects of the score itself constitute the methodological contribution, beyond the different application settings and theoretical assumptions?

3. Coverage under stronger temporal dependence. Could the authors report the empirical coverage corresponding to each setting in Figure 5? In particular, do Algorithms 1 and 2 continue to attain coverage close to the nominal 95% and 90% levels as the spectral radius of the VAR process increases?

3 comments

Thank you for the thoughtful review! The updated Table 1 shows the Algorithm 1 and 2 can achieve 90% target with efficient volume. The updated five-subset Traffic results in Appendix B.6 show that Algorithm 1 undercovers at the 95% target, while Algorithm 2 attains above-target mean coverage on both Traffic and Solar. Algorithm 1 remains effective on Solar. We have narrowed our efficiency claims accordingly, added coverage to the dependence-sensitivity experiment, and explained the technical distinction from DiffConf below. We address every point one by one below.

Coverage-volume trade-off on Traffic and Solar

We agree that the original operating points did not establish matched-coverage efficiency. We have updated Table 1 in Section 5, highlighted in blue, to report the 90% results and gray out entries below nominal mean coverage. On Traffic, Algorithms 1 and 2 achieve over 90%. And algorithm 1 can achieve the smaller volume in both dimensions: 0.258 ± 0.184 at 90.7 ± 6.4% coverage and 0.209 ± 0.176 at 92.2 ± 1.3 % coverage. On Solar, Algorithm 1 achieves the smallest qualifying mean volumes in both dimensions: 0.141 ± 0.033 at 90.5 ± 1.2% coverage and 0.106 ± 0.063 at 92.2 ± 1.1% coverage. Values report mean ± standard deviation across subsets. We report coverage and volume at both 90% and 95% nominal levels in Table 1 and Appendix B.6, respectively. At the 95% target, Algorithm 1 undercovers on Traffic, while Algorithm 2 achieves 96.7-97.2% mean coverage. Algorithm 2 (DDPM) has the smallest qualifying mean volume for dy=2, and Algorithm 2 (SMLD) has the second-smallest for dy=4, behind CONTRA-LOO. On Solar, Algorithm 1 achieves the smallest qualifying mean volumes in both dimensions: 0.399 ± 0.123 at 95.9 ± 0.8% coverage and 0.955 ± 0.327 at 97.5 ± 0.6% coverage. Values report mean ± standard deviation across subsets. We report coverage and volume at both 90% and 95% nominal levels in Table 1 and Appendix B.6. These results show that DISCO can produce competitive prediction-region volumes while attaining nominal mean coverage.

Technical distinction from DiffConf

DiffConf’s score has the form ∣y−f^(x)∣exp⁡{λΔθ(x,t)}|y-\widehat{f}(x)|\exp\{\lambda\Delta_\theta(x,t)\}. The diffusion discrepancy depends on the covariates and recent reference information. At a fixed context, this factor rescales the residual. DISCO instead perturbs each candidate response yy and evaluates its noise-prediction error. Its score can vary nonmonotonically with yy, and this gives us allowance for disconnected sublevel sets. Both methods use multiple noise levels, and this alone is not our claimed distinction. We will clarify this work in the related work section.

Coverage under stronger temporal dependence

We have added empirical coverage alongside volume in Appendix E. For Algorithms 1 and 2 with the OU backbone, mean coverage remains close to the nominal 90% and 95% levels across spectral radii {0,0.1,0.3,0.5,0.7,0.9}. At stronger dependence, Algorithm 1’s normalized volume increases while coverage remains approximately stable; Algorithm 2’s volume is more stable.

The revised Table 1 reports nominal-or-higher mean empirical coverage for 23 of 24 proposed-method configurations at the 90% target. The remaining configuration, Algorithm 1–DDPM on Solar with dy=2d_y=2, achieves 89.8% and is grayed out. Algorithm 2 meets or exceeds the target in all 12 configurations, and all bolded volume entries meet nominal mean coverage.

We’ve added a coverage-volume plot to make these comparisons easier to inspect. The plot shows that our methods achieve smaller mean prediction-region volumes than several baselines at comparable or higher mean coverage, placing these results toward the bottom right. We have added this figure as Figure 4 in the appendix (page 27).

