acceptodds
Under review as a conference paper at ICLR 2027

DAD4TS: Data-Augmentation-Oriented Diffusion Model for Time-Series Forecasting with Small-Scale Data

Abstract

Small-scale data is a critical problem in time-series forecasting tasks. Data augmentation is an effective strategy for this task, but it has a limitation in generating meaningful data. To address this limitation, we propose DAD4TS, a diffusion-model-based data augmentation method, designed for time-series forecasting with small-scale data. In DAD4TS, to generate samples that improve the predictive accuracy of a time-series forecasting model, the generator and the time-series forecasting model are trained simultaneously, and the data generator is updated while calculating the usefulness of the generated data. Furthermore, to support small-scale data, we use mathematical methods instead of conventional VAE methods to train the diffusion model by projecting the time-series data into the latent space. We validated the effectiveness of DAD4TS with eight comparative methods through qualitative and quantitative experiments on six real-world datasets and six time-series models. As a result, DAD4TS showed improvements across multiple datasets.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, data augmentation, diffusion models, small- scale data

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_g9YXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DAD4TS, a diffusion-based data augmentation framework for time-series forecasting with limited training data. It jointly trains a generator and a forecasting model, using prediction errors to guide sample generation and Thompson sampling to select augmentation strategies. The framework also introduces a training-free latent transformation and rectified flow for efficient generation. Experiments on six datasets with six forecasting models show an overall RMSE improvement of 12.05%, outperforming several existing augmentation methods.

Critical strengths
  1. The paper addresses a practical problem. Improving forecasting performance with limited training data is meaningful, and task-oriented data generation is a reasonable direction.

  2. I appreciate the adaptive augmentation design. Updating the generator alongside the forecaster allows sample generation to respond to the current training state.

  3. The experimental coverage is fairly broad, including six datasets, six forecasting models, and several recent augmentation baselines. The analysis of window normalization also provides some useful observations.

Critical weaknesses
  1. Novelty is somewhat limited. AutoDA-Timeseries and ReAugment already explore forecasting-guided augmentation. DAD4TS mainly combines related ideas with diffusion generation and Thompson sampling.

  2. The contribution of individual modules is not sufficiently clear. More direct ablations would help identify which components actually drive the improvements.

  3. The gains are uneven across forecasting models. Most improvements come from models without window normalization, while the benefits for stronger normalized models are relatively small.

  4. The augmentation policy requires additional validation rollouts during training. I am concerned about the computational overhead and possible overfitting to the small validation set.

Questions for authors
  1. What is the main methodological difference between DAD4TS and AutoDA-Timeseries beyond using a diffusion generator?

  2. Could the authors provide ablations for Thompson sampling, prediction-error weighting, and the training-free latent transformation?

  3. How does DAD4TS compare with AutoDA under the same computational budget? Are the improvements over AutoDA statistically significant?

  4. How do the authors prevent repeated validation-based augmentation selection from overfitting, especially on very small datasets?

4 comments

We sincerely appreciate the reviewer’s insightful comments and constructive suggestions. All new experiments cover the full grid (6 datasets, 6 TSF models, 4 horizons, 3 seeds); p-values are from Wilcoxon signed-rank tests on log RMSE.

Q1 / W1: Difference from AutoDA-Timeseries (and ReAugment).

They learn how to transform existing training windows. In contrast, DAD4TS learns a generator of new windows, weighted by the TSF model's current errors and regenerated every epoch. To isolate this, we kept DAD4TS's data processing and changed only how new windows are obtained. The results are as follows:

How new windows are obtainedRMSE change (%)
Processing applied to real windows, no generator−0.89
Generator trained once, sampled once (as comparison methods do)−7.55
Generator trained and resampled with the TSF model (DAD4TS-U; see Q2 / W2)−12.05
Reference: AutoDA-Timeseries−9.83

We confirmed that generating windows yields better performance than simply transforming existing windows. Furthermore, coupling the generator with the TSF model adds a significant 4.5 points (p = 2e-12). The difference is also statistically significant when compared to AutoDA-Timeseries (p = 2.4e-7).

Q2 / W2: Ablations of the three modules.

