acceptodds
Under review as a conference paper at ICLR 2027

Cross-Window Relations for Transferability Estimation in Time Series Forecasting

Abstract

Time series forecasting often benefits from transferring knowledge from auxiliary datasets, yet transfer performance depends strongly on which source dataset is selected. Existing transferability estimators mainly assess source–target compatibility at the level of individual forecasting windows, overlooking the local evolution between successive overlapping windows. In this paper, we introduce the Cross-Window Relation Estimator (CWRE) for source dataset selection in time series forecasting. CWRE evaluates each candidate source using its frozen source-pretrained model on the target dataset, jointly measuring how well the learned representations explain future sequences and how well representation changes explain the corresponding future changes across adjacent windows. To reduce noise amplification introduced by differencing, CWRE separates trend and seasonal components, applying cross-window evaluation to the smoother trend component while retaining complementary predictive information through instance-wise evaluation. These relations are integrated through closed-form ridge regression, enabling efficient scoring without target-domain fine-tuning. We further evaluate transferability under a cross-architecture protocol, where scoring and downstream forecasting models are distinct. Experiments across multiple target datasets, architectures, and missing-observation settings show that CWRE achieves higher overall agreement with downstream transfer-performance rankings than distribution-alignment, task-transferability, and time-series-specific baselines. The results demonstrate that cross-window evolution provides complementary information for identifying transferable source datasets beyond conventional instance-wise compatibility. To ease reproducibility, we release our implementation code online https://anonymous.4open.science/r/CWRE-submission-3FC1/.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, transfer learning, source dataset selection

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_9nq7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CWRE, a transferability estimator for source dataset selection in time-series forecasting. Instead of evaluating individual forecasting windows alone, CWRE considers how changes in learned representations correspond to changes in future sequences across adjacent windows. The method combines cross-window relations, trend–seasonal decomposition, and ridge regression to estimate transferability without target-domain fine-tuning. Experiments across several datasets and forecasting architectures show improved source-ranking agreement over existing methods. I find the cross-window perspective interesting, although its generalization and methodological novelty need further justification.

Critical strengths
  1. The paper studies a meaningful problem. Selecting transferable source datasets without expensive target fine-tuning has clear practical value.

  2. I find the cross-window relation idea interesting. Considering representation changes rather than only individual windows provides a useful additional perspective on transferability.

  3. The empirical evaluation is reasonably comprehensive. The cross-architecture experiments and missing-observation settings are useful, and the ablation results support the contribution of cross-window modeling.

Critical weaknesses
  1. The methodological novelty is moderate. Regression-based transferability estimation is well established, and CWRE mainly extends it with temporal differences and decomposition.

  2. The experiments use only three target datasets and six candidate sources per target. I am not sure whether the ranking improvements would remain stable with larger source pools.

  3. Hyperparameter selection needs more justification. Each source independently selects its best scoring configuration, which may affect the comparability of transferability scores.

  4. Adjacent windows overlap heavily. It is still unclear how much of the improvement comes from meaningful temporal evolution rather than this structural dependence.

Questions for authors
  1. How does CWRE differ fundamentally from existing transferability estimators such as LinMSE and TEMPLATE, beyond introducing adjacent-window differences?

  2. Could the authors clarify why independently optimized hyperparameters produce comparable scores across different source datasets?

  3. Would the improvements remain similar with non-overlapping windows or larger temporal gaps?

  4. Could the authors provide results on more target datasets or larger candidate-source pools to better demonstrate generalization?

7 comments
AuthorsOpenReview ↗

Response To Reviewer 9nq7 [Part 1]

Dear Reviewer,

We sincerely thank you for your careful review and constructive feedback. Your comments on methodological novelty, hyperparameter fairness, overlapping windows, and generalization have helped us clarify our design rationale and strengthen the experimental validation of CWRE.

Response to W1 and Q1: Novelty Beyond Regression-Based and Temporal-Pattern Transferability Estimation

Reviewer's Concern (W1): The methodological novelty is moderate. Regression-based transferability estimation is well established, and CWRE mainly extends it with temporal differences and decomposition.

Reviewer's Question (Q1): How does CWRE differ fundamentally from existing transferability estimators such as LinMSE and TEMPLATE, beyond introducing adjacent-window differences?

We agree that regression and temporal decomposition are established techniques. However, the novelty of CWRE lies in introducing a new transferability criterion: whether changes in source-pretrained representations correspond to changes in future forecasting targets across successive windows.

1. Beyond Conventional Transferability Criteria: Evaluating Predictive Local Evolution

Existing estimators assess source transferability from different perspectives:

  • Distribution alignment methods (e.g., CORAL) measure source–target distribution discrepancies, but distributional similarity does not necessarily indicate forecasting transferability.
  • Regression-based estimators (e.g., LinMSE and LogME), originally developed and primarily evaluated on visual tasks, assess the compatibility between pretrained representations and target labels through instance-wise evaluation, without explicitly modeling cross-window temporal evolution.
  • Time-series-specific estimators (e.g., TEMPLATE) explicitly consider temporal dependencies, temporal patterns, and task adaptation, but do not directly evaluate whether cross-window representation changes correspond to changes in future forecasting targets.

In time-series forecasting, overlapping windows represent successive states of the same temporal process, where historical contexts and future sequences evolve jointly. Thus, CWRE extends transferability estimation from evaluating individual forecasting states to additionally evaluating whether their predictive relations evolve consistently across successive windows.

2. Beyond Simple Temporal Differencing: Constrained and Noise-Aware Cross-Window Evaluation

CWRE evaluates whether representation changes explain the corresponding future changes, providing complementary evidence beyond instance-wise compatibility.

However, overlap can introduce correlations unrelated to prediction. Therefore, CWRE uses instance-wise relations to constrain cross-window relations through a shared linear mapping, limiting the independent fitting of potentially spurious cross-window correlations.

Furthermore, trend–seasonal decomposition mitigates noise amplification from differencing: the trend branch jointly evaluates both relations, while the seasonal branch retains complementary instance-wise information. Ridge regularization and closed-form solutions enable efficient scoring without target-domain fine-tuning.

