acceptodds
Under review as a conference paper at ICLR 2027

Projection-Based Alignment with a Fractional Trajectory Metric for Time Series Forecasting

Abstract

Time-series forecasting is predominantly trained with point-wise losses such as mean squared error (MSE), although multi-step forecasting requires predicting an entire future trajectory with joint temporal structure. Such objectives emphasize sample-wise accuracy but do not explicitly account for distributional structure across forecast horizons. We propose ProDA, a projection-based and model-agnostic objective that aligns the joint distributions of history-future and history-forecast trajectories through the Cramér-Wold discrepancy. While this provides a distribution-sensitive criterion, its pairwise geometry is governed by Euclidean trajectory distances, which can be dominated by slowly varying components under non-stationarity. To address this, ProDA incorporates fractional differencing directly into the alignment objective, inducing a structured trajectory metric that attenuates low-frequency discrepancies and relatively emphasizes localized temporal variation. Theoretically, we show that the induced Cramér-Wold kernel remains characteristic under the fractional transformation and derive its spectral reweighting, establishing that the geometry of distributional alignment can be modified without sacrificing distributional identifiability. Extensive experiments across multiple forecasting backbones, benchmark datasets, and zero-shot transfer settings demonstrate consistent improvements over standard training objectives and recent objective-centric baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, Distribution alignment, Cramér–Wold discrepancy, Fractional differencing, Trajectory metric

Reviews

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

This paper proposes ProDA, a model-agnostic objective for time-series forecasting that combines Cramér–Wold distribution alignment with a fractional trajectory metric. By applying fractional differencing to joint history–future trajectories, the method aims to better preserve local temporal variations and reduce oversmoothing. The authors provide theoretical analysis of the induced kernel and its spectral properties. Experiments across eight datasets and four forecasting models show improvements in 31 of 32 settings, with additional zero-shot transfer results.

Critical strengths
  1. The paper addresses a meaningful problem. Improving forecasting objectives without modifying model architectures is practically useful.

  2. I find the fractional trajectory metric interesting. Its frequency-domain interpretation provides a reasonable explanation for emphasizing local temporal variations.

  3. The experimental results are fairly strong. ProDA improves most backbone–dataset combinations, and the comparisons with recent forecasting objectives and zero-shot transfer experiments are encouraging.

Critical weaknesses
  1. The theoretical contribution seems relatively modest. The characteristic-kernel result largely follows from the invertibility of the fractional transformation.

  2. The oversmoothing claim needs stronger evidence. Most results focus on MSE and MAE. I would like to see direct measurements of high-frequency reconstruction or sharp-transition accuracy.

  3. Some experimental details remain unclear. The PEMS08 data split appears inconsistent with the stated protocol, and the additional training cost is substantial on ECL.

Questions for authors
  1. What is the main methodological difference between ProDA and DistDF beyond replacing the distribution discrepancy and introducing fractional differencing?

  2. Could the authors clarify the theoretical limit as K approaches infinity under the constraint K < N? How does this analysis relate to the finite-length implementation?

  3. Can the authors provide quantitative evidence that ProDA specifically reduces oversmoothing, rather than simply improving overall forecasting accuracy?

  4. Could the authors clarify the PEMS08 data split and report paired multi-seed significance tests against the strongest objective baselines?

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We thank the reviewer for the constructive feedback and insightful questions. We have carefully addressed the concerns and revised the manuscript accordingly.

Answer for Weakness 1 & Question 1 & Weakness 3

What is the main methodological difference between ProDA and DistDF beyond replacing the distribution discrepancy and introducing fractional differencing?

PRODA and DistDF both incorporate distributional information from joint trajectories, z=(x,y)\mathbf{z}=(\mathbf{x},\mathbf{y}) and z^=(x,y^)\hat{\mathbf{z}}=(\mathbf{x},\hat{\mathbf{y}}), into conventional MSE-based point forecasting models. However, the two methods differ in the following aspects.

1. Distribution Alignment and Geometric Structure.

While DistDF evaluates distributional differences between joint trajectories using a Wasserstein-based discrepancy, PRODA applies Fractional Differencing to joint trajectories and aligns the transformed trajectories using the Cramér–Wold discrepancy. In particular, PRODA integrates Fractional Differencing not merely as a preprocessing step but as a structured trajectory metric that defines distances between trajectories. Accordingly, pairwise distances in the induced kernel are determined by Δd,K⊤Δd,K\Delta_{d,K}^{\top}\Delta_{d,K}, resulting in the frequency-selective geometry described in Proposition 2.

2. Theoretical Contribution and Interpretation.

We agree that the characteristic-kernel result in Proposition 1 largely follows from the invertibility of the Fractional Differencing operator. However, our theoretical contribution does NOT lie in establishing the characteristic property itself, but in analyzing the structure of the trajectory distance metric induced by combining Fractional Differencing with Cramér–Wold distribution alignment. Proposition 2 shows that Δd,K⊤Δd,K\Delta_{d,K}^{\top}\Delta_{d,K} induces a frequency-selective geometry that relatively emphasizes high-frequency differences.

Furthermore, the Cramér–Wold discrepancy can be computed through the closed-form kernel representation derived in Appendix A.5 and can also be interpreted from the perspectives of kernelized data depth and diversity-preserving regularization of predicted trajectories. Thus, the key contribution of PRODA is not merely replacing the discrepancy measure, but designing an objective function that combines a frequency-selective trajectory metric with trajectory-level distribution alignment.