We removed or replaced each module, and built the latent transformation up step by step. Since the label sampler was ineffective, we call the variant that draws processing labels uniformly at random every epoch DAD4TS-U (same −12.05%), and used it as the base of all other ablations. The results are as follows:

VariantRMSE change (%)p vs DAD4TS-U
DAD4TS-U−12.05–
DAD4TS (submitted)−12.050.42
No TSF-model error weighting−11.600.12
Latent: none (raw windows)+2.114e-29
Latent: per-window normalization only−6.687e-26
Latent: + orthonormal basis, no standardization−6.025e-29
  • The latent transformation is essential; without it, augmentation hurts (+2.11%). Per-window normalization brings the gain to −6.68%, and only with the per-coordinate standardization on the orthonormal basis does it reach −12.05%.
  • Error weighting improves accuracy overall across our experiments (−12.05% vs −11.60%), but not significantly; we will present it as a modest component.
  • Thompson sampling gives no significant benefit: replacing it with uniformly random labels (DAD4TS-U) leaves accuracy unchanged (p = 0.42), so we will recommend DAD4TS-U and move the label sampler to the appendix.

W3: Uneven gains across TSF models.

iTransformer/PatchTST/TimesNet normalize each input window, removing the level and scale variation that DAD4TS generates; copying level and scale from the real window costs 12.65 points on RNN/LSTM/Transformer but only 0.95 points on these models. The comparison methods show the same pattern (AutoDA-Timeseries −12.74% → −1.94%; others near zero or worse). DAD4TS-U still gives the largest gain there (−2.77%), though not significantly above AutoDA-Timeseries (p = 0.60). We will present this as a limitation with its explanation.

W4 / Q4: Validation rollouts and overfitting.

We appreciate this well-founded concern, which changed our recommended configuration. If labels are drawn uniformly at random each epoch (DAD4TS-U), no rollouts are needed and validation data serve only for early stopping, as for every comparison method. Accuracy stays the same (DAD4TS-U −12.047% vs DAD4TS (submitted) −12.051%, p = 0.42), and the label sampler reduces validation MSE by only 1.8%. Given the cost (Q3) and fair use of validation data, we adopt DAD4TS-U as the recommended configuration.

(Continued in the next comment.)

(Continued from the previous comment.)

Q3: Same budget and significance vs AutoDA-Timeseries.

DAD4TS-U uses validation data only for early stopping, exactly like AutoDA-Timeseries. Under this matched budget, we compared accuracy and each method's own cost (training and generation; German, one T4 GPU, one job at a time, per TSF model and horizon; the total including the TSF model depends on early stopping and is a reference). The results are as follows:

MethodRMSE change (%)p vs DAD4TS-UOwn cost (s)Total, reference (s)
ReAugment−1.412e-2628.544.4
TimeVAE−1.722e-2852.066.7
AutoDA-Timeseries−9.832.4e-762.879.0
DAD4TS (submitted)−12.050.42401.3420.4
DAD4TS-U−12.05–27.246.5

Overall, DAD4TS-U is significantly more accurate than AutoDA-Timeseries at a lower cost. However, on the normalizing TSF models alone (as split in Table 2 of our paper), the improvement over AutoDA-Timeseries is not significant (−2.77% vs −1.94%, p = 0.60).

Revisions. We will make DAD4TS-U the recommended configuration, add the ablation, cost and significance results and the training protocol, and move the label sampler and its diagnostics to the appendix.

The additional ablations and efficiency comparisons substantially address my concerns, particularly by demonstrating the value of joint generator–forecaster training and eliminating costly validation rollouts; I will raise my rating from 2 to 3.

AuthorsOpenReview ↗

Thank you for acknowledging our rebuttal efforts.

Dear Reviewer g9YX,

We would like to thank you once again for your efforts and constructive comments. We would be happy to address any further questions you may have.

Best regards,

Authors

Reviewer_3u78OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an AutoDA like data augmentation method that creates data during training by selects standard augmentation methods, embeds them into latent space, then using a diffusion model (Rectified Flow) to generate data. The key point is that during training the generator is trained on a pool of samples.

Critical strengths

The proposed method is relevant, seems sound, and is reasonable. At a glance, the results look great, because of the different datasets and RMSE differences, but there are issues with that, see below.

Critical weaknesses
  1. Insufficient evaluation. Very little ablation is performed. THe proposed method has a lot of parts (Thompson sampler, different preprocessing methods, diffusion model, embedding, rollout, losses, DA amount, etc) but the only ablation is whether using a foundation model is used or not. In addition, there are many adversarial data augmentation (ADA) time series recognition methods that have a similar goal as the proposed method of "not aim to generate data that closely resembles real data; rather, it focuses on generating data that is difficult for predictive models to learn from". The authors don't compare against any ADA methods.

