acceptodds
Under review as a conference paper at ICLR 2027

CopuAdapt: An Attentional Copula Adapter for Improved Joint-Distribution Sampling in Probabilistic Time-Series Forecasting

Abstract

Time-series Foundation Models (TSFMs) have demonstrated strong forecasting performance and generalization across diverse domains. For multivariate forecasting, many TSFMs adopt channel-independent modelling. These approaches, however, can struggle to capture cross-variable dependencies in joint forecast samples. We introduce CopuAdapt, a novel attentional copula adapter that augments frozen TSFMs with a learned multivariate dependence model without requiring foundation-model retraining. CopuAdapt constructs empirical marginal distributions from the backbone's forecast samples or quantiles and couples them through an attentional copula architecture operating in probability space. To capture regime-aware dependence, we employ a sparse attention mechanism that infers a directed acyclic graph of conditional independences for each forecasting window, supported by a JEPA-inspired auxiliary objective that learns forecast-relevant temporal memory representations. By leveraging these structural details to dynamically re-order marginal sampling, CopuAdapt strictly preserves the backbone's empirical marginals. Consequently, CopuAdapt, evaluated across the TSLib datasets, exactly preserves the Continuous Ranked Probability Score (CRPS) of the underlying TSFM while significantly improving joint-sampling sensitive metrics, including the Variogram Score (VS), and RBF-MMD, with a median improvement of 4 to 18% across metrics. Code for this paper can be found at https://anonymous.4open.science/r/CopuAdapt-6C85.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time series foundation model, probabilistic forecasting, adaptor, copula

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

The paper addresses the problem of learning multivariate dependence in time-series models, where many current approaches operate under a channel-independence assumption. Unlike existing methods that adapt channel-independent models to multivariate settings by jointly learning marginal forecasting and multivariate dependencies, which can lead to a degradation in marginal performance, CopuAdapt exclusively focuses on recovering the missing multivariate structure by augmenting pre-trained, frozen time-series models with a learned multivariate dependence model.

The proposed end-to-end system includes five main components: (1) a frozen, pre-trained backbone that provides the channel-independent forecasts; (2) an encoder that processes the history window per channel and pools it using attention-based pooling to produce a global descriptor per variable, providing context for the dependence learning components and supervised via a Joint Embedding Predictive Architecture (JEPA)-inspired objective to predict representations of the future context from history representations; (3) a residual feedforward network that predicts features used to estimate pairwise edge probabilities in a directed acyclic graph, dictating which other variables' forecast values each variable can condition on; (4) an attention-based, graph-conditioned copula generation module that models joint dependence per variable by re-ranking samples; and (5) a non-parametric ranker that restores coordinate-wise empirical marginals so they remain identical to the underlying TSFM, transferring only the rank dependence structure.

The method is evaluated against two baselines, AdaPTS and ICA, across four time-series foundation models (Chronos, Moirai2, Sundial, and TiRex). CopuAdapt is shown to better preserve marginal forecasting accuracy, achieving substantially lower CRPS compared to AdaPTS and ICA, while showing a marginal degradation compared to the base model observed only for Chronos. It also improves metrics that measure joint modeling, such as the Variogram Score (VS) and RBF-MMD compared to AdaPTS and ICA. The proposed adapter adds roughly 0.7% to 8.5% additional parameters relative to the benchmarked base model backbones.

Critical strengths
  1. A novel approach to multivariate dependence learning in time series that decouples marginal forecasting from dependence modeling.
  2. The method is shown to substantially better preserve marginal forecasting accuracy compared to existing multivariate adaptation baselines while achieving superior (lower) VS and RBF-MMD scores.
Critical weaknesses
  1. The paper, and particularly the method section, is difficult to follow and digest, which obscures the paper's core contributions.
  2. The ablation studies are relatively limited given the complexity of the five-component pipeline. In addition, while the adapter is lightweight relative to large foundation models, it introduces more parameters than compact, strong architectures like PatchTST or TimeMixer. The paper does not clarify why the framework was not evaluated on such models.
Questions for authors
  1. Can the authors please provide a concise summary of their pipeline, with a particular focus on (a) the motivation and assumptions underlying the design on each component (b) input, output and purpose
  2. Can the authors please provide ablation studies isolating the contribution of each key design choice across the five proposed pipeline components?
  3. Have the authors considered or evaluated applying their method to smaller yet highly competitive architectures (e.g., PatchTST, TimeMixer)?