3. Experimental Evidence

Component ablations. Under MR = 0, averaged over four targets using PatchTST, the results are:

\hline \text{Method} & \tau_w & W \\\\ \hline \mathbf{CWRE} & \mathbf{0.3626} & \mathbf{0.3804} \\\\\\ \text{w/o Cross-Window} & 0.0921 & 0.2643 \\\\ \text{w/o Shared Mapping} & 0.1236 & 0.2839 \\\\ \text{w/o Trend-Seasonal} & 0.2118 & 0.3143 \\\\ \hline \end{array}$$ These results show that CWRE's gains cannot be attributed to regression alone: **cross-window modeling, shared mapping, and decomposition each provide substantial contributions.** **Temporal-step sensitivity.** We evaluate different temporal steps on ETTh1 under MR = 0, averaging results across both scoring architectures. $$\begin{array}{ccc} \hline \text{Temporal Step} & \tau_w & W \\\\ \hline s=0\\ (\text{State-only}) & 0.1347 & 0.4107 \\\\ \mathbf{s=1} & \mathbf{0.5420} & \mathbf{0.5429} \\\\ s=95 & 0.3342 & 0.4536 \\\\ \hline \end{array}$$ **Findings:** CWRE outperforms the state-only variant even at larger temporal gaps, indicating that **cross-window relations provide useful transferability information beyond immediately adjacent windows**, although their effectiveness varies with the temporal step. **Overall, CWRE's novelty is not the individual use of regression, differencing, or decomposition, but introducing cross-window evolution correspondence as an additional transferability criterion and modeling it jointly with instance-wise compatibility under a shared mapping.** We will make these methodological distinctions more explicit in **Sections 2.3–2.4**, supported by the ablation and temporal-step analyses in **Section 3.3 and Appendix C.8**.
AuthorsOpenReview ↗

Response To Reviewer 9nq7 [Part 2]

Response to W2 and Q4: Generalization Across Target Datasets and Larger Source Pools

Reviewer's Concern (W2): The experiments use only three target datasets and six candidate sources per target. I am not sure whether the ranking improvements would remain stable with larger source pools.

Reviewer's Question (Q4): Could the authors provide results on more target datasets or larger candidate-source pools to better demonstrate generalization? Our original experiments cover three targets, six candidate sources per target, two scoring architectures, and three downstream evaluators, under three missing rates with multiple mask seeds. Following the reviewer's suggestion, we further extend the evaluation to a fourth target and a longer forecasting horizon.

1. Additional Target Dataset: Weather (512→96)

We add Weather as a fourth target, evaluated across MR = 0, 0.3, and 0.6. Results are averaged over PatchTST and TimeEmb, with τw\tau_w additionally averaged across three downstream evaluators.

\hline \text{MR} & \text{CWRE }(\tau_w/W) & \text{ETran }(\tau_w/W) \\\\ \hline 0 & \mathbf{0.4093/0.4500} & 0.2497/0.4179 \\\\ 0.3 & 0.3148/\mathbf{0.5072} &\ \mathbf{0.3631}/0.4964 \\\\ 0.6 & \mathbf{0.3209/0.3429} & 0.1946/0.2857 \\\\ \hline \end{array}$$ **Findings:** CWRE achieves higher Kendall's $W$ than ETran across all three missing rates and higher $\tau_w$ in two of three settings, providing additional evidence of effectiveness on a new target domain. **2. Longer Forecasting Horizon: 512→512** We further extend the forecasting horizon from **96 to 512**, evaluating all four targets using two scoring architectures and three downstream evaluators under MR = 0. $$\begin{array}{llcc} \hline \text{Setting} & \text{Method} & \tau_w & W \\\\ \hline 512\rightarrow96 & \text{CWRE} & \mathbf{0.3820} & \mathbf{0.3946} \\\\ & \text{ETran (Best Baseline)} & 0.1978 & 0.3277 \\\\ \hline 512\rightarrow512 & \text{CWRE} & \mathbf{0.0914} & \mathbf{0.2393} \\\\ & \text{MMD (Best Baseline)} & 0.0616 & 0.2268 \\\\ \hline \end{array}$$ **Findings:** CWRE achieves the highest overall ranking agreement on both metrics, supporting its effectiveness beyond the original forecasting horizon. **3. Generalization to Larger Candidate-Source Pools** - **Current Coverage:** Our candidate sources span **five representative time-series forecasting domains**: electricity transformers, electricity consumption, transportation, exchange rates, and weather, providing diverse source distributions for evaluating transferability. - **Limitation:** We agree that evaluating a substantially larger candidate-source pool would provide stronger evidence regarding ranking stability as the number of candidates increases. Since expanding the pool requires source pretraining and downstream reference evaluation for every added dataset and architecture, we do not overclaim this aspect from the current experiments. The additional target and horizon results have been incorporated into **Sections 3.2 and 3.4 (Tables 1, 2, 3, and 4)**, extending our evaluation from **three to four targets and from one to two forecasting horizons**. Since Table 3 reports cross-target averages, we update it to include Weather for all methods, while keeping the original three-target results and evaluation protocols unchanged.
AuthorsOpenReview ↗

Response To Reviewer 9nq7 [Part 3]

Response to W3 and Q2: Shared Hyperparameter Selection and Source-Score Comparability

Reviewer's Concern (W3): Hyperparameter selection needs more justification. Each source independently selects its best scoring configuration, which may affect the comparability of transferability scores.

Reviewer's Question (Q2): Could the authors clarify why independently optimized hyperparameters produce comparable scores across different source datasets?

We apologize that our description may have led to this misunderstanding. No source-specific hyperparameter selection is performed in our experiments, either during source pretraining or CWRE scoring.

1. Actual Hyperparameter Selection Protocol

  • Source Pretraining: For a given scoring architecture, all six candidate sources use the same architecture, optimization settings, and training protocol. No source-specific hyperparameter tuning is performed during pretraining.
  • Selection Criterion: These parameters are selected using CWRE's own target-domain scoring objective, without downstream fine-tuning performance or ranking evaluation metrics.
  • Source Evaluation: All source-pretrained encoders remain frozen. Each source is evaluated using the same configuration:

Sk=T(Fk,Y;ω∗,ρtr∗,ρse∗).S_k = \mathcal{T}(F_k,Y;\omega^{\ast},\rho_{\mathrm{tr}}^{\ast},\rho_{\mathrm{se}}^{\ast}). 2. Source-Score Comparability

No source-specific score maximization occurs. All candidates share the same target data, scoring objective, and hyperparameter configuration. Only their source-pretrained representations and the corresponding derived projections differ.

Thus, source scores are compared under a unified evaluation protocol, eliminating the comparability concern arising from independently optimized scoring configurations.

We will clarify this protocol in Section 3.1 and Appendix C.1.2 (Eq. 44), and correct the misleading description that suggested independently optimized configurations for individual sources.