  2. I have concerns about actual usefullness of the proposed method. At a glance, the results look amazing. But, then when you look at the details (Table 2 and the Appendix tables), you find that the gains are all due to the old/poor models (RNN, LSTM, vanilla Transformer). On actual modern time series networks (iTransformer, PatchTST, TimesNet) that most people use, the results are about the same as all of the comparative methods.

Questions for authors

Clarification requests:

  1. Details of the training, especially about the competitors, are missing. To my understanding, the proposed method adds synthetic samples at the end of each epoch. 64 synthetic samples are used for evaluation, but it doesn't say how many syntetic samples are used for training. The same goes for the "without augmentation" case or the competitors case. The problem is, let's say the training set size is doubled, then each epoch is trained for longer than "without augmentation". This matters because the proposed method would have twice as many back propagation updates as "without augmentation". It would be more fair to compare to "identity" than purely without augmentation. Also, about the competitors, it's also not described about how much data augmentation is used and when. The proposed method generates new synthetic samples every epoch, does the same apply to all of the competitors? If not, then it''s not fair.

  2. I have concerns about the use of foundation models (Chronos, Tirex) and the conditioning vector for the model. The problem is how can the authors be sure there is no data leakage? The authors use very public datasets and how can they be sure that they weren't used in the training of the foundation models?

  3. Can the authors clarify the computational cost of the proposed method and competitors? In line 299, the authors say that they train a copy of the forcasting model on the pool of windows. If this happens at the end of every epoch, it seems like it would be very computationally expensive.

Minor suggestions:

  1. In line citations are often not used properly.

  2. The AI statement is the template/instructions.

  3. Fig 1 feels like filler, especially since most of the figure can be found elsewhere in the paper.

5 comments

We sincerely thank the reviewer for the critical and insightful questions. All new experiments cover the full grid (6 datasets, 6 TSF models, 4 horizons, 3 seeds); p-values are from Wilcoxon signed-rank tests on log RMSE.

W1: Ablation of every component.

We ablated each listed component, one at a time ("Latent": the embedding). As a result, DAD4TS-U becomes the default: it draws each window's processing label uniformly at random every epoch, needs no label sampler or rollouts, and uses validation data only for early stopping, like every comparison method. The results are as follows:

VariantRMSE change (%)p vs DAD4TS-U
DAD4TS-U (default)−12.05–
DAD4TS (submitted)−12.050.42
Identity (real windows as synthetic data)−1.016e-29
No generator: processing on real windows−0.898e-24
No generator, each processing operation alone−0.05 to −1.33< 1e-24
Static (generator trained and sampled once)−7.552e-12
No TSF-model loss weighting−11.600.12
No TSFM conditioning−9.932e-6
Latent: none (raw windows)+2.114e-29
Latent: per-window normalization only−6.687e-26
Latent: + orthonormal basis, no standardization−6.025e-29
Latent: level/scale copied from real data−2.545e-17

The gain comes from the generator, its joint updates with the TSF model, and the latent normalization and standardization; the label sampler and rollouts do not help, and loss weighting is not significant.

W1: ADA methods.

We ran two representative ADA methods with their published settings (ADA [1]: γ=1, η=1, T_min=100, T_max=15, K=2; MaxUp [2]: m=4). The results are as follows:

MethodRMSE change (%)p vs DAD4TS-U
ADA+0.172e-35
MaxUp−1.492e-25
DAD4TS-U−12.05–

With small-scale data, making samples hard for the current model helps little; learning new windows does.

W2: Usefulness on modern TSF models.

We agree that the gains are concentrated on RNN/LSTM/Transformer.

  • Why. Since iTransformer/PatchTST/TimesNet normalize each input window, the differences in level and scale between windows generated by DAD4TS are less likely to be passed on to these models. In an experiment where the level and scale are copied from the real window rather than generated, the gain shrank by 12.65 points on RNN/LSTM/Transformer but by only 0.95 points on these models.
  • Other methods show the same pattern. On the normalizing models AutoDA-Timeseries's gain also drops sharply, and the other comparison methods have almost no effect or even hurt:
MethodRNN/LSTM/TransformeriTransformer/PatchTST/TimesNet
ReAugment−1.90−0.06
TimeVAE−2.55+0.56
DiffAT−1.17+0.34
AutoDA-Timeseries−12.74−1.94
DAD4TS-U−15.46−2.77
  • Significance vs AutoDA-Timeseries. On RNN/LSTM/Transformer, DAD4TS-U beats AutoDA-Timeseries by 2.72 points, which is significant (p = 5.4e-10). On iTransformer/PatchTST/TimesNet, DAD4TS-U still gives the largest gain, but its 0.83 point margin over AutoDA-Timeseries is not significant (p = 0.60). We will state this limitation in the paper.
  • Our claim. Generating data dynamically, in step with the TSF model's training, instead of once before training as previous methods do, yields data that improve accuracy more than those methods. As the table shows, this ranking also holds on the normalizing models.