Post-rebuttal: The authors have thoroughly addressed my concerns by providing a per-component ablation study that clarifies each module's contribution, along with additional baseline comparisons that align well with the results reported in the original submission. I am raising my score from 2 to 3 accordingly.

2 comments

We thank you for noting the novelty of our approach, our substantial improvement in marginal preservation, and superior performance on VS and MMD.

Concise pipeline summary: motivation, assumptions, inputs, outputs and purpose of each component.

We will add the following summary of the five components to Section 4.

ComponentInputOutputPurpose and motivationAssumption
1. Frozen TSFM marginals (Sec. 3, Eqs. 4–5)Each variate's historyForecast samples or quantiles per variate and step; PITs via an empirical CDF of backbone support and historyKeep the pretrained marginal accuracy; makes the adapter backbone-agnosticBackbone marginals are worth keeping; marginals stay channel-independent
2. Temporal encoding with JEPA (Secs. 4.1–4.2, Eqs. 6–8)Standardised history of each variateWindow memory z=[h;m]z = [h; m]; JEPA aligns mm with an embedding of the future (training only)Condition the graph and copula on the current regime; JEPA makes the memory forecast-relevant, giving sparser graphsDependence varies with context and is predictable from history
3. DAG inference (Sec. 4.3, Eqs. 9–10)Refined memory featuresBernoulli edge probabilities, strictly lower-triangular in the dataset's variate order; one graph per trajectoryChoose each variate's parents; fewer topological levels allow parallel samplingA fixed variate order; graphs are a modelling device, not causal claims
4. Graph-conditioned copula generation (Sec. 4.4, Eqs. 11–15)Child queries, parents' PITs, sampled gatesHistogram conditional densities on [0,1][0, 1] per variate and stepModel cross-variate dependence in probability space, independent of marginal scaleA variate's horizon steps are conditionally independent given its parents and context
5. Rank matching (Sec. 4.5, Eqs. 16–17)Raw adapter draws and a fixed backbone reference ensembleFinal joint ensemble with identical empirical marginalsPreserve marginals and CRPS exactly while transferring dependence through ranks; trajectory-aware matching also keeps backbone pathsRank dependence is what transfers; reference ties coarsen it

Additional Ablations isolating core components.

ComponentAblationResult
1. Frozen TSFM marginalsFour TSFMs in the paper (Table 1) and, new, PatchTST trained in-domainOn PatchTST: +18.6% MMD, +3.3% TVS, +1.1% VS (27 of 28 cases)
2. Temporal encoding with JEPAJEPA on vs off, joint scores (new)No accuracy effect, solely improves sparsity
3. DAG inferenceLearned sparse DAG vs matched dense reference (new); sparsity vs latency (Figure 2)Same sample scores; sparse has better NLL and cheaper sampling
4. Graph-conditioned copula generationReplaced by a Gaussian copula and by TACTiS-2's attentional copula, same marginals (new)CopuAdapt better than TACTiS-2 on VS and TVS for both backbones, and better than Gaussian for Sundial
5. Rank matchingPer-step vs trajectory-aware rank matching (new)Chronos: +30.6% MMD, +10.0% TVS, +3.1% VS; PatchTST: +13.6% MMD; Sundial: no difference

Compact architectures.

We now evaluate CopuAdapt on PatchTST, using the official implementation and per-dataset hyperparameters with a 19-quantile output head (levels 0.05 to 0.95).

  • PatchTST-Q: the quantile PatchTST trained in-domain.
  • PatchTST-Q + CopuAdapt: CopuAdapt trained on PatchTST-Q samples. Its training PITs come from 2-fold cross-fitting, so no PatchTST model ever forecasts a window it was trained on.