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

This paper proposes the Cross-Window Relation Estimator (CWRE), a transferability estimation method for source-dataset selection in time series forecasting. CWRE evaluates each candidate source through both the ability of its pre-trained representations to explain future sequences and the ability of representation changes across adjacent windows to explain the corresponding future-sequence changes. It further decomposes the future sequence into trend and seasonal components and performs cross-window evaluation on the smoother trend component, so as to mitigate the noise amplification introduced by differencing. Finally, closed-form ridge regression enables efficient scoring without target-domain fine-tuning. Experiments across multiple target datasets, model architectures, and missing-observation settings show that, compared with distribution-alignment, task-transferability, and time-series-specific baselines, CWRE achieves higher overall agreement with downstream ground-truth transfer-performance rankings.

Critical strengths

S1. The paper constructs the consistency between representation changes and the corresponding future-sequence changes across adjacent forecasting windows as an evaluation signal for source-dataset transferability in time series forecasting. It extends instance-wise evaluation, which traditionally operates only at the level of individual windows, to the local evolution across adjacent windows, which is something no existing method has done. S2. The paper evaluates under the setting of missing observations in the target sequence, and CWRE attains the highest τw in three of the four experimental settings and the highest W in all four. This demonstrates the method's robustness under adverse conditions and that it can better transfer to practical scenarios. S3. The paper's evaluation method separates the scoring model from the downstream model, so that the source ranking obtained from scoring stays consistent with the true rankings of different downstream models. This means that in practice there is no need to know the downstream model in advance or to change the scoring model accordingly. This shows that CWRE captures a consistently transferable signal rather than the specific preference of a particular model.

Critical weaknesses

W1. CWRE performs a grid search over (ω, ρtr, ρse) and takes the maximum over each candidate source, while for the baselines LogME, LinMSE, ETran, and TEMPLATE the paper does not report doing any hyperparameter tuning on them, so the two sides do not have a comparable tuning budget. Taking the maximum over the grid for each source itself inflates the reported scores; the paper does not rule out the possibility that the baselines could also catch up given equal tuning. Yet the paper's headline conclusion is exactly that CWRE exceeds the strongest baseline ETran, and under this currently asymmetric comparison the conclusion does not stand. W2. The target datasets are ranked over only 6 candidate sources, so the sample size of the rank-correlation evaluation is limited. Moreover, the full paper conducts no statistical significance test on the performance differences between methods, and reports no confidence intervals, bootstrap, permutation tests, or p-values, so the core quantitative gains lack support from statistical reliability. The standard deviations in Table 1 are computed from only 3 downstream forecasting architectures; under the complete-observation (MR=0) setting no multi-random-seed repeated experiments are reported, and for a large number of entries in the table the standard deviation exceeds the absolute value of the mean, reflecting that the ranking consistency fluctuates very strongly across different downstream architectures.

Questions for authors

Q1. Can you provide the distribution of CWRE's score over this grid, to verify whether the τ_w and W results obtained by taking the max significantly inflate the reported score? Q2. For the relevant gains in Table 1, can you provide a permutation test or bootstrap confidence intervals, and provide the per-downstream-architecture score breakdown in Table 1, to verify whether the method is genuinely good, or individual scores pull the average? Moreover, both the scoring models and the downstream models are relatively few, and whether the conclusion can really be supported on diverse models is still difficult to determine. Q3. When Traffic or Electricity is used as the target dataset, can the obtained scores still outperform other baselines, or does the performance drop substantially? Q4. As a motivating experiment for noise suppression, the gain of trend projection relative to the random control is 28.6% vs 30.0% on PatchTST, and is clearly present only on TimeEmb. Please provide the comparison of trend projection against the random control under more settings or more scoring architectures, to verify whether this gain holds generally, or mainly depends on TimeEmb.

7 comments
AuthorsOpenReview ↗

Response to Reviewer eeXa [Part 1]

Dear Reviewer,

We sincerely thank you for your careful review and constructive suggestions. Your comments raise important questions regarding hyperparameter fairness, statistical reliability, target-domain generalization, and the effectiveness of trend projection. We have clarified the evaluation protocol and conducted additional experiments to address these concerns, as detailed below.

Reviewer's Concern (W1): W1. CWRE performs a grid search over (ω, ρtr, ρse) and takes the maximum over each candidate source, while for the baselines LogME, LinMSE, ETran, and TEMPLATE the paper does not report doing any hyperparameter tuning on them, so the two sides do not have a comparable tuning budget. Taking the maximum over the grid for each source itself inflates the reported scores; the paper does not rule out the possibility that the baselines could also catch up given equal tuning. Yet the paper's headline conclusion is exactly that CWRE exceeds the strongest baseline ETran, and under this currently asymmetric comparison the conclusion does not stand.

Response: We thank the reviewer for raising this important concern and clarify the actual protocol with additional baseline tuning experiments.

(1) Clarification of source-specific tuning. We apologize that our description may have led to this misunderstanding.

First, source pretraining uses the same architecture and training settings across all six candidate sources, without source-specific hyperparameter tuning.

Second, CWRE scoring uses one shared hyperparameter configuration for all six sources under each target × missing rate × scoring architecture:

\theta^{\star}=(\omega^{\star},\rho_{\mathrm{tr}}^{\star},\rho_{\mathrm{se}}^{\star}).$$ **No source-specific maximization is performed**, and downstream MAE, $\tau_w$, and Kendall W are not used for parameter selection. **(2) Tuned-baseline comparison.** We additionally tune LinMSE, ETran, and TEMPLATE. For each target and scoring architecture, **one configuration is selected by maximizing the mean method-specific score across six sources**, then applied uniformly to all candidates. LogME performs internal evidence optimization without an external search grid. Results are averaged over **four targets and two scoring architectures** under MR = 0 (512→96); $\tau_w$ is additionally averaged over three downstream evaluators. $$\begin{array}{lccc} \hline \text{Method} & \text{Search Grid} & \text{Default }(\tau_w/W) & \text{Tuned }(\tau_w/W)\\\\ \hline \text{LinMSE} & \lambda:0,10^{-6},\ldots,10 & 0.0357/0.2723 & 0.1483/0.3161\\\\ \text{ETran} & \text{SVD: }0.1\text{--}1.0 & 0.1978/0.3277 & 0.1743/0.3259\\\\ \text{TEMPLATE} & \\{5,13,25,49,97\\} & −0.0964 / 0.2107 & -0.0410/0.2375\\\\ \text{LogME} & \text{No external grid} & 0.0162/0.2598 & 0.0162/0.2598\\\\ \mathbf{CWRE} & \text{---} & \mathbf{0.3820/0.3946} & \text{---}\\\\ \hline \end{array}$$ **Findings.** Tuning substantially improves LinMSE, while CWRE still exceeds the **strongest default-or-tuned baseline (ETran)** by **+0.1842/+0.0669 in $\tau_w$/W**. Thus, the aggregate advantage persists when baselines are given additional tuning opportunities under a shared source-selection rule. We will clarify this protocol in **Section 3.1 and Appendix C.1.2 (Eq. 44)** and include the tuned-baseline results.
AuthorsOpenReview ↗

Response to Reviewer eeXa [Part 2]

Response to Q1: Hyperparameter Sensitivity and Score Distribution

Reviewer's Question (Q1): Q1. Can you provide the distribution of CWRE's score over this grid, to verify whether the τ_w and W results obtained by taking the max significantly inflate the reported score?

