acceptodds
Under review as a conference paper at ICLR 2027

Quantile-Guided Diffusion for Probabilistic Time Series Forecasting

Abstract

Quantile forecasts summarize uncertainty at each future time step, while diffusion forecasters represent it through sampled future trajectories. We study how predicted quantiles can serve as distributional controls for diffusion forecasting. Specifically, we investigate how the mechanism of quantile integration affects its effectiveness in guiding a pre-trained diffusion forecaster. We find that residual control throughout the denoising network yields the most consistent improvements. We instantiate this approach as Control-TS, which learns a quantile-control pathway while keeping the pretrained backbone frozen. Across eight benchmarks, Control-TS consistently improves the backbone’s CRPS, achieves competitive performance against state-of-the-art forecasting models, and retains its gains in capacity-matched comparisons. Further analyses examine how these gains depend on quantile granularity, predictor choice, and forecast quality. Finally, we introduce the Conditional Linear Predictive Score (CLPS), which complements distributional metrics by evaluating the downstream forecasting utility of generated futures. Together, these findings support predicted quantiles as an effective interface between explicit uncertainty estimates and diffusion-based trajectory generation.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic Time Series Forecasting, Diffusion Models, Quantile Guidance, Conditional Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_T9XaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the use of externally predicted quantiles to guide diffusion-based forecasting. The author proposes Control-TS, which introduces a trainable residual control pathway into a frozen pretrained diffusion backbone. Experiments shows that Control-TS outperforms to baselines like TSDiff, with additional analyses of model capacity, quantile granularity, robustness, and computational overhead.

Critical strengths

S1: Clear motivation and good writing. The paper identifies meaningful connections between explicit quantile forecasts and diffusion-based generation. The Comparison of different integration strategies provides useful empirical insights into how external predictive information can guide diffusion models.

S2: Extensive evaluation. The paper use eight benchmarks to evaluate the performance of Control-TS. The capacity-matched comparisons, additional backbone experiments, and robustness analyses strengthen the empirical evidence.

Critical weaknesses

W1: Limited evaluation under multivariate and long-horizon forecasting. The main experiments focus on univariate forecasting with relatively short prediction horizons (24–48). Although additional multivariate experiments are provided using MG-TSD, improvements are observed on only five of eight datasets. The lack of evaluation on widely adopted benchmarks such as ETT under longer prediction horizons limits the evidence for the method's broader applicability.

W2: Insufficient isolation of the benefits of quantile-specific uncertainty information. Although the experiments suggest that multiple quantiles can improve forecasting performance, increasing the number of quantiles also changes the dimensionality of the control input. Therefore, it remains unclear how much of the improvement comes specifically from uncertainty information rather than the additional input representation. Given that the proposed residual controller are similar to ControlNet-style architectures, stronger evidence isolating the quantile-specific contribution would help establish the methodological novelty.

Questions for authors

1.Quantile information: Can the authors further clarify whether the performance gains come from the additional uncertainty information provided by multiple quantiles, rather than simply using more input features? 2.Generalization: Could the authors provide additional experiments on standard long-term forecasting benchmarks, such as ETT? 3.CLPS evaluation: What additional insights does CLPS provide compared with conventional forecasting metrics? Could the authors further clarify its advantages and limitations?

1 comment

We thank the reviewer for recognizing the motivation, clarity, and breadth of our evaluation. We appreciate the suggestions on isolating uncertainty information and extending evaluation to longer horizons.

Comment: Regarding uncertainty information versus additional input features.
Response: The submitted observed-context ablation partially addresses this concern by repeating history across one, three, or five channels with the same controller architecture (Section 3.3 and Appendix B.11). We agree that further isolation requires controls based on predicted futures.

We trained controllers using three or five repetitions of the predicted median, matching the input widths of Q3 and Q5. This holds the number of conditioning channels fixed while replacing the additional quantiles with the same point forecast. We also evaluated mean-variance controls to compare predicted quantile trajectories with a summary of predictive location and scale.

The table reports the mean dataset-level CRPS reduction (%) of Control-TS over each new alternative, across all eight datasets. Positive values indicate improvement by Control-TS.