CopuAdapt significantly improves every dependence metric over PatchTST-Q, with CRPS identical by construction:

PatchTST-Q + CopuAdapt vs PatchTST-QVSTVSMMD
Trajectory-aware restoration+1.1% (26/27)+3.3% (26/27)+18.6% (25/27)
Per-step restoration+0.3% (27/27)+0.9% (27/27)+5.8% (24/27)

All Holm-corrected p<1e-5p < 1\text{e-}5. Results cover 27 of the 28 cases; weather H192 is still training and will be added. As with Chronos, trajectory-aware restoration is clearly better than per-step (+13.6% MMD, +2.4% TVS, +0.8% VS; p<1e-5p < 1\text{e-}5).

We thank you for your thorough review improving the clarity and empirical rigour of our work. We hope our reply has clarified our proposed approach, demonstrated the value of our method even on a learned in-domain backbone, and that in light of this you consider raising your score and supporting our acceptance.

I thank the authors for their response. They have thoroughly addressed my concerns by providing a per-component ablation study that clarifies each module's contribution, along with additional baseline comparisons that align with the results reported in the original submission. I will raise my score accordingly.

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

This paper introduces a copula model (CopuAdapt), which especially meant for Time Series Foundation Models (TSFMs). For that the auhtors use the attentional copula introduced in Tactis paper. However, what is not in the original Tactis archtitecure is a Directed Acyclic Graph (DAG) which models which variables depend on which other variables. The DAG based on the history of observed values. It is then used to create the respective attention masks for the attentional copula. The copula model is trained with a objective that consists of four components:

i )The copula NLL (refering to the copula density),

ii) a term that punishes every edge in the DAG ,

iii) a KL divergence for the univariate marginals (necassary, as attentional copulas do not guarantee to not change the marginals),

iv) a JEPA loss, in which one trains a seperate model to recover the time point from an encoded obersvaiton.

In their experiments the authors that a trained CopuAdapt together with a TSFMs (which models channels as independend), is able to recover the CRPS (which is expected from a copula). Additionally, CopuAdapt helps to getter a Variogram and Engergy Score, compared to the TSMFs that model channels as independend.

In their ablations, the authors provided evidence for the usefulness of their JEPA auxiialry loss. While the accuracy is barely effected, the JEPA loss leads to DAGs with higher sparsitiy. Additionally, the authors show that sparser DAGs lead to faster inferences, when at least 60% of the edges are dropped. There is no analysis on how the inductvie bias of the DAG effects accuracy. There is no comparison with Tactis-2 coupled with TSFMs, an experiment which, at first glance should be easily feasable.

Critical strengths

S1

The idea to use a copula to couple the probabilistic forecasts of channel-independent models, such as many foundation models, is very promising. The learned dependencies seem to be valuable and lead to lower RBF-MMD and Variogram Scores.

S2

The authors put a lot of effort into designing the objective function, and every term is well motivated and partially evaluated with corresponding ablation studies.

S3

The idea to mask the attentional copula with a learned DAG should provide a useful inductive bias and can also lead to cheaper sampling when the DAG is sufficiently sparse.

Critical weaknesses

W1

CopuAdapt models the dependence between a channel and its parent channels in the DAG across all time points. However, it cannot directly model the dependence of forecasting targets within a channel at different time points; this dependence is only captured indirectly through shared parents. For channels without parents (which always includes the first channel in the fixed ordering), the forecasting targets at different time points are conditionally independent given the history. This is a very strong limitation.

W2

While the DAG is intuitively useful and provides faster sampling at a certain sparsity level, the authors refrained from evaluating its advantages in terms of accuracy.

W3

The authors frame their work as a copula model for TSFMs. However, it should work with any probabilistic time series forecasting model that provides samples or quantiles. Therefore, one should compare its forecasting performance with other probabilistic multivariate forecasters. For example, a comparison with TACTiS-2 on the datasets that the TACTiS-2 paper used to evaluate their model would greatly improve the evaluation section of this work.

W4

The authors should also compare their model with other time series copulas (e.g., a Gaussian copula inferred from the embeddings of CopuAdapt's encoder).

  • Some minor typos (e.g., l. 139, l. 413, l. 423).
Questions for authors

Q1: Which of the evaluated TSFM do model intra-channel dependcies in their distributions ?

Q2: Could you report the NLLs?

Q3: Why did you not used the encodings of the TSFMs?