Response: We thank the reviewer for this suggestion. As clarified in W1, CWRE does not perform source-specific maximization, and downstream ranking metrics are not used for hyperparameter selection.

Existing Evidence. Appendix C.8 (Figures 14–16) shows that ranking performance is relatively stable across moderate ω\omega values, while being more sensitive to the trend ratio than the seasonal ratio. These findings indicate that performance is not confined to a single hyperparameter configuration, motivating further analysis of its distribution across the grid.

Additional Analysis. To further examine whether the reported improvements rely on exceptionally favorable configurations, we analyze the ranking performance distributions across 63 shared hyperparameter configurations for each of four targets and two scoring architectures under MR = 0.

\hline \text{Target} & \text{Backbone} & \text{Main }(\tau_w/W) & \text{Median }(\tau_w/W) & \text{IQR }(\tau_w/W) \\\\ \hline \text{ETTh1} & \text{PatchTST} & 0.4807/0.5071 & 0.2848/0.4000 & 0.0848/0.0429 \\\\ & \text{TimeEmb} & 0.6032/0.5786 & 0.1841/0.4786 & 0.4567/0.1286 \\\\\\ \text{ETTm1} & \text{PatchTST} & 0.2281/0.2286 & 0.0830/0.1571 & 0.2440/0.0929 \\\\ & \text{TimeEmb} & 0.2141/0.2500 & 0.1070/0.1786 & 0.0392/0.0286 \\\\ \text{Exchange} & \text{PatchTST} & 0.3891/0.3786 & 0.2816/0.3429 & 0.4921/0.2357 \\\\ & \text{TimeEmb} & 0.3224/0.3143 & 0.3143/0.3143 & 0.0259/0.0143 \\\\ \text{Weather} & \text{PatchTST} & 0.3524/0.4071 & 0.3546/0.4214 & 0.1587/0.1000 \\\\ & \text{TimeEmb} & 0.4662/0.4929 & 0.4871/0.4929 & 0.1646/0.0214 \\\\ \hline \end{array}$$ **Findings.** Ranking performance varies across hyperparameter configurations, with some settings exhibiting greater sensitivity than others. Nevertheless, the average grid medians reach **0.2621/0.3482** in $\tau_w/W$, exceeding ETran's aggregate **0.1978/0.3277**. These results suggest that competitive ranking agreement is not restricted to the reported configurations. Importantly, the reported configurations are selected using CWRE's own scoring objective rather than downstream ranking metrics. We will include these distribution statistics in **Appendix C.8**.
AuthorsOpenReview ↗

Response to Reviewer eeXa [Part 3]

Response to W2.1 & Q3: Candidate-Source Coverage and Target-Domain Generalization

Reviewer's Concerns (W2.1): The target datasets are ranked over only 6 candidate sources, so the sample size of the rank-correlation evaluation is limited.

Reviewer's Question (Q3): When Traffic or Electricity is used as the target dataset, can the obtained scores still outperform other baselines, or does the performance drop substantially?

Response: We thank the reviewer for highlighting the importance of broader source coverage and target-domain generalization. We clarify the existing evaluation scope and provide additional results.

(1) Candidate-source coverage and expansion.

  • Current Coverage: The original evaluation covers three targets, six candidate sources per target, two scoring architectures, and three downstream evaluators, across three missing rates with multiple mask seeds. The dataset pool spans five representative forecasting domains: electricity transformers, electricity consumption, transportation, exchange rates, and weather.
  • Expansion Challenges: Expanding the source pool requires additional pretraining and reconstruction of reference rankings. Evaluating each new target requires 12 source–scoring-backbone and 18 source–downstream-architecture evaluations in the basic setting, plus baseline comparisons and missing-rate experiments.
  • Further Evaluation: We prioritize additional target and horizon experiments during rebuttal and plan to expand the candidate-source pool in the final version.

(2) Additional target dataset

We add Weather as a fourth target, evaluating both scoring architectures across three missing rates. Results are averaged over PatchTST and TimeEmb, with τw\tau_w additionally averaged across three downstream evaluators.

\hline \text{Missinig Rate} & \text{CWRE }(\tau_w/W) & \text{ETran }(\tau_w/W) \\\\ \hline 0 & \mathbf{0.4093/0.4500} & 0.2497/0.4179 \\\\ 0.3 & 0.3148/\mathbf{0.5072} & \mathbf{0.3631}/0.4964 \\\\ 0.6 & \mathbf{0.3209/0.3429} & 0.1946/0.2857 \\\\ \hline \end{array}$$ **Findings.** On Weather, CWRE exceeds ETran in **Kendall W at all three missing rates** and in **$\tau_w$ at two of three rates**. Traffic and Electricity are already included among the candidate source domains, and we recognize the value of further evaluating them as additional target domains. **(3) Longer Forecasting Horizon: 512→512** We extend the forecasting horizon from **96 to 512**, evaluating all four targets using two scoring architectures and three downstream evaluators under complete observations (MR = 0). $$\begin{array}{llcc} \hline \text{Setting} & \text{Method} & \tau_w & W \\\\ \hline 512\rightarrow96 & \text{CWRE} & \mathbf{0.3820} & \mathbf{0.3946} \\\\ & \text{ETran (Best Baseline)} & 0.1978 & 0.3277 \\\\ \hline 512\rightarrow512 & \text{CWRE} & \mathbf{0.0914} & \mathbf{0.2393} \\\\ & \text{MMD (Best Baseline)} & 0.0616 & 0.2268 \\\\ \hline \end{array}$$ **Findings:** CWRE achieves the highest overall $\tau_w$ and Kendall's $W$, providing further evidence that its ranking advantages extend beyond the original forecasting horizon.
Reviewer_pGAzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This papre proposes CWRE as a transferability estimation framework that selects suitable source datasets for TS forecasting without any target-domain fine-tuning. Unlike existing methofs that evaluate the compatibility between learned representations and forecasting windows, the proposed method capture the relationship between representation changes and the future-sequence changes acroess adjacent windows. A series of modules such as the combination of instance-wise and cross-window relations, trend-seasonal decomposition, spectral projection, and closed-form ridge regression are proposed to enhance transferability estimation. Experiments acoress multiple datasets and model architectures demonstrate that the proposed method achieve better overall aggreement with downstream transfer-performance rankings than existing baselines.