(Continued in the next comment.)

(Continued from the previous comment.)

Q1: Training details and fairness.

We will add the full protocol to the paper.

  • Amount of synthetic data. Synthetic windows are made from the training windows, and there are as many of them as training windows. This holds for the comparison methods and for DAD4TS alike.
  • Number of TSF-model updates. Every method, including no augmentation, performs the same number of updates per epoch: one per real mini-batch. With augmentation, each update uses one real and one synthetic mini-batch.
  • When data are generated. The comparison methods generate their data once, before the TSF model is trained. DAD4TS updates its generator and regenerates the whole synthetic set at every epoch.
  • The 64 windows in the paper are the pool used to score labels during the rollouts, not training data. DAD4TS-U has no rollouts at all.
  • Two controls for your concerns (table in W1, ablation of every component). The "identity" row replaces the synthetic data by the real windows, with the same number of updates and samples per update, and reaches only −1.01%. The "static" row generates once, like the comparison methods, and reaches −7.55%. Regenerating at every epoch therefore adds a significant 4.5 points (p = 2e-12).

Q2: Data leakage through the TSFMs.

Five datasets (Inventories, German, Consumption, Employees, Forest Fires) were introduced by TempusBench [3] as absent from existing TSFM pretraining corpora. Neither these nor ILI is listed in the training data of TiRex [4] (Chronos datasets, GIFT-Eval pretraining data; Tables 5–6) or Chronos [5] (Table 3). Test windows are never passed to the TSFMs.

Q3: Computational cost.

As you pointed out, copying the TSF model itself is inexpensive, but in the submitted version, we trained a copy for each epoch (100 updates for each of the five candidate labels), and these rollouts dominated its cost. DAD4TS-U does not use them. We are grateful that your comment led us to review DAD4TS and greatly reduce its cost. Own cost (training and generation) on German (one T4 GPU, one job at a time, per TSF model and horizon; the total including the TSF model depends on early stopping and is a reference):

MethodOwn cost (s)Total, reference (s)
ReAugment28.544.4
TimeVAE52.066.7
AutoDA-Timeseries62.879.0
DAD4TS-U27.246.5
DAD4TS (submitted)401.3420.4

Minor. We will fix the in-line citations, replace the AI-use statement (template text left in by mistake), and revise Figure 1, considering merging it with Figure 2.

References (numbered for this response only; they differ from the numbering in our paper):
[1] Volpi et al., Generalizing to Unseen Domains via Adversarial Data Augmentation, NeurIPS 2018.
[2] Gong et al., MaxUp: Lightweight Adversarial Training with Data Augmentation Improves Neural Network Training, CVPR 2021.
[3] Goktas et al., TempusBench: An Evaluation Framework for Time-Series Forecasting, Recent Advances in Time Series Foundation Models Have We Reached the 'BERT Moment'?, 2025.
[4] Auer et al., TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning, NeurIPS 2025.
[5] Ansari et al., Chronos: Learning the Language of Time Series, TMLR 2024.

I've read the responses and some of the issues are addressed, but I still cannot recommend acceptance.

  1. I am satisfied with the new ablations and happy that the authors could improve their method based on them. However, the rebuttal period isn't for improving your method, it's for clearing up confusions. The ablation should have been done before the paper submission and this issue should have been found before. It really weakens the paper when one part of the author's claims turns out to have no difference in the end.

  2. The authors justify by saying that the results are lessened on modern transformers. There are two problems with this. First, while TimesNet, iTrasformer, and PatchTST are more modern time series transformers, they are all from 2022 and 2023. Second, is the contribution really that high if it only works on old models? Even now, people are moving away from the old models. I can't remember the last time a vanilla LSTM was competitve, nevermind a vanilla RNN. How useful would the proposed method be if no one uses those models? Another key thing is that time series recongnition has moved towards patch-based models (or frequency based models). The common thing among the models that the proposed method worked well on was element based models (RNN, LSTM, vanilla Transformer).

