acceptodds
Under review as a conference paper at ICLR 2027

GeNCoup: Cross-Channel Coupling for Probabilistic Forecast Trajectories

Abstract

Probabilistic time-series forecasters generate ensembles of future trajectories to represent predictive uncertainty. For these ensembles to form coherent multivariate scenarios, they require accurate marginals and realistic dependence across channels over time. Channel-independent generative forecasters already capture within-channel temporal dynamics, but leave cross-channel coupling unspecified. Existing plug-and-play methods recover this coupling post hoc, often at the expense of the temporal dynamics already encoded in the trajectories. We introduce Generative Noise Coupling (GeNCoup), which injects cross-channel dependence directly into the sampling process of a frozen generative forecaster. Instead of drawing channel noise independently, GeNCoup jointly samples the noise across channels at each noise feature using a cross-channel covariance (R), while preserving each channel's original noise distribution. This introduces cross-channel dependence without changing the forecaster's marginal predictions or within-channel temporal dynamics. Building on this framework, we introduce GeNCoup-Corr, which derives the coupling from empirical cross-channel correlations, and GeNCoup-Dyn, which learns it by optimizing a joint forecast score. Across six forecasters from four generative families, eight long-horizon datasets, and four prediction horizons, GeNCoup outperforms baselines in joint forecast quality while preserving both marginal accuracy and temporal dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Probabilistic Forecasting, Multivariate Time Series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9CvpOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of learning cross-channel dependencies in generative probabilistic time-series forecasters. The authors illustrate the importance of modeling cross-channel coupling through conceptual examples and data analysis of two correlated time series from the multivariate ETTh1 dataset.

The proposed method, named Generative Noise Coupling (GeNCoup), draws channel-wise noise using a cross-channel symmetric and positive semi-definite covariance matrix R, instead of drawing channel-wise noise values independently. To have R preserve the original noise distribution of each channel, the authors show that R must have a unit diagonal. By the definition of the original channel-independent sampling and the coupled seed, this method also preserves the temporal dependence within each channel. The authors propose two variants that differ in how R is set: (1) GeNCoup-CORR which sets R based on a linear interpolation between global and local cross-channel correlations and (2) GeNCoup-DYN which learns the coupling directly, while keeping the probabilistic forecaster frozen. Rather than sampling the noise directly from a gaussian distribution with a covariance matrix R, GeNCoup-DYN multiplies the original noise trajectory with a learnable matrix A, which is equivalent to sampling with R=AATR=AA^T. This is shown to guarantee symmetry, positive semidefinitenes and a unit diagonal of R by construction.

GeNCoup is different from post-hoc methods, which operate on generated forecast ensembles and from quantile methods which predict intervals per time point. Compared to dependence-reconstruction methods, which recover joint information and preserve marginals, GeNCoup also preserves temporal dependence and is integrated into the forecaster. GeNCoup-DYN is also E2E minimizable.

The authors evaluate GeNCoup on eight standard time series forecasting datasets and compare it to six representative generative probabilistic time series forecasters from four generative families (including recent baselines from 2025-2026) and five post-hoc methods. GeNCoup achieves the best variogram score on seven of eight datasets and the best cross-channel energy score on six of eight datasets, while better preserving CRPS/NMAE and temporal dynamics compared to post-hoc baselines. Main evaluation results are reported across three seeds. The authors further perform ablations and analyses related to key design choices in GeNCoup-CORR. They show that using a validation-optimized alpha value to blend local and global cross-correlations (a*) is better than using global only (a=0) or local only (a=1) alternatives, on average. In addition, to evaluate the choice of setting R to the empirical cross-channel correlation matrix of the observed data the authors evaluate the cross-channel correlation of the recovered source noise and compare it against the empirical correlation matrix.

Critical strengths
  1. A novel method that offers conceptual advantages over existing baselines, addressing a well-motivated problem.
  2. State-of-the-art results in terms of cross-channel performance compared to contemporary representative generative baselines and post-thoc methods, on multiple standard time series benchmarks. Results are reported across three seeds, method coverage appears sufficient, and the detailed tables underlying the aggregations presented in the main text are provided in the appendix. The authors provide ablations and empirical analyses to evaluate key design choices.
Critical weaknesses

The authors aim to preserve marginal behavior while injecting cross-channel dependence. However, for generative forecasters such as COBASE, the differences in CRPS and N-MAE (which measure marginal performance) are 0.000 on average, while for GeNCoup they are 0.012. While the authors interpret this as preserving the marginals, the practical implications of a difference of this magnitude are unclear and should be discussed. This is an important aspect since the authors set it as one the requirements of their method.

Questions for authors

Could the authors clarify their criterion for "marginal preservation" and discuss how their results compare to generative baselines that achieve lower marginal degradation (e.g., COBASE)?

3 comments

We thank the reviewer for their thoughtful and valuable feedback.

Q:

(W1) The authors aim to preserve marginal behavior while injecting cross-channel dependence. ...

(Q1) Could the authors clarify their criterion for "marginal preservation" and discuss how their results compare to generative baselines that achieve lower marginal degradation (e.g., COBASE)?

A:

To assess whether the observed marginal differences are practically meaningful, we introduce an empirical preservation criterion based on the uncoupled forecaster's sampling noise across random seeds. This allows us to assess marginal preservation by comparing the observed changes against the expected sampling variability, rather than relying solely on their absolute magnitude. We have formalized this criterion in the Experiments section of the revised manuscript.