Alternative controlControl-TS Q3 improvement (%)Control-TS Q5 improvement (%)
Channel-matched repeated median+5.22+13.17
Mean-variance control+4.33+7.72

Control-TS outperforms both alternatives on average for Q3 and Q5. The channel-matched repeated-median comparison supports the value of uncertainty information beyond simply increasing the number of conditioning channels, while the mean-variance comparison further supports the usefulness of the quantile representation. We will include the results and precise control constructions in the revised paper.

Comment: Regarding long-horizon and multivariate forecasting.
Response: We appreciate this suggestion. We extended the evaluation to multivariate forecasting on ETTh1 and ETTm2 at horizons of 96 and 192. The table reports the relative CRPS reduction (%) achieved by Control-TS over the MG-TSD backbone, positive values indicate improvement.

DatasetHorizonControl-TS reduction over MG-TSD backbone (%)
ETTh196+31.10
ETTh1192+8.89
ETTm296+15.94
ETTm2192+9.29

Control-TS improves CRPS in all four settings, by 8.89-31.10%, providing additional evidence of its applicability to longer-horizon multivariate forecasting. Results average three sampling seeds per fixed trained checkpoint, with 100 forecast paths per test window.

This probabilistic evaluation is computationally demanding: each test window requires generating 100 complete multivariate trajectories through iterative diffusion. We therefore extended the original setting to horizons of 96 and 192 within the available budget. We will include the full protocol and distinguish these additional multivariate results from the original univariate evaluation in the revised paper.

Comment: Regarding the additional insights, advantages, and limitations of CLPS.
Response: Thank you for this question. CLPS measures transferable forecasting utility: can a separate predictor learn useful context-future relationships from generated training targets and generalize to real observations? (Section 3.5 and Appendix B.16).

This provides information beyond directly scoring the generator’s forecasts. Indeed, Control-TS improves direct forecasting CRPS on all eight datasets, but improves CLPS on six, with small regressions on Electricity and UberTLC (Table 4). Thus, better forecast accuracy does not necessarily translate into better training targets for this downstream learner.

The advantages are inexpensive, deterministic, closed-form ridge fitting without backpropagation, reproducibility across hardware up to floating-point differences, and a controlled comparison with fixed training contexts and evaluator. Table 4 additionally includes DeepAR and TFT evaluators to examine whether the findings extend beyond a linear predictor.

The limitation is that CLPS measures utility accessible to the chosen evaluator, rather than complete distributional fidelity. Its deterministic predictions make the reported CRPS equivalent to normalized absolute error. Consequently, CLPS does not directly establish uncertainty calibration or cross-horizon dependence. Its additional information comes from the train-on-generated-targets, test-on-real-data protocol, not from a different error formula. We will further clarify this scope in the main text.

We appreciate the reviewer’s feedback and hope our clarifications and additional evidence resolve their concerns. If so, we hope that the reviewer would consider raising their score.

Reviewer_S3ojOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how externally predicted quantiles can guide diffusion models for probabilistic time-series forecasting. It proposes Control-TS, which injects quantile-conditioned residuals throughout a frozen diffusion denoising network via a trainable control pathway. The paper’s main contribution is to use explicit quantile trajectories as distributional controls for implicit trajectory generation and to systematically compare alternative integration mechanisms. Experiments on eight time-series benchmarks show that Control-TS improves the TSDiff backbone on all eight datasets. Among the four integration strategies, full-depth residual control is the most consistent design, achieving the best or tied-best result on five datasets.

Critical strengths

1. The paper introduces an effective interface between explicit uncertainty estimates and implicit generative forecasting. Control-TS uses predicted quantile trajectories as distributional controls and injects quantile-conditioned residuals throughout a frozen diffusion backbone. As shown in Table 1, full-depth residual control is the only integration mechanism that improves the TSDiff backbone on all eight datasets. This is important because it provides a practical and empirically consistent way to combine explicit probabilistic forecasts with diffusion-based trajectory generation.

2. The empirical evaluation is systematic and includes several meaningful controls. The paper compares four integration strategies, evaluates Control-TS on eight benchmarks, and includes capacity-matched baselines, observed-context controls, architectural ablations, degraded-control experiments, and experiments with additional diffusion backbones. Table 2 shows that Control-TS improves upon TSDiff on all eight datasets, with an approximately 35% reduction in CRPS on M4-Hourly. It also outperforms the capacity-matched TSDiff-Wide baseline on six datasets. These results provide evidence that the gains are not solely attributable to increased model capacity.