Critical strengths
  1. The method provides a novel perspective on transferability estimation. Different from existing methods that assess instance-wise representation compatibility, the proposed method CWRE explicitly examines whether changes in learned representations correspond to changes in the future sequences across adjacent forecasting windows.
  2. The authors conducted comprehensive empirical validation with sufficient ablation studies. Diverse target datasets, scoring architectures, downstream models and missing-observation settings are adopted. In addition, the ablation study demonstrates the contribution of individual components, providing empirical evidence that cross-window modelling contributes complementary information beyond instance-wise evaluation.
Critical weaknesses
  1. The method is not a completely unsupervised source-selection approach. It requires access to historical target training sequences with corresponding forecasting targets. Such as it would be interesting to compare it against a simpler frozen-encoder linear probing baseline, where a lightweight forecasting head is trained on the target training data and evaluated on a validation set. Such a comparison would help establish whether the proposed methods provide meaningful benefits beyond simple and conventional lightweight target-domain adaptation.
  2. The experiments cover three target datasets and six source datasets. This provides some evidence, but the scope is relatively narrow for a method intended to estimate transferability in TS forecasting generally. It is unclear whether the results extend to substantially different domains, such as weather, healthcare, traffic, or electricity markets.
Questions for authors
  1. I understand CWRE as a proxy evaluation method. However, it is unclear why directly evaluating frozen pretrained models on the target validation set would not provide a simpler and more straightforward source-selection strategy without requiring any fine-tuning. Could the authors clarify the practical limitations of this direct evaluation approach and demonstrate the advantages of CWRE over such a baseline?
2 comments
AuthorsOpenReview ↗

Response to Reviewer pGAz [Part 1]

Dear Reviewer,

Thank you for your thoughtful and constructive review. We sincerely appreciate your suggestions regarding CWRE's supervision setting, baseline comparisons, and experimental generalizability. In response, we have conducted additional experiments and clarified our methodology to address your concerns, as detailed below.

Response to W1 and Q1: Supervision Setting and Comparisons with Lightweight Adaptation and Direct Evaluation

Reviewer's Concern (W1): The method is not a completely unsupervised source-selection approach. It requires access to historical target training sequences with corresponding forecasting targets. It would be interesting to compare CWRE against a simpler frozen-encoder linear probing baseline, where a lightweight forecasting head is trained on target training data and evaluated on a validation set. Such a comparison would help establish whether CWRE provides meaningful benefits beyond conventional lightweight target-domain adaptation.

Reviewer's Question (Q1): Why not directly evaluate frozen pretrained models on the target validation set as a simpler source-selection strategy without fine-tuning? What are the practical limitations of this approach, and what advantages does CWRE offer over direct evaluation?

1. Clarification of the Supervision Setting

CWRE uses labeled target training sequences but does not claim completely unsupervised source selection. Instead, source-pretrained encoders remain frozen, and transferability scores are computed through closed-form linear estimation using target training data, without encoder fine-tuning or separate validation-set forecasting evaluation.

2. Comparisons with Linear Probing and Direct Frozen-Model Evaluation

  • Experimental Setup: Following the reviewer's suggestions, we introduce two baselines using the same source-pretrained models:

    • Linear Probing (LP): Fits a Ridge forecasting head on frozen representations using target training data.
    • Frozen Pretrained Model (FPM): Directly evaluates the complete source-pretrained model with its forecasting head, without target-domain adaptation.

Both rank sources by negative validation MAE:

skLP=−MAE⁡val(g^k(Fk(X)),Y)s_k^{\mathrm{LP}}=-\operatorname{MAE}_{\mathrm{val}}(\widehat{g}_k(F_k(X)),Y)

skFPM=−MAE⁡val(gksrc(Fk(X)),Y)s_k^{\mathrm{FPM}}=-\operatorname{MAE}_{\mathrm{val}}(g_k^{\mathrm{src}}(F_k(X)),Y)

Here, g^k\widehat{g}_k is the target-fitted head, while gksrcg_k^{\mathrm{src}} is the source-pretrained forecasting head.

We evaluate four targets, two scoring architectures, and three downstream evaluators (512→96, MR = 0), reporting weighted Kendall's τw\tau_w (averaged across evaluators) and Kendall's WW.

  • Results: Each entry reports τw/W\tau_w/W.
\hline \text{Architecture} & \text{Target} & \text{FPM} & \text{LP} & \text{CWRE} \\\\ \hline \text{PatchTST} & \text{ETTh1} & -0.2082/0.3071 & 0.2449/0.4786 & \mathbf{0.4807/0.5071} \\\\ & \text{ETTm1} & 0.1832/0.1929 & 0.1093/0.1929 & \mathbf{0.2281/0.2286} \\\\ & \text{Exchange} & 0.0571/0.2429 & -0.2109/0.1786 & \mathbf{0.3891/0.3786} \\\\ & \text{Weather} & 0.0104/0.2857 & -0.0490/0.3000 & \mathbf{0.3524/0.4071} \\\\ & \textbf{Average} & 0.0106/0.2572 & 0.0236/0.2875 & \mathbf{0.3626/0.3804} \\\\ \hline \text{TimeEmb} & \text{ETTh1} & 0.2000/0.4429 & 0.2757/0.4929 & \mathbf{0.6032/0.5786} \\\\ & \text{ETTm1} & -0.0463/0.1143 & 0.0707/0.1714 & \mathbf{0.2141/0.2500} \\\\ &\text{Exchange} & -0.1819/0.1643 & -0.0277/0.2357 & \mathbf{0.3224/0.3143} \\\\ & \text{Weather} & 0.3338/0.4714 & 0.2463/0.3929 & \mathbf{0.4662/0.4929} \\\\ & \textbf{Average} & 0.0764/0.2982 & 0.1413/0.3232 & \mathbf{0.4015/0.4090} \\\\ \hline \end{array}$$ - **Findings:** **CWRE outperforms both baselines on both metrics across all eight target–architecture settings**, providing consistent empirical evidence that relation-based evaluation yields more accurate source rankings than lightweight linear probing and direct frozen-model evaluation. **3. Why Relation-Based Evaluation?** **FPM** measures zero-shot forecasting performance, which is affected by the compatibility of the source-trained forecasting head with the target domain. **LP** reduces this dependency by fitting a target-domain head but requires validation-set evaluation to rank sources. In contrast, **CWRE jointly evaluates instance-wise and cross-window predictive relations through a shared linear mapping**, assessing whether frozen representations explain both future sequences and their temporal changes. **Together, these results demonstrate that CWRE provides more reliable source rankings than direct forecasting-performance-based approaches**, without target-domain encoder fine-tuning or separate validation-set forecasting evaluation. We will clarify the supervision setting and incorporate both baselines in Sections 3.2–3.4 (Tables 1, 3, and 4).
AuthorsOpenReview ↗