Preservation criterion. For each metric, we first average the uncoupled forecaster's scores across experimental settings for each seed, then compute the standard deviation across seeds. Since our results report relative percentage changes rather than absolute scores, we express the standard deviation as a percentage of the mean across seeds, yielding a directly comparable noise threshold. Changes within this threshold are treated as consistent with empirical preservation. These thresholds are reported in the Noise row of the results tables and apply to both marginal (CRPS, NMAE) and temporal (ES-Time) metrics.

Comparison with post-hoc baselines. The zero CRPS and NMAE changes reported for post-hoc reordering methods such as COBASE are guaranteed by construction. These methods use exactly the same generated values, changing only their assignment across scenarios. Since CRPS and NMAE are computed marginally for each channel and time step, they are invariant to scenario ordering. Consequently, their marginal score differences are exactly zero. In contrast, GeNCoup modifies the sampling process itself, generating new trajectories rather than reordering existing ones. Therefore, we expect marginal score changes to remain within the uncoupled forecaster's sampling noise threshold, rather than being exactly zero.

Results. Both GeNCoup variants satisfy our empirical preservation criterion, with aggregate CRPS, NMAE, and ES-Time changes all falling within their respective noise thresholds (approximately 0.02%). Notably, GeNCoup is the only method whose aggregate ES-Time changes remain within the 0.025% noise threshold, compared with 0.076–3.047% for post-hoc baselines. Thus, GeNCoup improves cross-channel dependence while empirically preserving both marginal accuracy and temporal dynamics.

In addition, we would like to note that, following Reviewer 3xqd's observations, we revised two aspects of our experimental setup. First, to address a potential information leakage risk, we revised the estimation of R_local to use each window's observed lookback independently, rather than pooling observations across evaluation minibatches. This eliminates the identified leakage risk and ensures that predictions are independent of batch size. Second, in response to the reviewer's concerns about the fidelity of our COBASE implementation, we revised it to best match the original method. We reran all experiments and analyses and updated the manuscript accordingly ( previous results are marked in red, while the updated results are marked in blue). The updated results generally preserve the relative rankings across the tables.

I thank the authors for addressing my initial concerns and appreciate their effort in revising the paper and re-running experiments. Given the identified leakage risk and potential fidelity issues with the COBASE implementation, could the authors elaborate on their claim that "the updated results generally preserve the relative rankings across the tables"? Comparing the red and blue numbers across the tables is hard to parse and summarize into new conclusions, so a summary of the main trends and specific discrepancies (especially where they may impact the original conclusions) would strengthen confidence in the updated result and the submitted revision.

AuthorsOpenReview ↗

Response to Reviewer 9Cvp

We thank the reviewer for their follow-up. By stating that the updated results generally preserve the relative rankings across the tables, we were primarily referring to the "1st count" row in Table 2, which counts the number of datasets on which each method achieves the best score among all compared methods. Before the corrections, the two GeNCoup variants achieved the best VS on 7/8 datasets and the best ES-Chan on 6/8 datasets. Following the corrections, the VS count remains unchanged at 7/8, while the ES-Chan count decreases to 5/8. This difference is driven by ETTh2, where Schaake now achieves the best ES-Chan score.

Regarding the quantitative changes, GeNCoup-Corr is only marginally affected. Averaged across the 8 datasets, its VS improvement over the frozen forecaster changed from −10.71% to −10.43%, while its ES-Chan improvement changed from −1.57% to −1.49%. These correspond to reductions of only 0.28 and 0.07 percentage points, respectively, in the average improvements over the frozen forecaster. The largest changes are concentrated in ETTh2 and ETTm2, while the remaining datasets exhibit generally smaller differences. GeNCoup-Dyn remains entirely unaffected. For COBASE, the average improvements over the frozen forecaster decreased by 0.50 and 0.23 percentage points in VS and ES-Chan, respectively.

Overall, the main experimental conclusions remain unchanged, with GeNCoup continuing to achieve the strongest overall cross-channel performance among the compared methods.

To facilitate comparison, we provide below the full per-dataset breakdown of GeNCoup-Corr and COBASE before and after the corrections, corresponding to Table 2 of the main paper. Values represent percentage changes relative to the uncoupled frozen forecaster (negative is better), while differences are expressed in percentage points.

Method, metricETTh1ETTh2ETTm1ETTm2WeatherECLTrafficSolarMean
Corr VS prev−11.994−4.395−11.750−3.722−7.723−0.878−1.597−43.582−10.705
Corr VS now−11.905−3.524−11.741−2.958−7.647−0.844−1.559−43.236−10.427
Corr VS diff+0.089+0.871+0.009+0.764+0.076+0.034+0.038+0.346+0.278
Corr ES-Chan prev−1.404−0.937−1.436−0.841−1.561−0.509+0.141−5.970−1.565
Corr ES-Chan now−1.358−0.741−1.351−0.707−1.465−0.481+0.095−5.944−1.494
Corr ES-Chan diff+0.046+0.196+0.085+0.134+0.096+0.028−0.046+0.026+0.071
COBASE VS prev−11.057−3.837−10.844−4.069−6.033−1.459−1.468−31.215−8.748
COBASE VS now−10.255−4.340−9.860−4.030−4.799−1.026−1.113−30.575−8.250
COBASE VS diff+0.802−0.503+0.984+0.039+1.234+0.433+0.355+0.640+0.498
COBASE ES-Chan prev−1.337−0.848−1.378−0.970−1.160−0.446−0.848−5.356−1.543
COBASE ES-Chan now−1.267−0.768−1.234−0.792−0.781−0.225−0.540−4.862−1.309
COBASE ES-Chan diff+0.070+0.080+0.144+0.178+0.379+0.221+0.308+0.494+0.234

We also reran the affected analyses where necessary, and their conclusions remain unchanged. All additional rebuttal experiments used the corrected GeNCoup-Corr implementation.