Q4: Could you provide an ablation study that shows the importance of the DAG ? What happens if we use the most naive version of DAG with 1->2->-3 ... ->D ?

Q5: Could you compare with the full attentional copula from Tactis ?

3 comments

Thank you for noting the promise of the proposed copula based approach, the value of the learned dependancies, and the effective use of sparsity for more efficient sampling.

In response to the questions you raised:

No direct within-channel temporal dependence.

We agree this is a property of the decoder (see the response to Reviewer MZNv regarding the restriction on temporal dependence). In the final forecast, within-channel dependence comes from the backbone's own sample paths, which the trajectory-aware restoration keeps intact. The temporal variogram score shows this works: CopuAdapt beats the channel-independent backbone on TVS in 24/28 (Chronos) and 28/28 (Sundial) cases, and beats the TACTiS-2 copula on TVS for both backbones (see the response to Reviewer MZNv regarding the simple copula comparison).

Accuracy benefit of the DAG.

We compared the learned sparse DAG against its matched dense reference, the on identical marginal samples. The two match on every sample score (differences within ±0.8%\pm 0.8\%), and the sparse model has a better test copula NLL in 20/28 (Chronos) and 23/28 (Sundial) cases. The DAG's contribution is therefore cheaper sampling at no loss of accuracy and improved NLL.

Comparison with other probabilistic multivariate forecasters.

We agree CopuAdapt applies to any model that produces samples or quantiles, and we now show this on PatchTST, a compact model trained in-domain (see the response to Reviewer RTKX). To isolate dependence modelling, we compare against TACTiS-2's attentional copula on the same marginals.

Other time-series copulas.

We added a Gaussian copula fitted on the same training PITs, and TACTiS-2's attentional copula, both on identical marginals (results in our response to Reviewer MZNv).

Q1. Which TSFMs model intra-channel dependence in their distributions?

  • Chronos: yes. It is univariate and obtains each sample by autoregressively sampling tokens, so every sample is a temporally coherent path.
  • Sundial: yes. It is pre-trained univariately, and its flow-matching head generates a whole multi-step output block per sample. The released Sundial-Base generates up to 720 steps per block, so every horizon we evaluate is one jointly sampled path.
  • TiRex: no. It is univariate and outputs nine marginal quantile levels (0.1 to 0.9) per step. Long horizons are carried by its recurrent state with future inputs treated as missing, so it defines no joint distribution over the horizon.
  • Moirai 2.0: no. It treats multivariate data as independent univariate series and outputs nine quantile levels per step. Its recursive decoding conditions each step on the previous step's quantiles, but the output remains per-step marginal quantiles, not a joint distribution over paths.

None of the four model cross-variate dependence. For the quantile models, CopuAdapt supplies all of the joint structure, consistent with its largest gains being on TiRex and Moirai2 (as shown in Table 1).

NLLs.

We report the test copula NLL in probability space, where all models share the same empirical marginal transform. Units are nats per forecast coordinate; independence scores 0, and lower is better.

ModelChronosSundialBetter than independence
CopuAdapt (sparse)−0.554-0.554−0.722-0.72228/28 and 28/28 (p≈1e-7p \approx 1\text{e-}7)
CopuAdapt without JEPA−0.548-0.548−0.730-0.73028/28 and 28/28 (p≈1e-7p \approx 1\text{e-}7)
CopuAdapt dense reference−0.506-0.506−0.679-0.67928/28 and 28/28 (p≈1e-7p \approx 1\text{e-}7)

Sparse improves on dense by 0.049 (Chronos, 20/28, p<0.001p < 0.001) and 0.043 nats (Sundial, 23/28, p<0.001p < 0.001). JEPA has no significant effect. NLLs are reported relative to the independence copula (0) on shared marginals. The TSFMs we adapt do not provide tractable joint likelihoods, so comparisons against the foundation models' own forecasts use proper scoring rules (see the General response).

Why not use the TSFM encodings?

CopuAdapt needs only forecast samples, this keeps it backbone-agnostic: the same adapter architecture works with Chronos, Sundial, TiRex, Moirai2 and PatchTST, and with models available only through an API, without depending on encoder shapes or internals that differ between models. Using encodings is a natural extension we will mention.