3. Empirical Performance and Computational Efficiency.

In Table 2 of Section 4.2, PRODA achieves a lower average MSE than DistDF across all 12 backbone–dataset settings.

We also acknowledge the additional training cost of PRODA. On ECL, the per-epoch training time increases from 26.29 to 39.22 seconds for iTransformer and from 17.82 to 30.96 seconds for DLinear. To provide a more comprehensive evaluation, we have extended Table 13 in Appendix C.6 by adding GPU memory usage and theoretical complexity analysis to the existing per-epoch training time. The additional time and memory complexities of PRODA are O(CBN(K+1)+CB2N)\mathcal{O}(CBN(K+1)+CB^2N) and O(CBN+CB2+CK)\mathcal{O}(CBN+CB^2+CK), respectively, where BB, CC, NN, and KK denote the batch size, number of variables, joint trajectory length, and maximum differencing lag. We further compared PRODA with the publicly available dense implementation of DistDF [1], whose time and memory complexities are O(CBN2+CN3)\mathcal{O}(CBN^2+CN^3) and O(CBN+CN2)\mathcal{O}(CBN+CN^2), respectively. Since PRODA uses a B×BB\times B pairwise distance matrix instead of an N×NN\times N covariance matrix, it may have a relatively advantageous computational structure when B<NB<N.

We have incorporated this analysis into Appendix C.6 to clarify the computational costs and limitations associated with the forecasting performance improvements of PRODA.

Reference
[1] Wang, H., Pan, L., Lu, Y., Chu, Z., Li, X., He, S., Chen, Z., Li, H., Wen, Q., & Lin, Z. (2026). DistDF: Time-Series Forecasting Needs Joint-Distribution Wasserstein Alignment. International Conference on Learning Representations.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Answer for Question 2

Could the authors clarify the theoretical limit as K approaches infinity under the constraint K < N?

We acknowledge that K→∞K\to\infty and N>KN>K alone were insufficient to establish the convergence of the Boundary Remainder in Appendix A.4. The result requires the joint limit K→∞K\to\infty and N−K→∞N-K\to\infty.

In Equation (25), the Boundary Remainder is bounded by the absolute sum of fractional differencing coefficients and the sum of squared trajectory differences. Their absolute and square summability, respectively, ensure that the bound converges to zero as N−K→∞N-K\to\infty. We have revised Proposition 2 and Appendix A.4 accordingly.

How does this analysis relate to the finite-length implementation?

In practice, we use finite NN and K<NK<N. Finite-length trajectory comparisons are defined by Δd,K\Delta_{d,K} (Section 3.3) and its distance metric Δd,K⊤Δd,K\Delta_{d,K}^{\top}\Delta_{d,K}. Proposition 2 retains the Boundary Remainder for finite lengths, while the limiting frequency-weighting representation describes its asymptotic properties. Thus, the result explains the frequency-selective properties without claiming that boundary effects vanish in finite implementations.

Answer for Weakness 2 & Question 3

Can the authors provide quantitative evidence that ProDA specifically reduces oversmoothing, rather than simply improving overall forecasting accuracy?

We agree that more direct quantitative evidence is needed to support the claim that PRODA alleviates oversmoothing. While Figure 3 and Proposition 2 provide qualitative and theoretical support, respectively, we conducted an additional analysis of high-frequency reconstruction errors following [2].

Following Figure 3, we applied the Discrete Fourier Transform to decompose ground-truth and predicted trajectories into low-frequency (LF) and high-frequency (HF) components at ∣ω∣/π=1/3|\omega|/\pi=1/3. Using Parseval's theorem, we decomposed overall MSE into LF-MSE and HF-MSE and compared them across four datasets and four backbones. These results and additional qualitative analysis are included in Appendix C.8.

Overall MSE (MSE)Overall MSE (+PRODA)Gain(%)LF-MSE (MSE)LF-MSE (+PRODA)Gain(%)HF-MSE (MSE)HF-MSE (+PRODA)Gain(%)
ETTm2iTransformer0.300 ± 0.0030.296 ± 0.002+1.33%0.240 ± 0.0020.235 ± 0.002+2.08%0.060 ± 0.0020.061 ± 0.001-1.67%
PatchTST0.282 ± 0.0030.274 ± 0.002+2.84%0.225 ± 0.0020.224 ± 0.002+0.44%0.057 ± 0.0020.050 ± 0.001+12.28%
TimeMixer0.308 ± 0.0040.271 ± 0.003+12.01%0.234 ± 0.0030.218 ± 0.002+6.84%0.074 ± 0.0020.053 ± 0.002+28.38%
DLinear0.321 ± 0.0030.270 ± 0.002+15.89%0.241 ± 0.0020.218 ± 0.002+9.54%0.080 ± 0.0020.052 ± 0.001+35.00%
WeatheriTransformer0.256 ± 0.0020.245 ± 0.002+4.30%0.191 ± 0.0020.190 ± 0.002+0.52%0.065 ± 0.0010.055 ± 0.001+15.38%
PatchTST0.235 ± 0.0020.227 ± 0.001+3.40%0.177 ± 0.0020.174 ± 0.001+1.69%0.058 ± 0.0010.053 ± 0.001+8.62%
TimeMixer0.242 ± 0.0020.236 ± 0.002+2.48%0.182 ± 0.0020.177 ± 0.002+2.75%0.060 ± 0.0010.059 ± 0.001+1.67%
DLinear0.248 ± 0.0030.241 ± 0.002+2.82%0.188 ± 0.0020.185 ± 0.002+1.60%0.060 ± 0.0020.056 ± 0.001+6.67%
ECLiTransformer0.168 ± 0.0020.167 ± 0.001+0.60%0.123 ± 0.0020.122 ± 0.001+0.81%0.045 ± 0.0010.045 ± 0.0010.00%
PatchTST0.178 ± 0.0030.171 ± 0.002+3.93%0.130 ± 0.0020.128 ± 0.002+1.54%0.048 ± 0.0020.043 ± 0.001+10.42%
TimeMixer0.187 ± 0.0030.178 ± 0.002+4.81%0.131 ± 0.0020.129 ± 0.002+1.53%0.056 ± 0.0020.049 ± 0.001+12.50%
DLinear0.170 ± 0.0020.166 ± 0.001+2.35%0.122 ± 0.0020.122 ± 0.0010.00%0.048 ± 0.0010.044 ± 0.001+8.33%
PEMS08iTransformer0.236 ± 0.0040.213 ± 0.003+9.75%0.144 ± 0.0030.135 ± 0.002+6.25%0.092 ± 0.0020.078 ± 0.002+15.22%
PatchTST0.243 ± 0.0040.237 ± 0.003+2.47%0.148 ± 0.0030.146 ± 0.002+1.35%0.095 ± 0.0020.091 ± 0.002+4.21%
TimeMixer0.214 ± 0.0030.213 ± 0.002+0.47%0.129 ± 0.0020.129 ± 0.0020.00%0.085 ± 0.0020.084 ± 0.001+1.18%
DLinear0.333 ± 0.0050.336 ± 0.004-0.90%0.208 ± 0.0030.210 ± 0.003-0.96%0.125 ± 0.0030.126 ± 0.002-0.80%