Response to Reviewer pGAz [Part 2]

Response to W2: Generalization Across Target Domains and Forecasting Horizons

Reviewer's Concern (W2): The experiments cover three target datasets and six source datasets. This provides some evidence, but the scope is relatively narrow for a method intended to estimate transferability in TS forecasting generally. It is unclear whether the results extend to substantially different domains, such as weather, healthcare, traffic, or electricity markets.

Our original experiments cover three targets, six candidate sources, two scoring architectures (PatchTST and TimeEmb), and three downstream evaluators (iTransformer, DUET, and FilterNet). They further examine three missing rates (MR = 0, 0.3, 0.6), with three masking seeds for incomplete observations, evaluating both cross-architecture ranking consistency and robustness to missing data.

To further validate CWRE across different domains and forecasting settings, we conduct two additional sets of experiments.

1. Generalization to a New Target Domain: Weather

  • Experimental Setup: We introduce Weather as a fourth target, extending the evaluation to meteorological forecasting alongside the original energy-related and financial targets. We evaluate 512→96 across all three missing rates, using both scoring architectures and three downstream evaluators.

  • Results: Each entry reports τw/W\tau_w/W, averaged across PatchTST and TimeEmb, with τw\tau_w further averaged across three downstream evaluators. MR = 0.3 and 0.6 additionally include three masking seeds.

\hline \text{MR} & \text{CWRE }(\tau_w/W) & \text{ETran }(\tau_w/W) \\\\ \hline 0 & \mathbf{0.4093/0.4500} & 0.2497/0.4179 \\\\ 0.3 & 0.3148/\mathbf{0.5072} & \mathbf{0.3631}/0.4964 \\\\ 0.6 & \mathbf{0.3209/0.3429} & 0.1946/0.2857 \\\\ \hline \end{array}$$ - **Findings:** On the additional Weather target, **CWRE outperforms ETran on both metrics at MR = 0 and 0.6, and achieves higher Kendall's $W$ at MR = 0.3**, supporting its effectiveness under both complete and incomplete observations. Detailed results for both scoring architectures (PatchTST and TimeEmb) are provided in Table 1 (MR = 0) and Appendix Table 9 (MR = 0.3 and 0.6). **2. Generalization to Longer Forecasting Horizons** - **Experimental Setup:** We extend the forecasting horizon from **96 to 512**, evaluating all **four targets and two scoring architectures** under complete observations (512→512, MR = 0). - **Results:** Averaged across four targets and both scoring architectures, with $\tau_w$ averaged across three downstream evaluators: $$\begin{array}{llcc} \hline \text{Setting} & \text{Method} & \tau_w & W \\\\ \hline 512\rightarrow96 & \text{CWRE} & \mathbf{0.3820} & \mathbf{0.3946} \\\\ & \text{ETran (Best Baseline)} & 0.1978 & 0.3277 \\\\ \hline 512\rightarrow512 & \text{CWRE} & \mathbf{0.0914} & \mathbf{0.2393} \\\\ & \text{MMD (Best Baseline)} & 0.0616 & 0.2268 \\\\ \hline \end{array}$$ - **Findings:** **CWRE achieves the highest ranking agreement on both metrics**, outperforming the strongest competing baseline under the extended forecasting horizon. This provides additional evidence of its effectiveness beyond the original 96-step forecasting setting. **Overall, the expanded experiments provide further evidence of CWRE's effectiveness across different target domains, missing-data conditions, scoring architectures, and forecasting horizons.** We will incorporate these results in Sections 3.2–3.4 (Tables 1, 2, 3, and 4). Since Table 3 reports cross-target averages, we update it to include Weather for all methods, while keeping the original three-target results and evaluation protocols unchanged. Traffic and Electricity are already included among the candidate source domains, and we recognize the value of further evaluating them and healthcare as additional target domains. We thank the reviewer again for these valuable suggestions. We will clarify the supervision setting, incorporate the additional baseline comparisons, and update the experimental results and corresponding analysis in the revised manuscript. We hope these responses address your concerns and welcome any further questions or suggestions. *Respectfully and sincerely,* **The Authors**
Reviewer_qDf3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CWRE for ranking source datasets in time series transfer learning without target-domain fine-tuning during scoring. It combines instance-wise and adjacent-window difference relations through shared ridge regression, with trend-seasonal decomposition and spectral projection. Experiments assess ranking agreement across scoring and downstream architectures, including missing-input settings.

Critical strengths

1.The paper introduces cross-window correspondence between representation changes and future changes as a useful signal for source selection. 2.The cross-architecture evaluation tests source selection with downstream forecasting models distinct from the scoring model. 3.The ablations show that cross-window relations complement instance-wise scoring across the three target datasets.

Critical weaknesses

1.The key technical challenges in cross-window transferability estimation and the rationale connecting these challenges to the proposed design are insufficiently developed. 2.The comparability of source scores is not validated when each source independently selects projection ratios and target scaling to maximize its training-fit score. 3.Evidence for general source-selection gains is limited to three targets and one forecasting horizon, with nine wins, six ties, and three losses against TEMPLATE across eighteen comparisons. 4.Figure 2 provides insufficient visual detail about the construction of cross-window relations and their integration with the trend and seasonal branches, obscuring the overall framework.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response To Reviewer qDf3 [Part 1]

Dear Reviewer,

We sincerely thank you for your careful review and constructive suggestions. Your comments highlight important aspects of our work, including the technical motivation and design rationale, source-score comparability, generalization across evaluation settings, and framework clarity. We appreciate these observations, which have helped us identify opportunities to strengthen both the presentation and experimental validation of CWRE.