Reviewer_3xQDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Channel-independent generative forecasters capture within-channel temporal dynamics well but do not model cross-channel dependencies. There are existing methods that attempt to recover cross-channel couplings, but these are often unable to do so without deteriorating the temporal dynamics of the trajectories. The authors introduce GeNCoup, a plug-and-play method that injects cross-channel dependencies into the sampling of frozen forecasters. Through this approach, the source noise is sampled jointly across channels using a correlation matrix whose unit diagonal leaves each channel's own noise distribution unchanged. This introduces cross-channel correlations without deteriorating the forecaster's marginal accuracy or the temporal dynamics inside channels. GeNCoup has two variants: GeNCoup-Corr and GeNCoup-Dyn, where the former gets its cross-channel dependencies from lookback correlations and the latter learns its coupling by optimizing for the variogram score (VS). The authors validate their method using six forecasters and eight datasets, comparing both variants with existing post-hoc methods. They show that for most forecasters and tasks, both variants improve the VS and ES-Chan metrics (cross-channel dependency metrics) while leaving CRPS, NMAE, and ES-Time (marginal and temporal dynamics metrics) essentially unchanged.

Critical strengths
  1. By design, the method's correlation matrix (R) has a unit diagonal. This ensures that only the dependence between channels changes and that within-channel distributions stay the same. So the guarantee that marginals and within-channel temporal dynamics are unaffected holds by construction, which is a clear advantage over post-hoc approaches. This is exactly the problem the authors set out to solve, since existing methods improve the cross-channel dependencies but at the cost of the temporal dynamics. The authors also show this in their experiments, where CRPS, NMAE, and ES-Time only change by about 0.01% for both variants, while the post-hoc baselines degrade ES-Time by up to a few percent.

  2. The method is simple and cheap to run, especially GeNCoup-Corr, which only adds one matrix multiplication on the noise. GeNCoup-Dyn adds a training step on top of this to optimize the correlation matrix for VS. It works on every forecaster that meets the clearly defined requirements (independently drawing each channel's noise from a Gaussian source and deterministic sampling), and is especially useful for channel-independent foundation models such as Sundial and Aurora, which cannot easily be retrained. This allows users to add cross-channel dependencies to a model they cannot or do not want to retrain, at almost no extra cost, and without losing the marginal accuracy and temporal dynamics from point 1. It should be noted, however, that the gains on these two foundation models are the smallest among all forecasters.

Critical weaknesses
  1. The authors motivate their method with claims that joint dependence across channels through time is important for simulations and risk evaluations in systems like energy, weather, and finance. However, they only evaluate their method for introducing these cross-channel dependencies using joint scores, and by showing that marginal accuracies and within-channel temporal dynamics do not change. The authors do not show any downstream task results that would indicate that their method improves these forecasters in such practical applications. Since GeNCoup leaves the marginal accuracy unchanged by design, these applications are the only place where its practical benefit could actually show. Additionally, there is no comparison between their method and the alternative of training a natively joint model, such as TACTiS-2 (which the authors cite in their related work). Such a comparison could demonstrate if GeNCoup is a viable method to achieve joint dependence across channels, or if training a joint model is a better trade-off.

  2. The current evaluations of GeNCoup make it unclear how much of its advantage comes from the method itself, because the comparison favours it. Both GeNCoup-Corr and GeNCoup-Dyn are fitted on the main metric (VS), while the baselines are not. Large VS gains can thus partly be expected. ES-Chan, the cross-channel metric they were not fitted on, shows only small and mixed gains (around 1% on average, with Dyn being best on only 2 of 8 datasets compared to 5 of 8 for VS, and Traffic getting worse). The authors argue that ES is less sensitive to dependence than VS, which is fair, but it does mean the main results rest mostly on the metric the method was optimized for. Additionally, GeNCoup-Corr also gets local correlations, while the baselines only get global ones. COBASE even had its original rolling refit removed (Appendix A.8). The balance between local and global correlations (α) is also chosen on VS. So while the method likely works, the size of its advantage over the baselines is uncertain.

Minor: The paper has no limitations section. Its requirements (Gaussian source noise, deterministic sampling, independent channel noise), the static linear correlation it uses, and the weak results on the widest datasets would fit there.