Comparison with TACTiS's full attentional copula.

We trained TACTiS-2's attentional copula on the same TSFM marginals, with its default hyperparameters and the same training windows and budget. CopuAdapt is significantly better on both VS and TVS for both backbones: VS +0.5% (Chronos, p=2e-4p = 2\text{e-}4) and +0.3% (Sundial, p=0.025p = 0.025); TVS +2.1% (Chronos, p=0.025p = 0.025) and +1.0% (Sundial, p=0.01p = 0.01). MMD differences are not significant.

We thank you for the time you spent crafting a useful and insightful review, we hope our response has demonstrated the value of our work and that in light of this you can support our acceptance and increase your score.

I thank the authors for their extensive answer to my questions. While the authors did add the full attentional copula (Tactis) as a baseline I do have the following issues in terms of fairness:

  • Tactis' hyperparameters were not tuned
  • NLL comparison (the metric on which these copulas were trained) is absent.
  • No comparison with previously published numbers. One could have reproduced at least a part of Tactis experiments and compare the two copulas respectivly.

I will therefore keep my score.

Clarifications on TACTiS-2 Fairness and NLL Comparisons

We thank the reviewer for their follow-up. To directly address your remaining concerns regarding NLL metrics, hyperparameter fairness, and alignment with previously published experiments, we evaluated both adapters on TACTiS-2's setup.

Specifically, we evaluated on FRED-MD (107 series, the Monash version utilized by TACTiS-2; evaluated at H=12H = 12 as in their original paper, and H=24H = 24). To ensure the fairest possible out-of-the-box comparison, we trained TACTiS-2's attentional copula using its official open-source code and the provided default hyperparameters. Both models were trained on the exact same TSFM marginals and training windows, with an identical training budget.

Under this setting, CopuAdapt achieves a lower test copula NLL in all four cases (for both Chronos and Sundial, across both horizons), improving by 0.11–0.40 nats per coordinate, and outperforming the baseline in 11 of 11 test blocks for each backbone.

We appreciate you pushing us to provide this NLL comparison on the FRED-MD dataser, as it strongly reinforces the efficiency and accuracy of our proposed architecture. We hope this thoroughly resolves the fairness concerns and gives you the confidence to support the acceptance of our work.

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

This paper introduces CopuAdapt, a post-hoc adapter for improving the joint forecast samples produced by frozen time-series foundation models. The method separates dependence modeling from marginal forecasting: a temporal encoder, trained with an auxiliary JEPA-inspired objective, extracts contextual representations; a stochastic sparse DAG, constrained by a fixed channel ordering, controls an attentional decoder operating in probability-integral-transform space. Sampling proceeds by topological level to enable parallel generation across eligible variables. Finally, coordinate-wise rank matching transfers the learned dependence structure to a fixed backbone ensemble, exactly preserving its empirical marginal distribution at every variable and forecast time. The evaluation shows unchanged marginal CRPS and improvements in dependence-sensitive metrics, particularly for Moirai2 and TiRex, although benefits are uneven across backbones.

Critical strengths
  1. Well-motivated practical objective. Strong univariate forecasts can still produce incoherent multivariate scenarios, making dependence adaptation a meaningful problem. Augmenting a frozen backbone while preserving its marginal forecasts provides a clear and practically useful objective.

  2. Sound preservation mechanism with promising empirical benefits. Equations (16)–(17) correctly preserve coordinate-wise empirical distributions through rank matching. Table 1 reports approximately 24.8% and 27.6% lower MMD for Moirai2 and TiRex, respectively, demonstrating that useful dependence adaptation is possible without changing marginal CRPS.

Critical weaknesses
  1. Unexamined restriction on temporal dependence. As specified by Eqs. (9) and (12), the first variable has no parents and its future coordinates are conditionally independent across forecast steps. Rank matching cannot restore this missing temporal dependence, potentially compromising the stated joint-trajectory objective. Neither temporal-dependence diagnostics nor channel-order sensitivity experiments establish whether this restriction is acceptable.