Critical weaknesses

The evaluation provides strong evidence that Control-TS improves horizon-wise marginal forecast quality, but it provides limited evidence about the joint temporal quality of the generated trajectories. The primary metric, CRPS, is computed separately for the predictive distribution at each future timestep and therefore does not directly evaluate temporal dependence across the forecast horizon. CLPS provides useful complementary evidence about downstream forecasting utility, but it does not directly resolve this issue: its default evaluator is a deterministic linear predictor whose output is replicated across forecast samples, while the additional DeepAR and TFT probes improve over TSDiff on only four and six datasets, respectively. Consequently, it remains unclear whether Control-TS improves the coherence and dependence structure of complete sampled trajectories, rather than primarily improving their timestep-wise marginal distributions. This limits the evidence for one of the paper’s main motivations for retaining a diffusion-based trajectory generator.

Questions for authors
  1. Could the authors provide an evaluation of the joint temporal quality of the generated trajectories? Results using an appropriate multivariate or path-level scoring rule, together with measurements of temporal correlations or spectral properties, would address my main concern and could increase my rating.

  2. Table 12 shows that each fixed DeepAR/TFT configuration improves TSDiff on at most seven of the eight datasets, whereas the headline 8/8 result relies on dataset-specific validation selection between the two external forecasters. Could the authors clarify the total model-selection budget for Control-TS and whether the competing baselines received a comparable dataset-specific tuning budget? Please also discuss how the main conclusion should be interpreted under a single pre-specified external forecaster.

4 comments

We thank the reviewer for recognizing the practical value of our approach and the breadth of our empirical controls. We appreciate the suggestions on joint trajectory evaluation and the inquiry about the model-selection procedure. We address each point below.

Comment: Regarding joint trajectory quality beyond marginal CRPS.
Response: We evaluated Control-TS against its exact frozen TSDiff backbone using energy and variogram scores, together with temporal correlation and spectral diagnostics.

The table reports relative reductions (%) over the TSDiff backbone, with positive values indicating improvement.

DatasetEnergy scoreVariogram scoreTemporal correlation errorSpectral error
Electricity+2.63+19.63+0.07+2.07
Exchange+8.55+4.82+1.43+1.04
KDDCup+4.56+14.55+3.59−0.71
M4-Hourly+26.85+24.21+12.01+6.96
Solar+4.25+15.30+7.44+8.33
Traffic−1.68+0.36+7.59+5.92
UberTLC+1.77+4.41+1.90+2.36
Wikipedia+2.31+6.20+3.08+2.29
Mean relative improvement+6.16+11.19+4.64+3.53

Results average three sampling seeds per fixed trained checkpoint. The final row is the unweighted mean of the eight dataset-level percentage reductions.

Control-TS improves energy score on seven of eight datasets, variogram score and temporal correlation error on all eight, and spectral error on seven of eight. These results provide direct evidence of improved temporal trajectory quality beyond marginal CRPS.

Compared with standalone DeepAR, the mean relative reductions across all eight datasets are approximately 12.64% in energy score, 24.97% in variogram score, 20.21% in temporal correlation error, and 29.07% in spectral error.

As a supplementary check, we reassigned sampled values between forecast paths independently at each future timestep. This preserves each timestep’s predictive distribution while disrupting the original temporal dependence. Marginal CRPS remains unchanged, while variogram, temporal correlation and spectral errors increase on all eight datasets. This supports the usefulness of temporal structure that marginal CRPS does not assess. We will include these results in full and the shuffling analysis in the final version.

Comment: Regarding the model-selection procedure and baseline tuning.
Response: For Control-TS, the configuration grid comprises six forecaster - quantile combinations per dataset, namely DeepAR/TFT × Q1/Q3/Q5, selected using validation CRPS (Section 3 and Appendix B.10). The remaining controller settings follow Appendix B.4. Controller training runs for up to 1,000 epochs, with validation every 50 epochs using 16 forecast paths, or four for Traffic/Wikipedia (Appendices B.3–B.4). These six combinations describe the forecaster-quantile grid, with external-forecaster fitting and checkpoint selection forming additional components of the procedure.