  3. While the authors did respond, the manuscript hasn't been fixed. With the deadline for rebuttal ending in a few hours, I cannot judge the paper to change my score because the paper is still old and still has the issues I commened on previously.

Reviewer_uJC9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DAD4TS, a diffusion-based data augmentation framework for time-series forecasting with limited training data. It combines training-free latent transformation, rectified flow, foundation-model conditioning, and forecasting-guided augmentation. The generator and forecaster are jointly updated, while Thompson sampling selects augmentation strategies based on validation performance. Experiments across six datasets and six forecasting models report improvements over existing augmentation methods.

Critical strengths
  • Improving time-series forecasting under data scarcity is a practically relevant problem.
  • Optimizing synthetic data for downstream forecasting performance rather than distribution fidelity is a reasonable direction.
  • The paper evaluates multiple forecasting architectures and augmentation baselines, including the recent AutoDA-Timeseries.
Critical weaknesses
  1. Limited novelty and unclear necessity of the complex framework. The proposed method combines several existing techniques, including rectified flow, conventional time-series transformations, foundation-model conditioning, and Thompson sampling. It remains unclear whether diffusion-based generation is necessary compared with simpler task-aware augmentation strategies.

  2. Insufficient ablation studies. The paper only provides a limited ablation of foundation-model conditioning. Critical comparisons are missing, including augmentation without diffusion, static versus dynamically updated generation, and predictor-loss weighting versus uniform training. Consequently, the contribution of the central design choices cannot be established.

  3. Potentially unfair optimization budgets. DAD4TS repeatedly evaluates augmentation strategies using validation MSE after additional forecaster training rollouts. This introduces substantial validation-guided optimization that may not be available to the baselines. The paper does not control for this additional optimization budget or investigate overfitting to the small validation sets.

  4. Unclear practical efficiency. The framework requires pretrained foundation-model embeddings, a generative model, repeated synthetic data generation, and multiple forecasting rollouts. Despite this complexity, the overall improvement over AutoDA-Timeseries is relatively modest (−12.05% versus −9.83% relative RMSE change). No end-to-end runtime or computational cost comparison is provided.

  5. Limited generality and incomplete statistical evidence. The experiments rely on six relatively small datasets, mostly univariate series. Performance improvements are concentrated in forecasting models without internal window normalization, while gains on models with normalization are considerably smaller. Statistical significance is reported against no augmentation rather than directly against the strongest competing method, and the temporal ordering of the Forest dataset requires clarification.

Questions for authors
  1. How does DAD4TS perform when the diffusion generator is replaced with the same conventional transformations and augmentation-selection mechanism?

  2. What are the individual contributions of dynamic generator updates, predictor-loss weighting, and Thompson sampling?

  3. How does the method compare with AutoDA-Timeseries under matched computational and validation-feedback budgets? Is the performance difference statistically significant?

  4. What are the total training time and GPU costs? How sensitive are the results to the number of validation rollouts and the amount of available training data?

  5. Does the Forest dataset have a meaningful chronological ordering for forecasting evaluation?

4 comments

We thank the reviewer for the careful review and for questions that directly improved the method. All new experiments cover the full grid (6 datasets, 6 TSF models, 4 horizons, 3 seeds); p-values are from Wilcoxon signed-rank tests on log RMSE.

Q1 / W1: Replacing the generator with the same transformations and selection mechanism.

We removed the generator and applied DAD4TS's four processing operations, selected by its label sampler, directly to the real windows. We also tested each processing operation alone. The results are as follows:

VariantRMSE change (%)
DAD4TS (submitted)−12.05
Same processing operations + submitted label sampler, no generator−1.10
Each processing operation alone, no generator (4 variants)−0.05 to −1.33

These results confirm that the improvement is attributable to the generator.

W3: Validation-guided optimization and overfitting.

This concern was well founded, and reviewing and testing DAD4TS in response to it led us to change our recommended configuration. With processing labels drawn uniformly at random every epoch (we call this DAD4TS-U), there are no rollouts, and the validation data are used only for early stopping, exactly as for every comparison method. Accuracy is unchanged (DAD4TS-U −12.047% vs DAD4TS (submitted) −12.051%, p = 0.42). We also confirmed that the label sampler lowers the validation MSE by only 1.8% relative to uniform labels. Considering both the cost (see our answer to Q3 / Q4 / W4) and a fair use of the validation data, we will make DAD4TS-U the recommended configuration and move the label sampler to the appendix.