Questions for authors
  1. Section 4.2 states that R_local is estimated "from the observed lookbacks in the current minibatch", while Section 6 refers to it as a "window-specific estimate", and Appendix B.3 describes it as estimated "from the current lookback window". Appendix B.2.3 (lines 1103-1105) describes the correlations in the inversion analysis as computed "matching the construction used by GeNCoup-Corr: all test windows are pooled along the temporal dimension", which suggests R_local might also be estimated from pooled test windows. How exactly is R_local computed: per window, per minibatch, or over the full test set? And what is the batch size? Since the test windows are taken with a stride of 96 and a lookback of 96, the lookbacks of later windows contain (parts of) the futures of earlier ones. Can R_local for a given window therefore include data from after its forecast origin? If so, the GeNCoup-Corr results for α > 0 would include information that is not available in practice.

  2. Appendix A.7 describes the optimizer for GeNCoup-Dyn, but not on which split A is trained, for how many steps, with what batch size, or whether a separate A is learned per horizon. Could the authors clarify this? Since the training objective uses the true futures, it is important to confirm that these come only from the training split.

  3. The method is presented as plug-and-play, but no runtime or memory costs are reported, apart from the GPU requirements in Appendix A.1. Could the authors report the overhead of GeNCoup-Corr and the training cost of GeNCoup-Dyn compared to the uncoupled sampler?

  4. On ECL and Traffic, the datasets with the most channels, both variants show little to no improvement in VS, and ES-Chan gets slightly worse on Traffic. What do the authors think caused this? For example, for GeNCoup-Corr, is R_local reliably estimated when the number of channels is large relative to the number of observations it is estimated from? And for GeNCoup-Dyn, is learning A harder at this scale (e.g. 862 × 862 for Traffic)?

My assessment depends mostly on the answers to questions 1 and 2. If R_local and the training of A use no information from after the forecast origin, I would view the soundness of the results more positively.

6 comments

We thank the reviewer for their insightful comments and thoughtful questions, which helped us clarify and strengthen the revised manuscript.

W1

Q:

The authors motivate their method with claims that joint dependence across channels through time is important for simulations and risk evaluations in systems like energy, weather, and finance. ...

A:

We thank the reviewer for raising these points.

Our work focuses specifically on the probabilistic forecasting task, aiming to generate more realistic joint trajectories by introducing cross-channel dependence while preserving marginal accuracy and within-channel temporal dynamics. Accordingly, our evaluation focuses on the quality of the resulting predictive distributions rather than their performance in downstream applications. We recognize that downstream evaluations would provide further evidence of the practical benefits of improved cross-channel dependence. However, given the limited rebuttal period, we were unable to construct a suitable downstream simulation framework with appropriate data and evaluation protocols that would allow us to assess these benefits reliably**.**

Regarding the comparison with natively joint models such as TACTiS-2, we would like to emphasize that GeNCoup serves a fundamentally different purpose. Our goal is not to replace joint forecasting architectures, but to introduce cross-channel dependence into existing frozen generative forecasters that do not explicitly model it, while leveraging their already strong marginal and temporal forecasting capabilities.

This distinction is central to our motivation. GeNCoup is a plug-and-play mechanism applicable to any generative forecaster satisfying the three requirements outlined in our paper, without modifying or fine-tuning the underlying forecaster. Consequently, it can benefit directly from advances in marginal and temporal forecasting, including recent foundation models, without requiring the development of a new joint forecasting model.

Moreover, GeNCoup-Corr can operate without learned parameters or target-domain training, using the configuration (\alpha=1) and estimating the required correlations from the available lookback observations. This enables its application to pretrained foundation models in a fully zero-shot setting, introducing cross-channel dependence at inference time without fine-tuning the forecaster or learning an additional coupling model.

Therefore, a direct comparison with TACTiS-2 would not isolate GeNCoup's contribution, as the results would also reflect differences in the underlying forecasting architectures and training regimes. We view these approaches as complementary rather than competing: natively joint models learn cross-channel dependencies during training, whereas GeNCoup introduces these dependencies into existing pretrained forecasters without fine-tuning or modifying their parameters. This allows GeNCoup to leverage the strong forecasting capabilities of existing models while benefiting directly from future advances in probabilistic forecasting, including foundation models.


W2

Q:

The current evaluations of GeNCoup make it unclear how much of its advantage comes from the method itself, because the comparison favours it. ...

A:

We thank the reviewer for raising these concerns. We would first like to clarify an important distinction between the two GeNCoup variants. While GeNCoup-Dyn is trained by directly minimizing VS, GeNCoup-Corr does not optimize VS or any other scoring function. Its coupling matrix is constructed directly from empirical correlations, without loss minimization. Its only connection to VS optimization is that the mixing coefficient α\alpha is selected on the validation split. However, the contribution of this selection is modest. As shown in Table 3, even without the fusion mechanism or local correlations (alpha=0), GeNCoup-Corr achieves a 10.11% average VS reduction, compared with 10.25% using the validation-selected alpha. Thus, validation-based selection contributes only 0.14 percentage points, indicating that nearly all of GeNCoup-Corr’s VS improvement is obtained without directly optimizing VS.

Baseline fairness. We thank the reviewer for highlighting this issue. We acknowledge that fitting COBASE’s copula only once on the training set did not faithfully reproduce its original rolling-refit procedure. We have therefore revised the implementation to best match the original method, refitting a separate copula for every forecast window using its preceding (L=96) observations. These are the same local observations available to GeNCoup-Corr when estimating its local correlation matrix. Further implementation details are provided in the paper.