The competing baselines also received dataset-specific validation tuning. Using the authors’ implementations, we evaluated K²VAE across four configurations and CN-Diff across two hidden widths, selecting hyperparameters and checkpoints by validation CRPS (Appendix B.8). Both used the same data splits, context/horizon lengths, rolling test windows, scaling, and 100 test forecast paths as Control-TS. Other baselines likewise used model-specific configurations under the shared evaluation protocol.

The submitted paper reports all six Control-TS configurations, making the effect of forecaster and quantile selection explicit. Selection follows validation performance: for example, on KDDCup, validation selects TFT-Q3 even though DeepAR-Q5 achieves a lower test CRPS. The fixed-configuration results below show broad gains without dataset-specific selection.

Comment: Regarding performance with a single fixed external forecaster.
Response: The 8/8 result describes the validation-selected system. Broad improvements also remain when both the external forecaster and quantile set are fixed across datasets. The submitted Table 12 gives the following numbers of strict improvements over TSDiff.

Fixed external forecasterQ1Q3Q5
DeepAR5/87/85/8
TFT5/87/86/8

Counts use the reported precision and exclude ties.

Every fixed configuration improves on at least five datasets. Both DeepAR-Q3 and TFT-Q3 improve on seven of eight, with Electricity and Traffic as their respective exceptions. Thus, broad improvements remain without switching forecasters or quantile sets between datasets. Validation selection increases consistency to 8/8, and we will make this distinction explicit in the main text.

We thank the reviewer again and hope our responses and additional results address their concerns. If so, we respectfully ask the reviewer to consider raising their score.

Reviewer_S3ojOpenReview ↗

Response to Authors

I thank the authors for the detailed response and the additional experiments. The rebuttal substantially addresses my primary concern regarding the joint temporal quality of the generated trajectories.

In particular, the newly reported energy and variogram scores, temporal-correlation error, and spectral error provide evidence beyond timestep-wise marginal CRPS. The path-shuffling experiment is especially informative: it preserves the marginal predictive distribution at each timestep while disrupting temporal dependence, leaving CRPS unchanged but degrading the dependence-sensitive diagnostics on all eight datasets. This directly demonstrates that these evaluations capture aspects of trajectory quality that marginal CRPS does not assess. Overall, Control-TS improves the variogram score and temporal-correlation error on all eight datasets, and improves the energy score and spectral error on seven of eight datasets. I therefore consider my main concern about the absence of joint trajectory evaluation to be largely addressed.

For completeness, the final paper should clearly define all newly introduced metrics and their aggregation procedures, including the variogram order and weighting scheme and the precise definitions of the temporal-correlation and spectral errors. Reporting absolute scores together with variability across sampling seeds would also make the evidence easier to assess. I also encourage the authors to phrase the conclusion carefully, since the energy score on Traffic and the spectral error on KDDCup do not improve.

The response also clarifies the model-selection procedure. I appreciate the explicit statement that the headline 8/8 result corresponds to validation-based selection among six forecaster–quantile configurations per dataset. The fixed-configuration results provide a useful qualification: DeepAR-Q3 and TFT-Q3 each improve upon TSDiff on seven of eight datasets, whereas dataset-specific validation selection increases the consistency to eight of eight. This distinction should be made explicit in the main paper.

My remaining concern is that the comparability of the model-selection budgets is not yet fully established. The response reports six configurations for Control-TS, four for K²VAE, and two widths for CN-Diff, but it does not provide the total number of training runs, external-forecaster configurations, checkpoint candidates, random seeds, or computational budget for each method. The statement that the other baselines received model-specific tuning is helpful, but it does not by itself demonstrate an equivalent tuning opportunity. I therefore view this part of my concern as only partially addressed. A compact summary of the search and selection budget for every method would improve transparency.

Overall, the rebuttal provides substantial new evidence for the joint temporal quality of Control-TS and resolves my primary technical concern. The remaining issue concerns the transparency and comparability of the model-selection budget rather than the central validity of the proposed method. Based on the additional evidence, I am inclined to keep my positive score.