I appreciate authors for conducting experiments, and clarifying details. The authors have added the requested plots and empirical results and clarified the distinction from DiffConf. While these revisions improve the transparency of the empirical claims, they do not fully resolve the practical limitations of the proposed method, such as the coverage instability on real-world data. Consequently, the methodological contribution remains moderate. My overall assessment remains unchanged.

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

This paper studies conformal prediction for multivariate time series, proposing a diffusion-based nonconformity score for constructing flexible prediction regions and analyzing approximate coverage under temporal dependence. The idea is interesting and relevant to the field. However, the connection between the theoretical analysis and the practical method needs clearer explanation. In particular, although the method avoids reverse-diffusion sampling, it still uses Monte Carlo perturbations to compute its score. The authors should clarify how the approximation bound in Theorem 4.1 informs the sampling-budget choice and affects the resulting prediction regions. More importantly, Table 1 shows that the proposed methods often produce smaller prediction regions without attaining the target coverage. Since smaller regions may simply reflect undercoverage, these results do not establish superior efficiency over the baselines. Comparisons at matched coverage levels are needed to substantiate the claimed advantages.

Critical strengths
  1. The paper proposes an interesting diffusion-based nonconformity score for constructing multimodal prediction regions without running reverse-diffusion trajectories. Its treatment of sampling randomness is thoughtful, combining a shared random bank with stratified diffusion times and antithetic perturbations. The theoretical analysis explicitly considers finite-budget score approximation and temporal dependence, while the experiments compare multiple diffusion formulations and investigate key sampling choices through ablations. The synthetic examples provide a clear illustration of the method’s ability to represent disconnected prediction regions.
Critical weaknesses
  1. The claimed volume efficiency is difficult to assess because the proposed methods frequently fail to attain the target coverage in Table 1. Smaller prediction regions can arise simply from undercoverage and therefore do not, by themselves, demonstrate improved efficiency. Although the theoretical analysis allows coverage penalties under temporal dependence, this does not resolve the empirical comparison. The authors should compare prediction-region volumes at matched coverage levels and report uncertainty in the coverage estimates to establish whether the efficiency gains persist after accounting for the coverage shortfall.
  2. The algorithm incorporates the sampling scheme analyzed in Theorem 4.1 through a fixed common-random-number bank, stratified diffusion times, and antithetic perturbations. However, the practical implications of the pointwise approximation bound remain unclear. The authors should explain how the sampling budget is selected and quantify how bank randomness affects calibration thresholds, coverage, and prediction-region volume. They should also distinguish population-score approximation from the coverage guarantee obtained by calibrating the finite-bank score.
Questions for authors

1, In Table 1, the proposed method with the smallest volume does not achieve the desired coverage, so it can not prove the effectiveness. 2, What is Algorithm 2 in Table 1, it should be clearly defined or explained in the main paper.

3 comments

Thank you for the thoughtful review! The updated 90% comparison in the Table 1 and the updated 95% comparison in Appendix B.6, highlighted in blue, includes all five Traffic location subsets. At the nominal level 90%, Algorithm 1 and 2 show competitive prediction-region volumes, and at the nominal level 95%, Algorithm 1 undercovers on Traffic, whereas Algorithm 2 attains above-target mean coverage on both datasets. Algorithm 1 remains effective on Solar. We have also moved Algorithm 2 into Section 3.3 and clarified the distinction between finite-budget score approximation and coverage. We address every point one by one below.