We reran all experiments with this refined COBASE implementation to ensure a fair comparison and updated the results throughout the main paper and appendix (previous results are marked in red, while the updated results are marked in blue). As we believe the revised implementation best reflects the original method, we intend to retain the updated results unless the reviewer suggests otherwise.


W3

Q:

Minor: The paper has no limitations section. ...

A:

We have added a dedicated Limitations section to the revised version of the paper.


Q1

Q:

Section 4.2 states that R_local is estimated "from the observed lookbacks in the current minibatch", while Section 6 refers to it as a "window-specific estimate", and Appendix B.3 describes it as estimated "from the current lookback window". ...

A:

We thank the reviewer for the careful reading of our manuscript and for identifying this important issue.

In the originally reported experiments, (R_{\text{local}}) was estimated by pooling the lookback observations across all windows in the current minibatch, rather than computing a separate correlation matrix for each window. The motivation was to obtain a more robust correlation estimate by leveraging additional observations and reducing the noise associated with individual lookback windows.

However, we agree with the reviewer's observation that this implementation introduces a potential risk of information leakage. Since test windows are extracted from the same time series using a sliding-window procedure, the lookback of a later window may contain observations that fall after the forecast origin of an earlier window in the same minibatch. Consequently, pooling lookbacks across a minibatch could inadvertently incorporate information unavailable at the time of prediction, particularly with larger batch sizes or datasets containing relatively few evaluation windows. We acknowledge that this issue was overlooked in our original implementation and sincerely thank the reviewer for bringing it to our attention.

To eliminate this risk, we have rerun all experiments with R_local estimated independently for each evaluation window, using only its own observed lookback. No information from other windows in the minibatch is used. This ensures that each correlation estimate is based exclusively on information available at the corresponding forecast origin, fully eliminating the identified leakage mechanism. Averaged across the 8 datasets, Corr's VS improvement over the frozen forecaster changed from −10.71% in the previous revision to −10.43% in the current revision, while its ES-Chan improvement changed from −1.57% to −1.49%. This corresponds to reductions of only 0.28 and 0.08 percentage points, respectively, in the average improvements relative to the frozen forecaster. We have updated the revised manuscript to include the corrected results, revised all affected analyses (including the ablation study and Appendices B.2 and B.3), and ensured that the descriptions are consistent throughout the paper.

Q2

Q:

Appendix A.7 describes the optimizer for GeNCoup-Dyn, but not on which split A is trained, for how many steps, with what batch size, or whether a separate A is learned per horizon. ...

A:

We thank the reviewer for raising this question. GeNCoup-Dyn is trained exclusively on the training split, with the training objective using only ground-truth futures from that split. GeNCoup-Dyn learns a single coupling matrix (A), which is independent of the lookback window and applied to all forecasting windows to introduce cross-channel dependence. A separate matrix is trained for each combination of forecaster, dataset, and prediction horizon, using the same training hyperparameters, including learning rate and number of epochs. The learned matrix is evaluated on the validation and test splits, with checkpoint selection based on validation performance. No ground-truth futures from the validation or test splits are used during training. We have also expanded Appendix A.7 in the revised manuscript to provide additional details on the training procedure.


Q3

Q:

The method is presented as plug-and-play, but no runtime or memory costs are reported, apart from the GPU requirements in Appendix A.1. ...

A:

We have added a new Memory and Compute subsection in Appendix A.8, reporting the parameters, inference, and memory overhead of GeNCoup.

Parameter overhead. Both variants introduce a single dense coupling matrix with about (C^2) parameters, where (C) is the number of channels. While the relative increase in parameters is negligible for FMPatch, DiffPatch, Sundial, and Aurora, it is more noticeable for compact models such as PPM and K2VAE. Nevertheless, the additional storage ranges from only 196 B for ETT to 2.83 MiB for Traffic, remaining modest relative to modern computational resources.

Inference overhead. The coupling introduces a similar number of additional FLOPs across models, with minor differences arising from the model-specific noise dimension (d). Consequently, the relative latency overhead depends primarily on the backbone's own sampling cost. For FMPatch, DiffPatch, Sundial, and Aurora, the measured increase in latency is less than 0.2%. The overhead is more noticeable for the lightweight samplers PPM and K2VAE. For the largest dataset, Traffic, the coupling adds approximately 35% latency for both variants on PPM, and 26% for GeNCoup-Corr and 10% for GeNCoup-Dyn on K2VAE. Across all measured settings, from ETT to Traffic, the added inference latency ranges from approximately 0.004 to 4.8 ms per window.

Training and memory overhead. Training GeNCoup-Dyn requires backpropagation through the frozen forecaster's sampling procedure, particularly for iterative diffusion and flow-matching models. We mitigate the resulting activation-memory overhead using gradient checkpointing. We measured peak GPU memory during Dyn optimization and compared it against gradient-free coupled sampling and VS evaluation. Across ETT and Weather, the median training-to-inference peak memory ratio is 1.1–1.2× (0.43-2 GiB) for K2VAE, PPM, Sundial, and Aurora, and approximately 2.5× (6.2 GiB) for FMPatch and DiffPatch. For Solar, ECL, and Traffic, the median ratios range from 2.1× to 4.1× (4.16-32.2 GiB) across models.