Thank you for your thoughtful follow-up and for recognizing that the additional experiments substantially address the main technical question. We will incorporate all your suggested additions and qualifications in the final paper.

Regarding the remaining point, we further clarify the complete search and selection accounting below. For the competing methods, we followed their published implementations and empirically tested training settings recommended by their original authors. These choices aim to support strong performance and effective convergence while respecting each model’s training requirements. CN-Diff and K²VAE do not use auxiliary external forecasters, so external-forecaster configurations do not apply to these methods. For all methods, each evaluated configuration was trained once. The three seeds reported in Table 2 were used for inference.

The following table reports the training schedules used in our experiments, following the released implementations of each method.

MethodEpochs × batches/epoch = updatesMax. checkpointsValidation
TSDiff1,000 × 128 = 128,00020Every 50 epochs
CN-Diff≤100 × 100 = ≤10,000100Every epoch
K²VAE50 × 100 = 5,00050Every epoch
SSSD1,200–2,000 × 128 = 153,600–256,000 updates24 - 40Every 50 epochs
TSFlow-Uncond.1,000 × 128 = 128,00020Every 50 epochs
TSFlow-Cond. (OU)400 × 128 = 51,20020Every 20 epochs
Control-TS≤1,000 × 128 = ≤128,00020Every 50 epochs

More training updates do not necessarily imply longer training time, since the computational cost of each update depends on the model and its implementation.

For Control-TS, the configuration grid, training and validation schedule, and computational profiling are reported in Appendices B.4, B.10, and B.15. Its budget above covers the adaptation stage.

The final paper will consolidate the accounting for all methods into one table covering candidate configurations, total training runs, checkpoint-selection opportunities, and the roles of training and sampling seeds. For Control-TS, this will cover backbone pretraining, external-forecaster fitting, and adaptation.

We hope these clarifications and the consolidated reporting resolve your remaining point. We will be happy to answer any further questions.

Reviewer_S3ojOpenReview ↗

Response to Authors

Thanks for your follow-up replies. I keep my original score towards acceptance.

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

This paper investigates the use of externally predicted quantiles as distributional controls for diffusion-based time-series forecasting. The authors systematically compare several integration mechanisms and propose Control-TS, which injects quantile-conditioned residuals into a frozen pretrained diffusion backbone. Experiments across eight benchmarks demonstrate consistent CRPS improvements over TSDiff, with additional evaluations on CSDI and MG-TSD. The paper also introduces the Conditional Linear Predictive Score (CLPS) to assess the downstream forecasting utility of generated trajectories.

Critical strengths
  1. Clear motivation and systematic design. The paper establishes a meaningful connection between explicit quantile forecasting and implicit diffusion-based trajectory generation. Comparing output ensembling, input concatenation, feature modulation, and residual control provides useful insights into how predictive uncertainty should be incorporated.

  2. Strong empirical performance. Control-TS consistently improves TSDiff across all eight benchmarks and achieves the best or tied-best CRPS on five datasets. Comparisons with TSDiff-Wide and alternative diffusion forecasters provide evidence that the gains are not solely attributable to increased parameter capacity.

  3. Extensive ablations and robustness analyses. The experiments investigate controller depth, initialization, quantile granularity, external predictor selection, and degraded control signals. The additional CSDI and MG-TSD evaluations provide further evidence regarding architectural transferability.

  4. Practical and interpretable integration. Freezing the pretrained diffusion backbone while learning a separate control pathway offers a modular adaptation strategy. The computational analysis also provides useful information about the accuracy-efficiency tradeoff.

Critical weaknesses
  1. Limited architectural novelty. The residual control mechanism closely follows established ControlNet-style conditioning and adapter-based architectures. The main contribution is their application to quantile-guided time-series diffusion rather than a fundamentally new generative modeling mechanism. The distinction from existing conditional diffusion approaches could be developed more explicitly.

  2. Insufficient isolation of distributional information. Although the paper compares Q1, Q3, and Q5, increasing the number of quantiles also changes the control input dimensionality. The observed-context control is informative but does not isolate the benefit of uncertainty information over point forecasts. A channel-matched repeated-median control would help establish whether improvements specifically arise from predictive dispersion and asymmetry.