Q2 / W2: Contributions of dynamic generator updates, loss weighting and Thompson sampling.

We removed each component in turn; as the label sampler had no effect, the other two were removed from DAD4TS-U (uniform labels, see W3; same −12.05%). The results are as follows:

Component removedRMSE change (%)p vs DAD4TS-U
Thompson sampling (= DAD4TS-U)−12.05–
Dynamic generator updates−7.552e-12
TSF-model loss weighting−11.600.12
  • Thompson sampling does not contribute significantly.
  • Dynamic updates contribute most. Training the generator once and sampling once, as the comparison methods do, loses 4.5 points. Based on this, the primary capability that enables DAD4TS is dynamic generation tailored to the training of the TSF model.
  • Loss weighting improves accuracy overall, but not significantly; we will present it as a modest component.

Q3 / Q4 / W4: Matched budgets, significance and cost.

DAD4TS-U uses validation data only for early stopping, exactly like AutoDA-Timeseries. Under this matched budget, we compared accuracy and each method's own cost (training and generation; German, one T4 GPU, one job at a time, per TSF model and horizon; the total including the TSF model depends on early stopping and is a reference):

MethodRMSE change (%)p vs DAD4TS-UOwn cost (s)Total, reference (s)
ReAugment−1.412e-2628.544.4
TimeVAE−1.722e-2852.066.7
AutoDA-Timeseries−9.832.4e-762.879.0
DAD4TS (submitted)−12.050.42401.3420.4
DAD4TS-U−12.05–27.246.5

DAD4TS-U is significantly better than AutoDA-Timeseries at a lower cost, as it needs no rollouts and no separate fixed-length generator training.

Sensitivity. Accuracy is unchanged without rollouts, so their number does not matter. A training-data-size study could not be completed in the discussion period; we will add it to the camera-ready version.

(Continued in the next comment.)

(Continued from the previous comment.)

Q5 / W5: Normalizing TSF models, statistics, Forest.

  • Normalizing TSF models. After data processing, DAD4TS generates not only the shape of each window but also its level and scale. However, iTransformer/PatchTST/TimesNet normalize every input window, so most of this level and scale variation is removed. Indeed, if the level and scale are copied from the real window instead of generated, the gain falls by 12.65 points on RNN/LSTM/Transformer but by only 0.95 points on the normalizing models. The smaller gain is not specific to DAD4TS, however. AutoDA-Timeseries's gain shrinks from −12.74% to −1.94%, and the other comparison methods give roughly zero or even hurt (ReAugment −1.90% → −0.06%, TimeVAE −2.55% → +0.56%). Our claim is that, unlike previous methods, DAD4TS generates data dynamically, following the TSF model's training, and that these data improve forecasting accuracy more than those of previous methods. Even on the normalizing models, DAD4TS-U gives the largest gain of all methods (−2.77%). Nevertheless, we recognize the smaller gain on the normalizing models as a limitation of our study and will state it as such in the paper.

  • Significance vs AutoDA-Timeseries. We compared DAD4TS-U with AutoDA-Timeseries, the strongest comparison method, separately for each type of TSF model. On RNN/LSTM/Transformer, DAD4TS-U reaches −15.46% and AutoDA-Timeseries −12.74%; the 2.72-point gap is significant (p = 5.4e-10). On iTransformer/PatchTST/TimesNet, the figures are −2.77% vs −1.94%; the 0.83-point gap is not significant (p = 0.60). We will add these results to the paper.

  • Forest. This dataset, introduced in TempusBench [1], was used as released. The data contain no timestamp information, so we could not verify their exact chronological order. In our experiments we used the series in the order in which it was obtained, without any reordering.

Novelty (W1). The ablations show the combination is not arbitrary: the generator is necessary (Q1), joint updates add significantly (Q2), and validation-guided selection can be dropped (W3).

References (numbered for this response only; they differ from the numbering in our paper):
[1] Goktas et al., TempusBench: An Evaluation Framework for Time-Series Forecasting, Recent Advances in Time Series Foundation Models Have We Reached the 'BERT Moment'?, 2025.

raise to 3

AuthorsOpenReview ↗

Thank you for acknowledging our rebuttal efforts.

Dear Reviewer uJC9,

Thank you again for your careful review and helpful comments.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.