Reviewer_6YhMOpenReview ↗
Rating
1(−2)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeNCoup, which introduces cross-channel dependence by coupling the source noise of a frozen generative forecaster while preserving each channel’s predictive trajectory distribution. The authors introduce an empirical correlation-based variant, GeNCoup-Corr, and a learned variant, GeNCoup-Dyn, evaluated across six forecasters, eight datasets, and four horizons. The approach is practical and well motivated, although its expressiveness, evaluation scope, and computational costs require further analysis

Critical strengths

Clear motivation and aligned design. Introducing dependence at the noise level avoids disrupting temporal structure through timestamp-wise sample reordering, making the approach useful for existing channel-independent generative forecasters.

Well-founded preservation guarantee. Under channel-wise deterministic mappings with randomness fully specified by the source noise, a unit-diagonal correlation matrix preserves each channel’s noise distribution and therefore its complete trajectory distribution, including temporal dependence.

Broad experimental coverage. Experiments cover multiple generative families and foundation models, with several post-processing baselines. Separate marginal, temporal, and cross-channel evaluations, together with pairwise diagnostics and noise-inversion experiments, help characterize the method’s effects.

Critical weaknesses

Limited coupling expressiveness. All noise coordinates share the same cross-channel correlation matrix, and Dyn does not predict an input-dependent matrix. Although nonlinear decoding can induce complex output dependence, the ability to capture changing relationships, cross-channel lagged dependence, and joint extreme events remains insufficiently established.

Joint evaluation focuses on contemporaneous dependence. VS and ES-Chan evaluate channels jointly at individual forecast times, while ES-Time evaluates each channel separately. These metrics do not fully assess cross-channel, cross-time relationships. Lagged-dependence diagnostics or probabilistic evaluation of multivariate path-dependent quantities would strengthen the evidence.

The additional benefit of learned coupling is inconsistent. Corr achieves larger average VS improvements than Dyn on PPM and FMPatch. On k2VAE and Sundial, Dyn performs better in VS but worse in ES-Chan, while both variants worsen ES-Chan on Aurora. The authors should clarify when learned coupling is preferable to empirical correlation and whether these limitations arise from optimization or the coupling parameterization.

Local correlation estimation depends on inference batches. Corr estimates its local matrix using lookback windows from the current minibatch, so the same input may receive different cross-channel predictive distributions depending on batch composition. Although future labels are not used, this raises deployment-consistency concerns. Comparisons with per-window estimation and sensitivity analyses for batch size and composition are needed.

Computational costs of learned coupling are insufficiently characterized. Dyn requires backpropagation through the entire sampling procedure despite freezing the forecaster. This may incur substantial training and memory costs for iterative generators and high-dimensional dense coupling matrices. Runtime and memory measurements, together with low-rank or sparse alternatives, would help establish practical scalability.

Questions for authors

see weakness

5 comments

We thank the reviewer for their careful reading and valuable feedback, and address each of their points below.

W1

Q:

Limited coupling expressiveness.
All noise coordinates share the same cross-channel correlation matrix, and Dyn does not predict an input-dependent matrix. ...

A:

We thank the reviewer for this observation. In GeNCoup, the same cross-channel correlation matrix is applied across all noise coordinates by design, as we intentionally preserve the independence between noise coordinates. This allows us to introduce cross-channel dependence while preserving the original forecaster's temporal modeling capabilities.

Sampling noise dependently across time steps would alter the temporal dependencies produced by the pretrained forecaster, which we explicitly aim to preserve. Exploring more expressive noise sampling strategies that jointly improve cross-channel and temporal dependencies is an interesting direction for future work, but beyond the scope of this study.


W2

Q:

Joint evaluation focuses on contemporaneous dependence.
VS and ES-Chan evaluate channels jointly at individual forecast times, while ES-Time evaluates each channel separately. ...

A:

We thank the reviewer for this suggestion. We focused on these evaluation metrics because they best highlight GeNCoup's main objective: improving cross-channel dependence while preserving the original forecaster's temporal structure. Specifically, VS and ES-Chan evaluate the cross-channel dependence targeted by GeNCoup, whereas ES-Time assesses the preservation of within-channel temporal structure.

We agree that these metrics do not directly evaluate lagged cross-channel dependence. To further investigate GeNCoup's effectiveness on datasets with substantial lagged dependencies, we conducted additional experiments using MixBench-TS (Abdelmalak et al., 2026), a recently proposed benchmark specifically designed to evaluate cross-channel modeling. According to this study, MixBench-TS datasets exhibit stronger lagged cross-channel dependencies than the standard datasets commonly used in forecasting research. Based on the analysis in Table 5 of their paper, we selected three datasets representing different levels of lagged dependence: CzeLan (11 channels), ZafNoo (11 channels), and HouseholdPower (7 channels), exhibiting high, moderate, and lower levels of lagged dependence, respectively.

We evaluated GeNCoup-Corr with (\alpha=1) on top of the pretrained Sundial foundation model. This configuration uses only the empirical correlation estimated from each input's lookback window. We used a lookback length of (L=96) and 4 standard prediction horizons. The following table reports the relative changes in VS and ES-Chan compared to the uncoupled Sundial baseline. Negative values indicate improvements.