  2. Insufficient evidence for the architectural contribution. Attentional copulas and marginal-preserving rank reordering have precedents in TACTiS and ensemble copula coupling, making the adaptive sparse DAG the principal advance requiring validation. The experiments lack simple copula and matched dense-decoder controls using identical marginal samples, while ICA and AdaPTS also alter marginal quality. Moreover, the JEPA ablation reports sparsity without predictive scores, leaving the motivation and demonstrated benefit of these added components insufficiently connected.

Questions for authors
  1. Clarification — dependence modeling: Does the implementation include a temporal-dependence mechanism absent from Eq. (12), and how sensitive are forecasts to channel ordering? Is copula validity claimed for the raw decoder or only for the final rank-matched ensemble? If the former, how does Eq. (14) ensure uniform marginals for each conditioning history?

  2. Evidence request — controlled comparisons: With identical marginal samples, how does CopuAdapt compare against a simple copula and its matched dense reference? What are the JEPA ablation’s joint predictive scores alongside the reported sparsity?

  3. Clarification — evaluation protocol: Are VS and MMD computed on complete conditional forecast trajectories? Please specify their computation, uncertainty estimates, and the meaning of sparse-selection fallbacks.

  4. Clarification — robustness and baseline validity: What explains the 41.2% aggregate MMD deterioration for Chronos in Table 1? Given AdaPTS’s substantial degradation, how were its implementation, configuration, and hyperparameter selection validated?

6 comments

We thank you for noting our meaningful problem suggestion, that maintaining frozen marginals and augmenting a TSFM is a useful objective, and the promising empirical benefits of our proposed approach.

In response to the questions you raised:

Restriction on temporal dependence; does the implementation have a temporal mechanism absent from Eq. (12)?

The decoder has no within-variable temporal mechanism: given its parents and the context, a variable's horizon coordinates are conditionally independent.

Within-channel temporal dependence is inherited from the backbone. The updated trajectory-aware restoration, proposed to mitigate this issue, assigns whole backbone sample paths to adapter draws before matching ranks, so each restored trajectory keeps a coherent backbone path. On the new temporal variogram score (TVS), the final CopuAdapt ensemble beats the channel-independent backbone in 24/28 cases for Chronos (+1.9%, p=5e-5p = 5\text{e-}5) and 28/28 for Sundial (+3.3%, p<1e-6p < 1\text{e-}6). Your concern is confirmed for the initial per-step matching: on Chronos it is 9.0% worse than independent on TVS, which the trajectory-aware variant fixes (+10.0% over per-step, p=1e-4p = 1\text{e-}4).

Is copula validity claimed for the raw decoder or the final ensemble? How does Eq. (14) ensure uniform marginals?

Only for the final rank-matched ensemble. Rank matching (Eqs. 16–17) makes its coordinate-wise empirical marginals identical to the backbone reference ensemble. Eq. (14) does not guarantee uniform marginals for each conditioning history: it penalises deviation of the batch-averaged histogram from uniform, so it only encourages approximate uniformity.

Simple copula and matched dense reference with identical marginal samples.

Comparison (positive = CopuAdapt better)VSTVSMMD
vs Gaussian copula, Sundial+1.2% (23/28, p=0.001p = 0.001)+4.0% (25/28, p=2e-4p = 2\text{e-}4)+0.3% (n.s.)
vs Gaussian copula, Chronos+0.1% (n.s.)+0.8% (23/28, n.s.)−0.3%-0.3\% (n.s.)
vs TACTiS-2 attentional copula, Chronos+0.5% (24/28, p=2e-4p = 2\text{e-}4)+2.1% (20/28, p=0.025p = 0.025)+14.6% (p=0.06p = 0.06)
vs TACTiS-2 attentional copula, Sundial+0.3% (20/28, p=0.025p = 0.025)+1.0% (22/28, p=0.01p = 0.01)−0.5%-0.5\% (n.s.)
Sparse vs matched dense reference, both backboneswithin ±0.1%\pm 0.1\%within ±0.15%\pm 0.15\%within ±0.8%\pm 0.8\%

Sparse and dense match on every sample score. The sparse DAG therefore delivers its sampling-cost reduction at no measurable loss of accuracy, which is the role the paper assigns to it. However, in our reply to reviewer zHbG the reported negative log-likelihoods show statistically significant improvements for the sparse model.