Reference
[2] Hu, Y., Yang, J., Zhou, T., Liu, P., Tang, Y., Jin, R., & Sun, L. (2026). Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting. International Conference on Learning Representations.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Answer for Weakness3 & Question 4

The PEMS08 data split appears inconsistent with the stated protocol. Could the authors clarify the PEMS08 data split?

We acknowledge that the description of the data splitting protocol for PEMS08 was inaccurate. In the actual experiments, we used a Train / Validation / Test ratio of 6:2:2, but it was incorrectly stated as 7:1:2 in the original Appendix B.2. Accordingly, we have corrected Appendix B.2 and Table 5 in the revised manuscript to match the actual implementation. This can also be verified in lines 108–115 of datasets/PEMS08_dataset.py in the Supplementary Material, where the data split is implemented as follows:

num_train = int(len(base) * 0.6)
num_test = int(len(base) * 0.2)
num_vali = len(base) - num_train - num_test

Could the authors report paired multi-seed significance tests against the strongest objective baselines?

We added a paired multi-seed significance test comparing PRODA with the objective-based baseline that achieved the lowest MSE for each dataset–backbone setting in the original Table 2. Based on the MSE values obtained using the same 10 random seeds, we report the mean and standard error, the paired t-test p-value, and the Holm-adjusted p-value to examine the statistical significance of the performance differences between PRODA and the strongest competing methods.

StrongestBaseline MSE ↓PRODA MSE ↓ΔMSE ↑Paired pHolm p
ETTh1iTransformerFreDF0.4440 ± 0.00180.4470 ± 0.0020-0.00300.03860.1543
PatchTSTDBLoss0.4350 ± 0.00180.4240 ± 0.0022+0.01100.00100.0119
TimeMixerFreDF0.4560 ± 0.00160.4510 ± 0.0018+0.00500.00420.0292
DLinearTime-o10.4230 ± 0.00180.4150 ± 0.0022+0.00800.00270.0292
WeatheriTransformerPSLoss0.2490 ± 0.00140.2440 ± 0.0017+0.00500.00420.0292
PatchTSTDBLoss0.2350 ± 0.00160.2280 ± 0.0019+0.00700.00270.0292
TimeMixerDBLoss0.2460 ± 0.00200.2370 ± 0.0025+0.00900.00310.0292
DLinearFreDF0.2430 ± 0.00150.2400 ± 0.0018+0.00300.03860.1543
ECLiTransformerTime-o10.1760 ± 0.00200.1670 ± 0.0025+0.00900.00310.0292
PatchTSTDBLoss0.1770 ± 0.00160.1710 ± 0.0019+0.00600.00370.0292
TimeMixerFreDF0.1800 ± 0.00150.1780 ± 0.0018+0.00200.09570.1914
DLinearDBLoss0.1660 ± 0.00130.1660 ± 0.0017+0.00001.00001.0000
Reviewer_LkWFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ProDA, a model-agnostic training objective for multi-step time-series forecasting. ProDA combines MSE with a Cramér–Wold discrepancy between joint history–future and history–forecast trajectories after truncated fractional differencing. The transformation induces a temporally structured metric that relatively emphasizes higher-frequency discrepancies. The theoretical analysis establishes preservation of the characteristic-kernel property under the finite-dimensional invertible transformation and provides a spectral interpretation of the induced metric.

Critical strengths
  1. Broad empirical support for a practical, model-agnostic objective. ProDA improves average MSE in 31 of the 32 backbone–dataset combinations in Table 1, with substantial gains in some settings, such as DLinear on ETTh2. The evaluation includes multiple architectures, alternative training objectives, transfer experiments, and selected five-seed results. This breadth makes the method potentially useful beyond a particular forecasting architecture.