DatasetHorizonVS (%)ES-Chan (%)
CzeLan96−8.27−0.81
192−7.55−1.11
336−6.48−0.76
720−6.35−0.24
Avg.−7.16−0.73
ZafNoo96−1.37+0.13
192−1.14+0.29
336−1.06+0.28
720−1.03+0.44
Avg.−1.15+0.29
HouseholdPower96−4.71−0.41
192−4.79−0.34
336−4.88−0.36
672−4.83−0.36
Avg.−4.80−0.37

GeNCoup-Corr improved VS across all three datasets and every prediction horizon, with the largest average reduction on CzeLan (7.16%). ES-Chan improves on two datasets and remains nearly unchanged on ZafNoo (+0.29%).

These results suggest that GeNCoup can improve cross-channel modeling even on datasets with lagged dependencies. We acknowledge that explicit lagged-dependence diagnostics and evaluation of multivariate path-dependent quantities could provide further insights and remain valuable directions for future work.

Reference. Abdelmalak, I., Putzke, M., Choi, J., Hanika, T., Yalavarthi, V. K., & Schmidt-Thieme, L. (2026). MixBench-TS: A Multivariate Time Series Forecasting Benchmark Where Channel Mixing Pays Off. arXiv:2609.32656. https://arxiv.org/abs/2609.32656.

W3

Q:

The additional benefit of learned coupling is inconsistent.
Corr achieves larger average VS improvements than Dyn on PPM and FMPatch. ...

A:

We thank the reviewer for raising this point. We would like to highlight that both variants achieve comparable VS improvements on FMPatch, while the most substantial difference appears on PPM. However, this gap is largely driven by Solar, where Corr achieves a VS improvement of −43.48%, compared to −23.07% for Dyn. Excluding Solar, the average VS improvements change from −10.89% and −7.95% to −5.47% and −5.44%, respectively, indicating nearly identical performance on the remaining datasets.

We believe these performance differences can be explained by two main factors.

First, the methods differ in their inductive biases and adaptability. GeNCoup-Corr constructs its coupling matrix from empirical correlations, which our analysis "What the Seed Would Have to Carry" identifies as an effective choice for the coupling structure. In contrast, GeNCoup-Dyn is initialized with the identity matrix (no cross-channel correlation) and learns its coupling matrix directly by minimizing VS, without an explicit correlation-based inductive bias. Consequently, the quality of the learned coupling matrix depends more heavily on the training process and its optimization. Furthermore, Dyn learns a single static coupling matrix shared across all forecasting windows, whereas Corr adapts its coupling using the empirical correlations of each window's lookback. Our ablation study shows that fusing the global correlation matrix (estimated from the training set) with local correlations computed from each window's lookback, using an alpha selected on the validation set, yields a modest but consistent improvement in average VS reduction, from −10.11% with the global matrix alone to −10.25% with the fusion mechanism.

Second, optimization may contribute to the remaining differences. In our experiments, GeNCoup-Dyn uses a fixed learning rate of 0.05 across all forecasters and datasets, without a dedicated hyperparameter search. Consequently, given its identity initialization and lack of a correlation-based inductive bias, Dyn may be more sensitive to optimization choices and potentially converge to suboptimal local minima. More extensive hyperparameter tuning could therefore improve its results.

We recommend selecting between GeNCoup-Corr and GeNCoup-Dyn based on validation performance. Across all forecaster–dataset configurations, the variant achieving the better validation VS also achieves the better test VS in 85.3% of cases. This selection accuracy increases with the difference in validation VS improvements, reaching 90.3%, 92.1%, and 93.0% when the difference exceeds 0.5, 1, and 2, respectively. Furthermore, when comparing the two variants at the dataset level, as reported in the main results table, validation-based selection identifies the better test variant on 7 of 8 datasets. The only exception is ETTh1, where the validation improvements are nearly identical (−6.925% for Dyn versus −6.931% for Corr). These results suggest that validation performance is a reliable predictor of test performance when selecting between the two variants.

W4

Q:

Local correlation estimation depends on inference batches.
Corr estimates its local matrix using lookback windows from the current minibatch, so the same input may receive different cross-channel predictive distributions depending on batch composition. ...

A:

As pointed out by reviewer 3xqd, our original implementation introduced a potential risk of information leakage, since (R_{\text{local}}) was estimated by pooling lookback observations across windows in the same evaluation minibatch. Since these windows originate from the same test split, the lookback of a later window may contain observations unavailable at the forecast origin of an earlier window. This risk is particularly relevant for large batch sizes or datasets with relatively few evaluation windows.

Following this observation, we have rerun all experiments and analyses using per-window correlation estimation. Specifically, (R_{\text{local}}) is now computed independently from each window's own observed lookback, without incorporating observations from other windows in the minibatch. This eliminates the identified leakage risk and ensures that the predictive distribution for a given input is independent of batch size. The manuscript has been updated accordingly, with previous table captions marked in red and new ones in blue.

In addition, we would like to note that, following Reviewer 3xqd's concerns regarding the fidelity of our COBASE implementation, we revised the implementation to best match the original method and reran the experiments. The updated results generally preserve the relative rankings across the tables.

W5

Q:

Computational costs of learned coupling are insufficiently characterized.
Dyn requires backpropagation through the entire sampling procedure despite freezing the forecaster. ...

A:

We have added a new Memory and Compute subsection in Appendix A.8, reporting the parameter, inference, and memory overhead of GeNCoup.