JEPA ablation: joint predictive scores.

With both arms retrained, JEPA has no significant effect on any joint score (VS, TVS, MMD; all Holm-corrected p≥0.16p \ge 0.16, mean differences ≤0.2%\le 0.2\%), nor on test copula NLL. Thus only providing the sparsity benefits mentioned in the original paper.

Evaluation protocol.

  • VS is computed per test window on the complete H×DH \times D forecast trajectory, over all coordinate pairs with p=0.5p = 0.5 and equal weights, then averaged over windows.
  • RBF-MMD compares the complete flattened H×DH \times D trajectories. It uses a random-Fourier-feature approximation (4,096 features, fixed seed), with the bandwidth fitted on validation targets only, pooled over all test windows.
  • Every method uses 100 joint samples per window and all test windows, with identical seeds per window.
  • Uncertainty: In the provided ablations we report per-comparison Wilcoxon signed-rank tests with Holm correction and win counts over cases. These will be expanded to TiRex and Moirai prior to the camera ready but have been constrained to Sundial and Chronos to give an example of a direct multi-step forecaster and an iterative multi-step forecaster.
  • Fallbacks: sparse checkpoint selection takes the fewest expected edges among checkpoints whose validation NLL is within the tolerance of the dense reference and whose edge fraction is strictly between 0.1% and 99.9%. If no checkpoint qualifies, we record a fallback and use the checkpoint with the smallest tolerance violation, then lowest validation NLL.

The 41.2% MMD deterioration for Chronos.

Per-step rank matching assigns ranks independently at each variable-step pair, which scrambles Chronos's autoregressive paths. Trajectory-aware restoration removes the effect: +30.6% MMD over per-step (26/28, p=1e-4p = 1\text{e-}4) and +2.5% over independent. Sundial, which is not autoregressive, is not affected in a statistically significant way by the choice of per step or trajectory-aware sampling.

Thank you for your time in providing this thorough and insightful review and contributing to the improvement of how we handle temporal dependence. We hope our reply can encourage you to raise your score supporting the publication of our work.

Thanks for the response. The questions regarding the scope of the copula guarantee, sampling settings, and fallback procedures have been addressed, but the following concerns remain:

  1. Consistency of the TVS results. TVS should remain unchanged if it uses only within-variable temporal pairs and restoration bijectively reorders the same reference paths. Therefore, the reported improvements on TVS require further clarifications on the exact TVS definition.
  2. Evidence for the architectural contribution. Sparse and dense models have indistinguishable joint scores, while JEPA provides no statistically significant predictive improvement. This makes efficiency central to the architectural contribution. However, the response failed to establish the end-to-end speedup after incorporating trajectory-aware restoration.
  3. Unanswered questions. Channel-order sensitivity and the implementation and hyperparameter validation of AdaPTS remain unanswered.

We thank you for your response and continued efforts in improving the quality of this work. To answer your questions:

1. Consistency of the TVS results and trajectory-aware restoration We confirm that for a strictly bijective whole-path restoration (λ=0\lambda = 0), TVS remains invariant. In the 9 Chronos cases where validation selected λ=0\lambda = 0, TVS matches the independent backbone to within 10−710^{-7}%. However, the reported TVS improvements occur because validation selected 0<λ≤10 < \lambda \le 1 in 47 of 56 cases (including all 28 on Sundial), accounting for the entire TVS gain. Under these settings, a draw can transition to a different reference path across steps, splicing paths together. This alters within-variable step pairs and changes the TVS, while strictly preserving coordinate-wise empirical marginals. Our TVS computation uses within-variable pairs only, lags 1 to H−1H-1 weighted equally, p=0.5p = 0.5.

2. Evidence for the architectural contribution The role of the sparse DAG is to make dependence sampling scale with graph depth rather than D2D^2, at no cost to accuracy. The sparse model and matched dense reference yield identical sample scores, but the sparse model achieves a better test copula NLL (+0.049 nats for Chronos in 20/28 cases; +0.043 nats for Sundial in 23/28 cases).