Response to W1: Technical Challenges and Design Rationale

Reviewer's Concern (W1): The key technical challenges in cross-window transferability estimation and the rationale connecting these challenges to the proposed design are insufficiently developed.

We thank the reviewer for this comment. The design of CWRE is motivated by two practical challenges when using adjacent-window changes for transferability estimation.

Challenge 1: adjacent-window differences can be misleading because neighboring windows heavily overlap.

  • Evidence: In Appendix A, we construct an independent time-series counterexample with zero mean and variance σ2\sigma^2. Let $ f_t $ and yty_t denote the means of historical and future windows of lengths LL and HH, respectively. Although the underlying observations are independent and Cov⁡(ft,yt)=0\operatorname{Cov}(f_t,y_t)=0, their adjacent-window differences share boundary observations and satisfy a nonzero correlation: $ \operatorname{Corr}(\Delta f_t,\Delta y_t)=-\frac{1}{2}. $ This shows that cross-window correlation may arise purely from window overlap, even without a genuine predictive relationship.

  • Design Rationale: Therefore, CWRE does not use the cross-window relation in isolation. Instead, it jointly evaluates the original instance-wise relation and the cross-window relation using a shared linear mapping. A candidate source is favored only when its representations can explain both the target futures themselves and their corresponding changes across adjacent windows, reducing the influence of overlap-induced correlations that are not predictive of the target task.

Challenge 2: differencing operation amplifies high-frequency noise.

  • Evidence: Appendix B.2 shows that, under temporally independent perturbations, first-order differencing doubles the expected noise energy. Appendix C.5 further confirms this effect empirically, with the noise MSE increasing by 1.997× for PatchTST and 2.000× for TimeEmb.

  • Design Rationale: To mitigate this effect, CWRE applies cross-window evaluation to the smoother trend component, where local evolution is less affected by high-frequency perturbations, while retaining complementary predictive information from the non-trend component through instance-wise evaluation.

Response to W2: Clarification of Source-Score Comparability and Shared Hyperparameter Selection

Reviewer's Concern (W2): The comparability of source scores is not validated when each source independently selects projection ratios and target scaling to maximize its training-fit score.

We apologize that our description may have led to this misunderstanding.

First, there is no source-specific hyperparameter selection during source pretraining. For a given scoring architecture, all six candidate sources are pretrained using exactly the same architecture, optimization settings, and training protocol. Thus, the source-pretrained models differ only in the source dataset used for pretraining.

Second, CWRE scoring also uses a shared hyperparameter configuration across all six candidate sources. For each target dataset, missing rate, and scoring architecture, the same (ω,ρtr,ρse)(\omega,\rho_{\mathrm{tr}},\rho_{\mathrm{se}}) is applied to every candidate:

Sk=T(Fk,Y;ω∗,ρtr∗,ρse∗).S_k = \mathcal{T}(F_k,Y;\omega^{\ast},\rho_{\mathrm{tr}}^{\ast},\rho_{\mathrm{se}}^{\ast}).

No per-source score maximization is performed. All candidates therefore share the same target data, scoring objective, and scoring hyperparameters, while differing only in their source-pretrained representations.

We will clarify this protocol in Section 3.1 (Experimental Setup) and correct Appendix C.1.2 (Architectures and Source-Ranking Protocol), including Eq. (44) to make this protocol explicit.

AuthorsOpenReview ↗

Response To Reviewer qDf3 [Part 2]

Response to W3: Broader Target and Horizon Evaluation, and Source-Selection Gains over TEMPLATE

Reviewer's Concern (W3): Evidence for general source-selection gains is limited to three targets and one forecasting horizon, with nine wins, six ties, and three losses against TEMPLATE across eighteen comparisons.

We thank the reviewer for suggesting broader evaluation. Our original experiments already cover six candidate sources, two scoring architectures, three downstream evaluators, and three missing rate settings. We further extend the evaluation along three dimensions.

1. Additional target dataset

We add Weather as a fourth target, evaluating both scoring architectures across three missing rates . Results are averaged over PatchTST and TimeEmb, with τw\tau_w additionally averaged across three downstream evaluators.

\hline \text{Missing Rate} & \text{CWRE }(\tau_w/W) & \text{ETran }(\tau_w/W) \\\\ \hline 0 & \mathbf{0.4093/0.4500} & 0.2497/0.4179 \\\\ 0.3 & 0.3148/\mathbf{0.5072} & \mathbf{0.3631}/0.4964 \\\\ 0.6 & \mathbf{0.3209/0.3429} & 0.1946/0.2857 \\\\ \hline \end{array}$$ **Findings:** CWRE outperforms ETran in Kendall's $W$ across all three missing rates and in $\tau_w$ in two of three settings. Although DTW achieves a higher $W$ at MR = 0, CWRE maintains competitive ranking agreement across the evaluated missing conditions. **2. Longer Forecasting Horizon: 512→512** We extend the forecasting horizon from **96 to 512**, evaluating all four targets using two scoring architectures and three downstream evaluators under complete observations (MR = 0). $$\begin{array}{llcc} \hline \text{Setting} & \text{Method} & \tau_w & W \\\\ \hline 512\rightarrow96 & \text{CWRE} & \mathbf{0.3820} & \mathbf{0.3946} \\\\ & \text{ETran (Best Baseline)} & 0.1978 & 0.3277 \\\\ \hline 512\rightarrow512 & \text{CWRE} & \mathbf{0.0914} & \mathbf{0.2393} \\\\ & \text{MMD (Best Baseline)} & 0.0616 & 0.2268 \\\\ \hline \end{array}$$ **Findings:** CWRE achieves the highest overall $\tau_w$ and Kendall's $W$, providing further evidence that its ranking advantages extend beyond the original forecasting horizon. **3. Source-selection gains over TEMPLATE** We clarify that the originally reported 9 wins, 6 ties, and 3 losses against TEMPLATE refer only to the downstream performance of the Top-1 selected source, rather than the quality of the full ranking. For the complete source ranking, CWRE achieves (0.3820/0.3947) in $(\tau_w/W)$, compared with (-0.0964/0.2107) for TEMPLATE. We further evaluate Top-$K$ coverage of the actual best source: $$\begin{array}{ccc} \hline \text{Selection Budget} & \text{CWRE} & \text{TEMPLATE} \\\\ \hline \text{Top-1} & \mathbf{8/18} & 5/18 \\\\ \text{Top-2} & \mathbf{10/18} & 7/18 \\\\ \text{Top-3} & \mathbf{12/18} & 8/18 \\\\ \hline \end{array}$$ CWRE achieves higher best-source coverage at all three budgets, showing that its ranking is more useful for narrowing the candidate pool under limited fine-tuning budgets. The additional target and horizon evaluations are included in **Sections 3.2–3.4 (Tables 1, 2, 3, and 4)**. We further incorporate the **Top-K source-selection analysis into Appendix C.7**. Since Table 3 reports cross-target averages, we update it to include Weather for all methods, while keeping the original three-target results and evaluation protocols unchanged. ### Response to W4: Clarifying Cross-Window Construction and Trend–Seasonal Integration in Figure 2 > **Reviewer's Concern (W4):** Figure 2 provides insufficient visual detail about the construction of cross-window relations and their integration with the trend and seasonal branches, obscuring the overall framework. We agree that **Figure 2 should more explicitly illustrate how cross-window relations are constructed and integrated into the two branches of CWRE**. We will revise Figure 2 to explicitly show: - **Cross-window construction:** Adjacent-window pairing and the computation of representation differences $\Delta F_t$ and corresponding future-target differences $\Delta Y_t^{\mathrm{tr}}$. - **Branch-specific evaluation:** Illustrate how the **Trend branch jointly evaluates instance-wise and cross-window relations using a shared linear mapping**, while the **Seasonal branch uses only instance-wise evaluation**, without differencing. - **Score fusion:** Show how the two branch scores are combined through **energy-based weighting** to obtain the final transferability score. We will also update the caption and discription in **Section 2.2**, making the data flow and its correspondence to **Sections 2.3–2.5** explicit. We hope our responses and additional evidence adequately address your concerns. We would be happy to provide further clarification or discuss any remaining questions. Respectfully and sincerely, **The Authors**
AuthorsOpenReview ↗