Undercoverage and prediction-region efficiency

We agree that the original operating points did not establish matched-coverage efficiency. We have updated Table 1 in Section 5, highlighted in blue, to report the 90% results and gray out entries below nominal mean coverage. On Traffic, Algorithms 1 and 2 achieve over 90%. And algorithm 1 can achieve the smaller volume in both dimensions: 0.258 ± 0.184 at 90.7 ± 6.4% coverage and 0.209 ± 0.176 at 92.2 ± 1.3 % coverage. On Solar, Algorithm 1 achieves the smallest qualifying mean volumes in both dimensions: 0.141 ± 0.033 at 90.5 ± 1.2% coverage and 0.106 ± 0.063 at 92.2 ± 1.1% coverage. Values report mean ± standard deviation across subsets. We report coverage and volume at both 90% and 95% nominal levels in Table 1 and Appendix B.6, respectively. At the 95% target, Algorithm 1 undercovers on Traffic, while Algorithm 2 achieves 96.7-97.2% mean coverage. Algorithm 2 (DDPM) has the smallest qualifying mean volume for dy=2, and Algorithm 2 (SMLD) has the second-smallest for dy=4, behind CONTRA-LOO. On Solar, Algorithm 1 achieves the smallest qualifying mean volumes in both dimensions: 0.399 ± 0.123 at 95.9 ± 0.8% coverage and 0.955 ± 0.327 at 97.5 ± 0.6% coverage. We report coverage and volume at both 90% and 95% nominal levels in Table 1 and Appendix B.6. These results show that DISCO can produce competitive prediction-region volumes while attaining nominal mean coverage.

Sampling budget, bank randomness, and Theorem 4.1

Theorem 4.1 bounds the pointwise Monte Carlo approximation error relative to the fitted denoiser’s population score. It motivates allocating samples across diffusion-time strata, but does not directly bound prediction-region volume or identify a universally optimal budget. Appendix D reports synthetic and real-data experiments with a fixed budget of 256 perturbations. Score variability and region volume improve substantially as the number of time strata increases from 4 to approximately 16, with smaller changes thereafter. This supports an empirical plateau rather than a proven optimum. Note that the coverage theorem applies to the finite-bank score itself, with the fitted model and bank frozen before calibration and reused at test time. It does not require convergence to the population or oracle score. In the revised manuscript, we specify the sampling budget used in each experiment and include ablations of common-random-number bank reuse, diffusion-time stratification, and antithetic sampling. We also report additional experiments examining robustness to increasing temporal dependence. We do not claim a separate analysis of bank-to-bank calibration-threshold variability or a universally optimal sampling budget.

Definition of Algorithm 2

We have moved Algorithm 2 and its explanation into Section 3.3, highlighted in blue. It fits a deterministic predictor, trains a diffusion model on training residuals, and calibrates the residual score on held-out observations. Prediction regions are translated by the point forecast.

Thank you for the detailed response and the additional coverage results. However, achieving the target coverage is a fundamental requirement in conformal prediction. Prediction-set volumes should therefore be compared at comparable coverage levels, as smaller volumes obtained through undercoverage do not demonstrate improved efficiency. The reported results indicate that the proposed method falls short of the target coverage in approximately half of the experiments, including those with the best volume results highlighted in bold. These findings leave my concerns unresolved, and I will maintain my original score.

Thank you for your follow-up! The revised Table 1 reports nominal-or-higher mean empirical coverage for 23 of 24 proposed-method configurations at the 90% target. The remaining configuration, Algorithm 1–DDPM on Solar with dy=2d_y=2, achieves 89.8% and is grayed out. Algorithm 2 meets or exceeds the target in all 12 configurations, and all bolded volume entries meet nominal mean coverage.

We’ve added a coverage-volume plot to make these comparisons easier to inspect. The plot shows that our methods achieve smaller mean prediction-region volumes than several baselines at comparable or higher mean coverage, placing these results toward the bottom right. We have added this figure as Figure 4 in the appendix (page 27).

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful feedback. We address each question and concern below and summarize the main revisions to the manuscript:

  • Coverage and efficiency: We addressed concerns about coverage-volume comparisons by reporting results at both 90% and 95% nominal levels, identifying undercoverage, and qualifying our efficiency claims.
  • Theoretical clarity: We clarified the learning assumptions, finite-bank coverage guarantees, and scope of localized conditional coverage.
  • Method presentation: We moved Algorithm 2 into the main text and explained the technical distinction between DisCo and DiffConf in our response.
  • Additional evaluation: We added optimal-transport baselines, updated ablations, empirical coverage under stronger temporal dependence, real-data prediction-region visualizations, and approximate training-time estimates.

Please let us know if there are any questions, happy to discuss them!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.