  2. A coherent and interpretable way to incorporate temporal structure into distribution alignment. Applying fractional differencing within the discrepancy changes the geometry of trajectory comparison without requiring architectural modifications. The finite-dimensional invertibility argument provides a sound justification for preserving distributional identifiability, while the spectral interpretation offers useful intuition. The comparison with CW alignment without differencing also provides evidence that the transformed geometry can contribute additional gains.

Critical weaknesses
  1. The evidence does not yet sufficiently isolate the contribution of the proposed fractional CW formulation. The existing CW ablation and alternative-discrepancy comparisons are useful, but it is unclear whether the alternatives in Table 4 use the same fractional transformation and comparable hyperparameter tuning. A particularly informative control would be MSE combined with a paired squared-error penalty on fractionally differenced trajectories, which would help distinguish the benefit of frequency reweighting from that of distribution alignment. In addition, Appendix B.4 uses recommended settings for competing objectives, whereas Table 7 specifies a search space for ProDA; the relative tuning budgets are not sufficiently clear. These issues matter because the paper’s central contribution is the particular objective design, rather than simply the usefulness of additional regularization.

  2. The statistical motivation and spectral analysis support narrower conclusions than some of the paper’s claims suggest. For a deterministic forecaster, the predicted conditional distribution is a point mass and generally cannot match a non-degenerate true conditional distribution. Moreover, a smooth conditional mean can be optimal under MSE even with perfect optimization. The characteristic-kernel result therefore supports identifiability of the discrepancy, but does not establish recoverability of conditional variability or improved prediction of high-frequency components. The spectral result likewise describes the underlying quadratic metric rather than directly guaranteeing the behavior of the full nonlinear CW objective. There is also a fixable gap in Appendix A.4: K tending to infinity with N > K does not itself imply that the tail sum starting at N−K+1 vanishes. I would welcome a more precise interpretation of ProDA as a regularizer and a corrected limiting argument.

Questions for authors
  1. Experimental protocol clarification: Were competing objectives given comparable validation-based tuning budgets? Does “validation loss” mean a common validation MSE or each method’s own training objective? For zero-shot transfer, were hyperparameters and checkpoints selected exclusively using source-domain data, and what data were used to fit normalization statistics?

  2. Contribution clarification and supporting evidence: Do the RBF-MMD and other alternatives in Table 4 use the same fractional transformation as ProDA? Can the authors provide a controlled comparison with MSE plus a paired fractional-difference error penalty, or explain whether existing results already isolate this effect? This would materially help assess the necessity of the proposed combination.

  3. Theoretical clarification: In Appendix A.4, how is the vanishing tail bound justified when N−K need not tend to infinity? For example, N=K+1 satisfies N>K but keeps the tail’s lower index fixed. Please clarify the intended joint limit or provide a revised argument. Separately, how should the distribution-alignment interpretation be understood when the forecasting model is deterministic but the true conditional distribution is non-degenerate?

  4. Reporting clarification: Table 4 reports a Weather MAE of 0.266 for DLinear with ProDA, whereas Tables 1, 2, and 11 report 0.286. Which value and experimental setting are correct? Please also reconcile the PEMS08 split counts in Table 5 with the stated 7:1:2 protocol.

5 comments
AuthorsOpenReview ↗

Rebuttal from Authors (1/3)

We thank the reviewer for the constructive comments and valuable suggestions. We have provided additional clarifications and experimental evidence to address the concerns.

Answer for Weakness 1 & Question 1: Experimental protocol clarification

Were competing objectives given comparable validation-based tuning budgets?

As described in Appendix B.4, the hyperparameter settings of the compared loss functions followed the official implementations and recommended settings provided by their original authors, rather than applying identical validation-based tuning budgets. We have also revised Table 7 to report the hyperparameter search spaces of the competing objectives. This approach maintains consistency with the previously validated configurations of the compared methods.

Does “validation loss” mean a common validation MSE or each method’s own training objective?

The "Validation Loss" specified in Appendix B.5 refers to the MSE on the validation dataset, which is commonly applied across all methods, rather than the method-specific training objectives. Therefore, model checkpoints were selected based on the same validation MSE criterion during training. Specifically, as shown in lines 108–115 of modules/train.py in the code repository provided in the Supplementary Material, the learning rate scheduler was adjusted and the best model was selected based on the MSE computed on the validation dataset.

"""validation"""
valid_mse, valid_mae, valid_cost = validation(model, valid_dataloader, config)
test_mse, test_mae, test_cost = validation(model, test_dataloader, config)
scheduler.step(valid_mse)
#%%
if valid_mse < best_valid_mse:
    best_valid_mse = valid_mse.item() if torch.is_tensor(valid_mse) else float(valid_mse)
    best_state_dict = copy.deepcopy(model.state_dict())

To avoid confusion, we have changed the term "Validation Loss" to "Validation MSE" in Appendix B.5 of the revised manuscript to clarify that the same validation MSE criterion was used for model selection across all methods.

For zero-shot transfer, were hyperparameters and checkpoints selected exclusively using source-domain data, and what data were used to fit normalization statistics?

For zero-shot transfer, the model was trained on the source-domain training data, and checkpoints were selected based on the source-domain validation MSE. Normalization statistics were estimated independently from the training splits of the source and target domains, and the target-domain test data were not used to estimate these statistics. This normalization protocol can also be found in the publicly available zero-shot implementations of One Fits All [1] and AutoTimes [2]. This implementation can be verified in the data preprocessing code under datasets/ in the Supplementary Material.