  3. Incomplete evaluation of joint trajectory quality. CRPS evaluates marginal predictive distributions at individual forecasting horizons but does not directly assess temporal dependence. Although CLPS provides a complementary evaluation, Appendix B.16 uses a deterministic ridge predictor with identical replicated forecast samples. Consequently, its CRPS reduces to absolute prediction error, limiting its ability to assess the joint distribution of generated trajectories. Additional dependence-sensitive metrics would strengthen the evaluation.

  4. Limited robustness and additional computational cost. The improvements are not uniform across architectures: quantile control improves MG-TSD on only five of eight datasets. Appendix B.13 also shows substantial degradation under inaccurate controls, particularly on Exchange and M4-Hourly. Furthermore, the controller nearly matches the backbone in parameter count, and inference latency increases by 17--61%. The practical benefit under unreliable quantiles and constrained computational budgets remains insufficiently established.

Questions for authors
  1. Distributional contribution: Can the authors compare Q3/Q5 against channel-matched repeated-median controls, as well as controls constructed from predicted means and variances? This would clarify whether the gains arise from richer uncertainty information rather than additional conditioning channels.

  2. Joint distribution quality: Could the authors evaluate generated trajectories using dependence-sensitive metrics, such as energy scores, variogram scores, or temporal correlation errors? How well does quantile guidance preserve cross-horizon dependence?

  3. CLPS interpretation: Since the linear evaluator produces deterministic forecasts, its CRPS is equivalent to absolute error. What additional information does CLPS provide beyond evaluating a ridge predictor trained on synthetic targets? Could the authors include a real-target training reference and explicitly quantify the utility gap?

  4. Robustness and efficiency: Can the authors quantify performance as a function of external quantile quality and investigate adaptive control gating? Additionally, how does Control-TS compare with simpler conditioning methods under matched training and inference budgets?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ETCq - Part 1/2

We thank the reviewer for recognizing the motivation, empirical performance, and extensive evaluation. We address the concerns below.

Comment: Regarding the distinction from existing conditional diffusion approaches.
Response: The distinction concerns the conditioning information and adaptation setting. Control-TS supplies horizon-aligned quantile predictions from a separate forecaster, conveying predictive location, dispersion, and asymmetry, and learns to incorporate them into an already trained diffusion forecaster while keeping its backbone frozen. Our study examines the value of these distributional summaries and compares mechanisms for integrating them. We will further emphasize this distinction from existing conditional diffusion approaches in the main text and clarify the scope of our contribution relative to the established residual-conditioning architecture.

Comment: Regarding uncertainty information versus additional conditioning channels.
Response: The submitted observed-context ablation partially addresses this concern by repeating history at matched channel widths (Section 3.3 and Appendix B.11). We agree that further isolation requires controls based on predicted futures.

We trained controllers using three or five repetitions of the predicted median, matching the input widths of Q3 and Q5. We also evaluated mean-variance controls to compare predicted quantile trajectories with a summary of predictive location and scale.

The table reports mean dataset-level CRPS reductions (%) of Control-TS’s submitted Q3/Q5 results (Table 2) over each new alternative across eight datasets. Positive values favor Control-TS.

Alternative controlControl-TS Q3 improvement (%)Control-TS Q5 improvement (%)
Channel-matched repeated median+5.22+13.17
Mean-variance control+4.33+7.72

Control-TS outperforms both alternatives on average for Q3 and Q5. The channel-matched repeated-median comparison supports the value of uncertainty information beyond simply increasing the number of conditioning channels, while the mean-variance comparison further supports the usefulness of the quantile representation. We will include the results and precise control constructions in the revised paper.

Comment: Regarding joint trajectory quality and cross-horizon dependence.
Response: We evaluated Control-TS against its exact frozen TSDiff backbone using energy and variogram scores, together with temporal correlation and spectral diagnostics.

The table reports relative reductions (%) over the TSDiff backbone, with positive values indicating improvement.