At the adapter level, sparsity provides a measurable speedup via DAG-parallel sampling. In our ETT ablations, the sparse sampler is faster than the fastest dense sampler in 8 out of 8 cases. For example, on Chronos with H=192H=192, the sparse sampler completes in 11.3 ms compared to 20.1 ms for the dense reference, representing a 1.78$\times$ speedup. This end-to-end computational benefit scales robustly in scenarios where DD is large and the learned graphs are sparse.

3. Unanswered questions: Channel-order sensitivity and AdaPTS validation Channel-order sensitivity: To evaluate sensitivity to the autoregressive channel conditioning, we retrained 48 models under 3 seeded random orders on the Illness and Exchange datasets across chronos and sundial and across 4 horizons. No permuted order differed significantly from the original fixed ordering on any metric. Furthermore, on Sundial, every random order maintained its significant lead over the independent backbone on VS (+1.50% to +1.81%) and TVS (+4.36% to +5.23%) in 8 out of 8 cases.

AdaPTS baseline validation: To ensure a strictly fair baseline, we utilized the official upstream AdaPTS implementation. The baseline was trained exclusively on the training contexts with a training-only scaler, leaving the backbone frozen. For the final ensemble, each of the 100 draws is a fresh stochastic projection paired with one backbone sample path, decoded back to the original variables with no added noise. We intentionally utilized the original authors' default hyperparameters without secondary tuning to guarantee an out-of-the-box comparison; validation was strictly limited to shape and finiteness checks.

We hope this clears up any remaining concerns and allows you to support the acceptance of this work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for noting the novelty of the proposed method, the efficacy of our marginal preservation, and the promising improvements in Variogram Score and Maximum Mean Discrepancy . To answer the questions posed we ran six new controlled experiments: on Chronos and Sundial over 7 datasets (including the 4 ETT datasets, exchange, Illness and weather) ×\times 4 horizons (28 cases per backbone), and on PatchTST.

Statistics. For each comparison we report the geometric-mean improvement over cases, wins out of 28, and a two-sided Wilcoxon signed-rank test on per-case log score ratios. P-values are Holm-corrected within each experiment and backbone.

New metric. In addition to the variogram score (VS, cross-variable dependence), RBF-MMD and CRPS, we report a temporal variogram score (TVS) over pairs of horizon steps within each variable. It directly measures the within-channel temporal dependence that Reviewers MZNv and zHbG asked about.

ExperimentQuestion it answersReviewersMain result
Trajectory-aware vs per-step restorationDoes rank matching destroy temporal dependence? Cause of the Chronos MMD drop?MZNv, zHbGChronos: +30.6% MMD, +10.0% TVS, +3.1% VS (all p≤1e-4p \le 1\text{e-}4). Sundial: no difference
Gaussian copula, same marginalsSimple copula controlMZNv, zHbGSundial: +1.2% VS, +4.0% TVS (p≤0.001p \le 0.001). Chronos: tie
TACTiS-2 attentional copula, same marginalsFull attentional copulaMZNv, zHbGBetter VS and TVS on both backbones (p≤0.025p \le 0.025)
Dense reference vs sparse DAGMatched dense control; value of sparsityMZNv, zHbG, RTKXSame accuracy (all differences <1%< 1\%); sparse has better test NLL
JEPA vs no JEPA, joint scoresPredictive effect of JEPAMZNv, RTKXNo significant difference on any metric
Test copula NLLLikelihoodzHbGAll CopuAdapt variants beat independence in 28/28 cases (p≈1e-7p \approx 1\text{e-}7)
PatchTST-Q vs PatchTST-Q + CopuAdaptCompact in-domain backboneszHbG, RTKX+18.6% MMD, +3.3% TVS, +1.1% VS (p<1e-5p < 1\text{e-}5; 27 of 28 cases)

One change to the method. The new experiments revealed that per-step rank matching discards the coherent sample paths of autoregressive backbones. We therefore use a trajectory-aware restoration: whole backbone trajectories are assigned to adapter draws before matching, with the restoration variant chosen on validation windows. Marginals stay exactly preserved. We report both variants throughout, and will state the change clearly in the camera ready version.

We thank the reviewers for their time in providing thorough reviews which have contributed to the improvement of our proposed adaptor and the experimental rigour of our work. We hope these changes can encourage them to consider increasing their scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.