"""Standardization"""
self.scaler = StandardScaler()

if self.scale:
    st, ed = border1_[0], border2_[0]
    train_data = df[st:ed]
    
    self.scaler.fit(train_data.values)
    data = self.scaler.transform(df.values)
else:
    data = df.values

Reference
[1] Zhou, T., Niu, P., Wang, X., Sun, L., & Jin, R. (2023). One Fits All: Power General Time Series Analysis by Pretrained LM. Advances in Neural Information Processing Systems.
[2] Liu, Y., Qin, G., Huang, X., Wang, J., & Long, M. (2024). AutoTimes: Autoregressive Time Series Forecasters via Large Language Models. Advances in Neural Information Processing Systems.

AuthorsOpenReview ↗

Rebuttal from Authors (2/3)

Answer for Weakness 1 & Question 2: Contribution clarification and supporting evidence

Do the RBF-MMD and other alternatives in Table 4 use the same fractional transformation as ProDA?

The experiments in Table 4 were indeed conducted using the same fractional differencing transformation. We have explicitly clarified this in the paragraph titled "Comparison of alternative discrepancies" on page 9 of the revised manuscript.

Can the authors provide a controlled comparison with MSE plus a paired fractional-difference error penalty, or explain whether existing results already isolate this effect?

To clarify the contribution of each component of the proposed objective function, we have added a w/o CW setting, which applies fractional differencing without distribution alignment, to the existing ablation study. By comparing four settings (MSE, w/o FD, w/o CW, and PRODA) we evaluated the individual and combined effects of fractional differencing and CW alignment. The additional results are presented in the table below and have also been incorporated into Figure 5 of the Ablation Study and Figure 8 of Appendix C.2 in the revised manuscript.

DatasetBackboneMSE ↓ (Baseline)MSE ↓ (w/o FD)MSE ↓ (w/o CW)MSE ↓ (PRODA)MAE ↓ (Baseline)MAE ↓ (w/o FD)MAE ↓ (w/o CW)MAE ↓ (PRODA)
ETTh1iTransformer0.4520.4530.4500.4470.4540.4550.4550.456
PatchTST0.4470.4360.4370.4240.4460.4440.4440.441
TimeMixer0.4900.4580.4720.4510.4680.4650.4660.458
DLinear0.4510.4460.4340.4150.4470.4400.4390.427
ETTh2iTransformer0.3960.3830.3920.3910.4240.4070.4200.419
PatchTST0.3790.3750.3800.3780.4060.4070.4100.414
TimeMixer0.4060.3880.3920.3630.4220.4120.4160.409
DLinear0.5260.4600.4860.4300.4970.4620.4710.449
ETTm1iTransformer0.3930.4040.3910.3870.4110.4120.4100.407
PatchTST0.3890.3830.3790.3630.4020.3990.3980.393
TimeMixer0.3970.4080.3900.3770.4090.4100.4060.395
DLinear0.4040.4020.3860.3540.4080.4040.3990.376
ETTm2iTransformer0.2990.2930.2990.2970.3540.3350.3520.351
PatchTST0.2830.2860.2800.2730.3280.3310.3290.326
TimeMixer0.3070.3030.2940.2700.3400.3410.3360.328
DLinear0.3220.3020.2960.2690.3750.3570.3510.332
WeatheriTransformer0.2570.2550.2520.2440.2870.2850.2850.282
PatchTST0.2340.2330.2310.2280.2760.2750.2720.267
TimeMixer0.2410.2380.2400.2370.2790.2770.2780.274
DLinear0.2470.2440.2450.2400.3000.2920.2940.286
ECLiTransformer0.1680.1670.1680.1670.2620.2620.2620.262
PatchTST0.1780.1740.1760.1710.2730.2680.2690.267
TimeMixer0.1870.1840.1830.1780.2890.2780.2820.275
DLinear0.1700.1690.1680.1660.2660.2650.2630.261
ExchangeiTransformer0.3680.3670.3640.3590.4180.4100.4140.409
PatchTST0.3950.3480.3730.3430.4240.4190.4150.398
TimeMixer0.4130.3520.3920.3480.4370.4120.4220.401
DLinear0.3080.3020.3020.2920.3850.3820.3850.381
PEMS08iTransformer0.2360.2340.2260.2130.3070.2920.3000.290
PatchTST0.2430.2380.2410.2370.3200.3120.3170.312
TimeMixer0.2140.2130.2150.2130.3050.3050.3050.305
DLinear0.3330.3330.3320.3360.3710.3700.3710.369
AuthorsOpenReview ↗

Rebuttal from Authors (3/3)

Answer for Weakness 2 & Question 3: Theoretical Clarification

In Appendix A.4, how is the vanishing tail bound justified when N−KN−K need not tend to infinity? For example, N=K+1N=K+1 satisfies N>KN>K but keeps the tail’s lower index fixed. Please clarify the intended joint limit or provide a revised argument.

As you pointed out, the conditions K→∞K\to\infty and N>KN>K alone were INSUFFICIENT to establish that the tail term converges to zero in the original proof in Appendix A.4. This result holds under the joint limit where both K→∞K\to\infty and N−K→∞N-K\to\infty. Specifically, in Equation (25) of Appendix A.4, the tail term is bounded as follows ∣RN,K∣≤(∑k=0∞∣(−1)k(dk)∣)2∑l=N−K+1∞∣zl−z^l∣2.|R_{N,K}|\leq\left(\sum_{k=0}^{\infty}\left|(-1)^k\binom{d}{k}\right|\right)^2\sum_{l=N-K+1}^{\infty}|z_l-\hat z_l|^2.