Follow-up on Reviewer Concerns

Dear Reviewer qDf3,

We sincerely appreciate your time and thoughtful feedback. We would be grateful to know whether our responses have adequately addressed your concerns. If any questions remain, we would be more than happy to provide further clarification or discussion.

Thank you again for your valuable time and consideration.

Respectfully and sincerely,

The Authors of Submission 39941

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair, Senior Area Chair, Program Chairs, and Reviewers,

We sincerely thank all reviewers for their careful evaluation and constructive feedback.

Research significance and core contribution. Transferability across time-series datasets is important but remains insufficiently understood: given multiple candidate sources, it is difficult to know which source will benefit a target forecasting task without costly adaptation and evaluation. To the best of our knowledge, prior work has not systematically studied dataset-level transferability estimation specifically for time-series forecasting.

Motivated by the temporal structure of forecasting data, CWRE introduces cross-window evolution correspondence as a time-series-specific transferability criterion, extending evaluation from individual-window compatibility to whether predictive relations remain consistent as forecasting windows evolve. At a high level, this provides an effective basis for source-data selection and efficient reuse of pretrained time-series models, without exhaustively fine-tuning every candidate source.

Reviewer assessments. All four reviewers recognized the paper's informativeness and rated their soundness concerns as minor or fixable, with two also rating the presentation as clear and easy to follow. Their feedback primarily focuses on clarifying methodological details, specifying experimental protocols, and strengthening fine-grained experimental validation. We have provided targeted clarifications and additional experimental evidence in response, as summarized below.


Methodological motivation and framework clarity (9nq7 W1/Q1; qDf3 W1/W4): We explain the motivation and implementation of cross-window evaluation, including how shared mapping reduces the influence of overlap-induced correlations and trend–seasonal decomposition mitigates noise amplification. Theoretical analyses and ablations support these designs (Sections 2.3–2.5, Table 2, Appendices A and B.2). We will also improve the illustration in Section 2.2 and Figure 2.

Evaluation protocol and baseline fairness (qDf3 W2; eeXa W1/Q1; 9nq7 W3/Q2; pGAz W1/Q1): We clarify that no source-specific hyperparameter maximization is performed, with all candidate sources evaluated under a shared configuration (Section 3.1, Appendix C.1.2). Additional baseline tuning, hyperparameter distribution analyses, and comparisons with Linear Probing and Frozen Pretrained Model further assess fairness and ranking performance (Table 1, Appendices C.8–C.9).

Broader experimental validation (qDf3 W3; pGAz W2; eeXa W2.1/Q3; 9nq7 W2/Q4): We add Weather as a fourth target and evaluate two forecasting horizons (512→96 and 512→512). The results provide additional evidence across different target domains, missing rates, and forecasting horizons (Sections 3.2–3.4, Tables 1–4). We are also actively extending our evaluation to additional target datasets.

Ranking consistency and statistical analysis (eeXa W2.2/Q2; qDf3 W3): We provide detailed per-architecture results (Appendix C.3.2, Table 7) and Top-K analyses demonstrating improved best-source coverage over TEMPLATE (Appendix C.7.2, Figure 13). Additional target-level bootstrap confidence intervals and exact paired permutation tests further characterize the positive average gains and their statistical uncertainty.

Temporal robustness and noise suppression (9nq7 W4/Q3; eeXa Q4; qDf3 W1): Additional temporal-gap evaluations and controlled overlap-correlated perturbations provide evidence that CWRE's benefits do not rely solely on window overlap (Section 3.5, Appendices C.4 and C.8). Extended matched-random projection experiments across 24 settings further support trend projection's noise-suppression benefits (Appendix C.6, Figure 8).


Reviewer follow-up. We appreciate that the reviewers generally recognized the importance of the problem and the value of CWRE's cross-window perspective. Reviewer eeXa explicitly acknowledged that our additional experiments and clarifications addressed their methodological questions, with no further methodological concerns, while Reviewer 9nq7 also acknowledged that several concerns were addressed.

Overall, the remaining discussion mainly focus on implementation details and finer-grained empirical validation, rather than the core motivation or contribution of introducing a time-series-specific transferability criterion. We hope the committee will consider this central contribution together with the additional evidence provided during rebuttal.

All the revised content has been marked in blue\textcolor{blue}{\text{blue}}.

We sincerely appreciate the reviewers' constructive suggestions, which have helped strengthen the presentation and empirical validation of CWRE. We hope this summary assists the committee in its assessment.

Respectfully and sincerely,

The Authors of Submission 39941

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.