DatasetEnergy scoreVariogram scoreTemporal correlation errorSpectral error
Electricity+2.63+19.63+0.07+2.07
Exchange+8.55+4.82+1.43+1.04
KDDCup+4.56+14.55+3.59−0.71
M4-Hourly+26.85+24.21+12.01+6.96
Solar+4.25+15.30+7.44+8.33
Traffic−1.68+0.36+7.59+5.92
UberTLC+1.77+4.41+1.90+2.36
Wikipedia+2.31+6.20+3.08+2.29
Mean relative improvement+6.16+11.19+4.64+3.53

Results average three sampling seeds per fixed checkpoint, with 100 paths per window, the final row averages dataset-level reductions.

Control-TS improves energy score on 7/8 datasets, variogram score and temporal correlation error on 8/8, and spectral error on 7/8. Compared with standalone DeepAR, the corresponding mean reductions are 12.64%, 24.97%, 20.21%, and 29.07%.

Independently permuting sample identities at each future timestep within each test window preserves marginal CRPS but increases variogram, temporal correlation, and spectral errors on all eight datasets. This demonstrates sensitivity to temporal structure beyond marginal CRPS.

Our responses on CLPS, adaptive gating, and matched-budget comparisons continue in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer ETCq - Part 2/2

Continuing our response to Reviewer ETCq from Part 1/2.

Comment: Regarding CLPS interpretation and a real-target training reference.
Response: CLPS measures transferable forecasting utility through the train-on-generated-targets, test-on-real-data protocol. Its deterministic ridge predictions make CRPS equivalent to normalized absolute error (Section 3.5 and Appendix B.16).

Below is the real-target reference alongside Control-TS. Lower scores are better, the gap is Control-TS minus real, calculated from displayed values.

DatasetReal targetsControl-TS targetsUtility gap
Electricity0.081660.07175−0.00991
Exchange0.010570.02847+0.01790
KDDCup0.551170.69878+0.14761
M4-Hourly0.038870.03809−0.00078
Solar0.555970.52649−0.02948
Traffic0.183740.15438−0.02936
UberTLC0.238410.30611+0.06770
Wikipedia0.897620.37364−0.52398

Control-TS targets improve over TSDiff targets on 6/8 datasets and the real-target reference on 5/8. The real-target reference is not an oracle for this finite-sample linear learner, since synthetic targets may provide useful regularization or denoising. These results characterize utility for the shared evaluator. CLPS does not establish joint distributional fidelity, which we assess separately through the trajectory evaluation in point 3. We will include the real-target reference and clarify this interpretation in the revised paper.

Comment: Regarding unreliable controls and adaptive gating.
Response: We thank the reviewer for his thoughtful suggestion. The submitted corruption experiments cover all eight datasets (Section 3.4 and Appendix B.13). We additionally tested an inference-only gating heuristic on Exchange without retraining.

The gate attenuates controller residuals when the median fluctuates excessively relative to history, using a clean-validation threshold without test targets. Gaussian corruption scales with the predicted 10–90% spread and preserves quantile ordering. Preliminary results use two paired sampling seeds and 100 paths per window.

Noise multiplierCRPS reduction from gating (%)
0.00−1.12
0.25−1.77
0.50+22.37
1.00+52.95

On Exchange, this simple inference-only gate reduces CRPS by 22.37% and 52.95% under strong corruption, without retraining. Attenuating controller residuals increases reliance on the pretrained backbone. The 1.12% and 1.77% degradations under clean and mild corruption are consistent with weakening still-useful guidance. Adaptive gating is already identified as future work in the paper (Section 6), these results motivate developing gates that better preserve useful controls while suppressing corrupted ones.

Comment: Regarding simpler conditioning methods under matched budgets.
Response: The comparisons directly test whether simpler conditioning can match Control-TS under the same training-update and sampling budgets. Concatenation, FiLM, and one-/two-block residual control share Control-TS’s predictive inputs, 128,000 updates (batch size 64), and 100 diffusion steps/100 paths, ensembling requires no training.

Control-TS improves TSDiff on 8/8 datasets, versus 4/8 for concatenation, FiLM, and ensembling, and 3/8 for each one-/two-block residual variant (Table 1). Thus, neither simpler conditioning interfaces nor restricting residual injection to the first one or two blocks achieves the same consistency.

We appreciate the reviewer’s constructive suggestions and hope our responses and additional evidence address their concerns. If so, we would be grateful if the reviewer would consider raising their score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.