By the absolute summability of the fractional differencing coefficients and the square summability of the trajectory differences, the right-hand side converges to zero as N−K→∞N-K\to\infty. Therefore, RN,K→0R_{N,K}\to0 is guaranteed under the corrected limit conditions. Accordingly, we have explicitly added these limit conditions to Proposition 2 and its proof in Appendix A.4 of the revised manuscript to address the incompleteness of the original argument.

Separately, how should the distribution-alignment interpretation be understood when the forecasting model is deterministic but the true conditional distribution is non-degenerate?

PRODA is not a distribution learning model that explicitly estimates the true conditional distribution p(y∣x)p(\mathbf{y}\mid\mathbf{x}), but rather a regularization method for deterministic point forecasting based on distribution alignment.

Specifically, the prediction function y^=f(x;θ)\hat{\mathbf{y}}=f(\mathbf{x};\theta) of PRODA outputs a single future trajectory for a given historical observation x\mathbf{x}, so its predictive conditional distribution corresponds to a point mass. Therefore, we do not claim that PRODA recovers or matches the non-degenerate true conditional distribution p(y∣x)p(\mathbf{y}\mid\mathbf{x}). Instead, PRODA compares the distributions of the true joint trajectory z=(x,y)\mathbf{z}=(\mathbf{x},\mathbf{y}) and the predicted joint trajectory z^=(x,f(x;θ))\hat{\mathbf{z}}=(\mathbf{x},f(\mathbf{x};\theta)) using the Cramér–Wold discrepancy and adds it as a regularization term to MSE-based training. In other words, distribution alignment should be interpreted not as a means of estimating the conditional distribution itself, but as a training constraint that encourages the point forecasting function to account for trajectory-level distributional characteristics.

Answer for Question 4: Reporting Clarification

Table 4 reports a Weather MAE of 0.266 for DLinear with ProDA, whereas Tables 1, 2, and 11 report 0.286. Which value and experimental setting are correct?

First, we apologize for the incorrectly reported experimental results for the Weather dataset in Table 4, which may have caused confusion. Upon verification, we found errors not only in the MAE of PRODA but also in the results for MMD@Linear, MMD@RBF, and KL divergence. In particular, the correct MAE of PRODA for the DLinear–Weather setting is 0.286, as reported in Tables 1, 2, and 11. Accordingly, we have re-examined the results for the Weather dataset in Table 4, corrected them using the verified values, and incorporated these corrections into the revised manuscript.

Please also reconcile the PEMS08 split counts in Table 5 with the stated 7:1:2 protocol.

We also acknowledge that the description of the data splitting protocol for PEMS08 was inaccurate. In the actual experiments, we used a Train / Validation / Test ratio of 6:2:2 for PEMS08, but Appendix B.2 incorrectly stated that a 7:1:2 ratio was applied to all datasets except ETT. Therefore, we have explicitly specified the splitting ratio for PEMS08 in Appendix B.2 of the revised manuscript and correct the numbers of samples in Table 5 to match the actual experimental settings. This splitting ratio can also be verified in lines 108–115 of datasets/PEMS08_dataset.py in the Supplementary Material, as shown below.

num_train = int(len(base) * 0.6)
num_test = int(len(base) * 0.2)
num_vali = len(base) - num_train - num_test
Reviewer_RJGxOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of loss function engineering in time-series analysis based on deep neural networks. Specifically, it finds that the MSE-based loss fail to capture the cross-horizon dependency in the future labels, and has poor extrapolation capability. To address these problems, it proposes PRODA, which is featured by an additional Cramér–Wold distribution-alignment term to the MSE loss. It compares the trajectory-level distributional properties while using fractional differencing to capture the dynamic variability of time series. Experiments demonstrate that it can improve MSE in most settings, while ablation and zero-shot transfer experiments offer additional evidence for its effectiveness.

Critical strengths

This paper investigates the loss function in time series forecasting, which is a relevant and interesting topic. The experiment settings, such as datasets, baselines, where the method improves forecasting across several backbones and datasets. The code implementation is provided and the implementation details are well introduced, which improves the reproducibility of this work. The overall paper is written well and easy to follow.

Critical weaknesses

Authors identified two limitations of MSE: oversmoothing issue and poor extrapolation capability in the first two paragraphs of introduction. However, the addressed issues claimed in the following paragraphs of introduction are handling non-stationary series and aligning trajectory-level distributional properties, which seems misaligned with the previous statements.

The evidence that the two identified limitations reduce forecasting performance is lacking. More importantly authors should prove that PRODA can effectively address the identified problems, with theoretical analysis and specific empirical studies. While there are some theoretical analysis in section 3.4, they focus on the induced kernel and spectral interpretation, not highly related to the effectiveness of the proposed method to handle the identified two issues. Finally, authors should prove that handling the two issues can improve the performance of forecasting task.

The empirical effectiveness of PRODA is not very persuasive. For example, in Table 1 (PEMS08), the MSE of PRODA is 0.213, while the MSE of MSE is 0.214. Given the implementation complexity and huge search space of PRODA, such improvement is not persuasive. Moreover, it is not clear why PRODA outperforms MSE in terms of MSE metric.

The search space of the hyper-parameters is very dense by Table 7. It is not clear whether the performance gain of the proposed method comes from its genuine effectiveness or exhaustive tuning. Is the hyperparameter tuning conducted on the validation dataset? Authors should report the process to tune the proposed method, and provide concrete hyperparameters used for each dataset. Moreover, is it feasible to provide a set of reproducible logs containing the searched hyperparameters and validation results?