Parameter overhead. Both variants introduce a single dense coupling matrix with about (C^2) parameters, where (C) is the number of channels. While the relative increase in parameters is negligible for FMPatch, DiffPatch, Sundial, and Aurora, it is more noticeable for compact models such as PPM and K2VAE. Nevertheless, the additional storage ranges from only 196 B for ETT to 2.83 MiB for Traffic, remaining modest relative to modern computational resources.

Inference overhead. The coupling introduces a similar number of additional FLOPs across models, with minor differences arising from the model-specific noise dimension (d). Consequently, the relative latency overhead depends primarily on the backbone's own sampling cost. For FMPatch, DiffPatch, Sundial, and Aurora, the measured increase in latency is less than 0.2%. The overhead is more noticeable for the lightweight samplers PPM and K2VAE. For the largest dataset, Traffic, the coupling adds approximately 35% latency for both variants on PPM, and 26% for GeNCoup-Corr and 10% for GeNCoup-Dyn on K2VAE. Across all measured settings, from ETT to Traffic, the added inference latency ranges from approximately 0.004 to 4.8 ms per window.

Training and memory overhead. Training GeNCoup-Dyn requires backpropagation through the frozen forecaster's sampling procedure, particularly for iterative diffusion and flow-matching models. We mitigate the resulting activation-memory overhead using gradient checkpointing. We measured peak GPU memory during Dyn optimization and compared it against gradient-free coupled sampling and VS evaluation. Across ETT and Weather, the median training-to-inference peak memory ratio is 1.1–1.2× (0.43-2 GiB) for K2VAE, PPM, Sundial, and Aurora, and approximately 2.5× (6.2 GiB) for FMPatch and DiffPatch. For Solar, ECL, and Traffic, the median ratios range from 2.1× to 4.1× (4.16-32.2 GiB) across models.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions to the manuscript

We thank all three reviewers for their careful reading and constructive suggestions. Their feedback helped us address a potential information leakage risk in our method, improve the fidelity of baseline implementations to ensure fair comparisons, and strengthen the experimental evidence, as well as improve the manuscript. Below, we summarize the main revisions and additional analyses. Changes are highlighted in the revised PDF.

  1. Corrected local correlation estimation and reran experiments. As pointed out by Reviewer 3xQD, our original implementation introduced a potential information leakage risk by pooling lookback observations across windows in the same evaluation minibatch. Since these windows originate from the same time series, later windows may contain observations unavailable at the forecast origin of earlier windows. We eliminated this risk by estimating the local correlation matrix independently using only each window's own observed lookback. We reran all affected experiments and analyses and updated the results throughout the manuscript. The correction does not change the proposed coupling mechanism.

  2. Improved the fidelity and fairness of the COBASE baseline. Following Reviewer 3xQD's concerns, we revised COBASE to refit its copula independently for each forecasting window using the preceding observations, rather than fitting it globally once on the training set, as in our original implementation. Subsequently, following the reviewer's suggestion, we included both global and local fitting variants, selecting the best configuration based on validation VS for each (forecaster, dataset) combination, consistent with the validation-based selection used by GeNCoup-Corr. This ensures a fairer comparison, and we updated the reported results accordingly. Following these corrections, GeNCoup's relative performance among the baselines remains largely unchanged: our methods achieve the best VS on 7/8 datasets (unchanged) and the best ES-Chan on 5/8 datasets (compared with 6/8 previously). Thus, the main contributions and conclusions of our work remain unchanged.

  3. Introduced an empirical preservation criterion. In response to Reviewer 9Cvp, we formalized marginal and temporal preservation using the uncoupled forecaster's sampling variability across random seeds. This provides a quantitative criterion for assessing whether changes in CRPS, NMAE, and ES-Time are consistent with sampling noise.

  4. Tested GeNCoup on datasets with lagged cross-channel dependencies. Following Reviewer 6YhM's concerns regarding the evaluation of cross-channel dependence across time, we tested GeNCoup-Corr on three MixBench-TS datasets exhibiting different levels of lagged dependence. GeNCoup-Corr improved VS across all evaluated datasets and prediction horizons, providing further evidence of its effectiveness beyond the standard benchmarks.

  5. Analyzed the cross-channel dependence already captured by the original forecasters. Following Reviewer 3xQD's questions regarding the limited improvements on high-dimensional datasets, we introduced an additional correlation analysis comparing uncoupled and coupled forecast correlations against the ground truth. This analysis provides insights into the smaller improvements observed on ECL and Traffic, suggesting that the original forecaster already captures substantial cross-channel dependence on these datasets.

  6. Expanded implementation and computational details. Following questions from Reviewers 3xQD and 6YhM, we provided a more complete description of GeNCoup-Dyn's training procedure, including data splits, optimization settings, and checkpoint selection. We also added a detailed analysis of parameter, inference-time, and memory overhead.

  7. Added a Limitations section. In response to the reviewers' feedback, we added a dedicated Limitations section to the revised manuscript.

Overall, the additional experiments, corrected evaluations, and expanded analyses reinforce the main conclusions of our work: GeNCoup provides an effective plug-and-play mechanism for improving cross-channel dependence in frozen probabilistic forecasters while empirically preserving their marginal accuracy and temporal dynamics.

We sincerely thank the reviewers again for their valuable feedback and for helping us improve the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.