The hyperparameter tuning space for other loss functions should be introduced in detail. Since exhaustive tuning is performed for PRODA, it is necessary to perform hyperparameter tuning for compared baselines under the same setting and tuning quota to ensure fair comparison.

The complexity analysis covers merely the training time in each epoch. A comprehensive complexity analysis is needed, which should cover memory usage, end-to-end training time, supported with theoretical analysis and comparison with modern baselines.

Questions for authors

Why do you employ a weight of MSE in the final loss, given its identified limitations?

7 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We thank the reviewer for the detailed feedback, which we address through additional analyses and revisions.

Answer for Weakness 1 & 2

Methodological Clarification and Supporting Evidence.

We agree that the original Introduction insufficiently connected the identified limitations to PRODA. We clarify this below.

First, PRODA regularizes point forecasting through distribution alignment rather than estimating conditional distributions.

MSE-based forecasting targets the conditional mean E[y∣x]\mathbb{E}[\mathbf{y}\mid\mathbf{x}] without explicitly constraining temporal structure or local variability. PRODA instead regularizes deterministic forecasts by aligning true and predicted joint trajectories, (x,y)(\mathbf{x},\mathbf{y}) and (x,f(x;θ))(\mathbf{x},f(\mathbf{x};\theta)), using the Cramér–Wold discrepancy, without estimating p(y∣x)p(\mathbf{y}\mid\mathbf{x}).

Second, Fractional Differencing and Cramér–Wold alignment jointly address oversmoothing and difficulties in predicting abrupt future variations.

MSE does not explicitly balance low- and high-frequency discrepancies, potentially causing oversmoothing or inadequate reconstruction of abrupt variations. PRODA uses Fractional Differencing to adjust their relative weighting and Cramér–Wold alignment to regularize trajectory distributions under this frequency-selective geometry. Proposition 1 establishes the characteristic kernel property, while Proposition 2 derives the asymptotic weighting (2∣sin⁡(ω/2)∣)2d(2|\sin(\omega/2)|)^{2d}, explaining the relative emphasis on high-frequency variations.

Third, we added frequency-decomposed forecasting results in Appendix C.8.

Using the Figure 3 boundary ∣ω∣/π=1/3|\omega|/\pi=1/3, we decomposed MSE into LF-MSE and HF-MSE via Parseval's theorem. PRODA reduced HF-MSE in 13 of 16 settings across four datasets and four backbones, including a 35.0% reduction (0.080 to 0.052) for ETTm2–DLinear. These results provide quantitative evidence that PRODA improves forecasting performance across frequency components, complementing the qualitative observations in Figure 3 and the ablation results in Figure 5.

Overall MSE (MSE)Overall MSE (+PRODA)Gain(%)LF-MSE (MSE)LF-MSE (+PRODA)Gain(%)HF-MSE (MSE)HF-MSE (+PRODA)Gain(%)
ETTm2iTransformer0.300 ± 0.0030.296 ± 0.002+1.33%0.240 ± 0.0020.235 ± 0.002+2.08%0.060 ± 0.0020.061 ± 0.001-1.67%
PatchTST0.282 ± 0.0030.274 ± 0.002+2.84%0.225 ± 0.0020.224 ± 0.002+0.44%0.057 ± 0.0020.050 ± 0.001+12.28%
TimeMixer0.308 ± 0.0040.271 ± 0.003+12.01%0.234 ± 0.0030.218 ± 0.002+6.84%0.074 ± 0.0020.053 ± 0.002+28.38%
DLinear0.321 ± 0.0030.270 ± 0.002+15.89%0.241 ± 0.0020.218 ± 0.002+9.54%0.080 ± 0.0020.052 ± 0.001+35.00%
WeatheriTransformer0.256 ± 0.0020.245 ± 0.002+4.30%0.191 ± 0.0020.190 ± 0.002+0.52%0.065 ± 0.0010.055 ± 0.001+15.38%
PatchTST0.235 ± 0.0020.227 ± 0.001+3.40%0.177 ± 0.0020.174 ± 0.001+1.69%0.058 ± 0.0010.053 ± 0.001+8.62%
TimeMixer0.242 ± 0.0020.236 ± 0.002+2.48%0.182 ± 0.0020.177 ± 0.002+2.75%0.060 ± 0.0010.059 ± 0.001+1.67%
DLinear0.248 ± 0.0030.241 ± 0.002+2.82%0.188 ± 0.0020.185 ± 0.002+1.60%0.060 ± 0.0020.056 ± 0.001+6.67%
ECLiTransformer0.168 ± 0.0020.167 ± 0.001+0.60%0.123 ± 0.0020.122 ± 0.001+0.81%0.045 ± 0.0010.045 ± 0.0010.00%
PatchTST0.178 ± 0.0030.171 ± 0.002+3.93%0.130 ± 0.0020.128 ± 0.002+1.54%0.048 ± 0.0020.043 ± 0.001+10.42%
TimeMixer0.187 ± 0.0030.178 ± 0.002+4.81%0.131 ± 0.0020.129 ± 0.002+1.53%0.056 ± 0.0020.049 ± 0.001+12.50%
DLinear0.170 ± 0.0020.166 ± 0.001+2.35%0.122 ± 0.0020.122 ± 0.0010.00%0.048 ± 0.0010.044 ± 0.001+8.33%
PEMS08iTransformer0.236 ± 0.0040.213 ± 0.003+9.75%0.144 ± 0.0030.135 ± 0.002+6.25%0.092 ± 0.0020.078 ± 0.002+15.22%
PatchTST0.243 ± 0.0040.237 ± 0.003+2.47%0.148 ± 0.0030.146 ± 0.002+1.35%0.095 ± 0.0020.091 ± 0.002+4.21%
TimeMixer0.214 ± 0.0030.213 ± 0.002+0.47%0.129 ± 0.0020.129 ± 0.0020.00%0.085 ± 0.0020.084 ± 0.001+1.18%
DLinear0.333 ± 0.0050.336 ± 0.004-0.90%0.208 ± 0.0030.210 ± 0.003-0.96%0.125 ± 0.0030.126 ± 0.002-0.80%

Reference
[1] Hu, Y., Yang, J., Zhou, T., Liu, P., Tang, Y., Jin, R., & Sun, L. (2026). Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting. International Conference on Learning Representations.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Answer for Weakness 3

The empirical effectiveness of PRODA is not very persuasive.

We agree that the MSE improvement from 0.214 to 0.213 for PEMS08–TimeMixer in Table 1 is marginal. However, this case does not represent PRODA's overall performance. In Table 1, PRODA improved the average MSE in 31 of 32 backbone–dataset combinations. Notably, for ETTh2–DLinear, PRODA reduced the MSE from 0.526 to 0.430.

Furthermore, in Table 14 of Appendix C.7, we extended the repeated experiments to include ETTm2 and Weather, reporting the mean and standard error over five random seeds across four datasets. We examined the consistency and variability of improvements and observed MSE improvements across all backbones on ETTm2, Weather, and ECL. For PEMS08, three backbones improved, while DLinear showed a slight deterioration.

Original MSE ↓PRODA MSE ↓Original MAE ↓PRODA MAE ↓
ETTm2iTransformer0.300 ± 0.0030.296 ± 0.0020.353 ± 0.0020.351 ± 0.002
PatchTST0.282 ± 0.0030.274 ± 0.0020.329 ± 0.0020.325 ± 0.002
TimeMixer0.308 ± 0.0040.271 ± 0.0030.341 ± 0.0030.329 ± 0.002
DLinear0.321 ± 0.0030.270 ± 0.0020.374 ± 0.0030.333 ± 0.002
WeatheriTransformer0.256 ± 0.0020.245 ± 0.0020.288 ± 0.0020.283 ± 0.002
PatchTST0.235 ± 0.0020.227 ± 0.0010.275 ± 0.0020.268 ± 0.001
TimeMixer0.242 ± 0.0020.236 ± 0.0020.279 ± 0.0030.275 ± 0.002
DLinear0.248 ± 0.0030.241 ± 0.0020.301 ± 0.0030.287 ± 0.002
ECLiTransformer0.168 ± 0.0020.167 ± 0.0010.262 ± 0.0020.262 ± 0.001
PatchTST0.178 ± 0.0030.171 ± 0.0020.273 ± 0.0030.267 ± 0.002
TimeMixer0.187 ± 0.0030.178 ± 0.0020.289 ± 0.0040.275 ± 0.003
DLinear0.170 ± 0.0020.166 ± 0.0010.266 ± 0.0020.261 ± 0.002
PEMS08iTransformer0.236 ± 0.0040.213 ± 0.0030.307 ± 0.0030.290 ± 0.003
PatchTST0.243 ± 0.0040.237 ± 0.0030.320 ± 0.0040.312 ± 0.003
TimeMixer0.214 ± 0.0030.213 ± 0.0020.305 ± 0.0030.305 ± 0.002
DLinear0.333 ± 0.0050.336 ± 0.0040.371 ± 0.0040.369 ± 0.003

Answer for Weakness 4 & Weakness 5

Authors should report the process to tune the proposed method, and provide concrete hyperparameters used for each dataset.

The main tuning hyperparameters of PRODA are the regularization weight γ\gamma and fractional differencing order dd. Specifically, γ∈[0,1]\gamma\in[0,1] balances the MSE and CW regularization terms, while d∈(0,1)d\in(0,1) relatively suppresses low-frequency components while preserving temporal dependencies. We defined the search ranges in Table 7 within these domains and selected hyperparameters based on validation MSE. We also updated the supplementary execution scripts under scripts/ to specify the selected hyperparameters for each dataset.

Is the hyperparameter tuning conducted on the validation dataset?

As shown in lines 108–115 of modules/train.py in the code repository provided in the Supplementary Material, the learning rate scheduler was adjusted and the best model was selected based on the MSE computed on the validation dataset.

"""validation"""
valid_mse, valid_mae, valid_cost = validation(model, valid_dataloader, config)
test_mse, test_mae, test_cost = validation(model, test_dataloader, config)
scheduler.step(valid_mse)
#%%
if valid_mse < best_valid_mse:
    best_valid_mse = valid_mse.item() if torch.is_tensor(valid_mse) else float(valid_mse)
    best_state_dict = copy.deepcopy(model.state_dict())

Is it feasible to provide a set of reproducible logs containing the searched hyperparameters and validation results?

We have specified the hyperparameter search ranges for PRODA in Table 7 of Appendix B.5 and provided the wandb_tuning_log.csv file in the Supplementary Material, which contains the searched hyperparameter configurations and corresponding validation results. Based on these tuning results, we recommend starting with γ=0.02\gamma=0.02 and d=0.2d=0.2, as these values were most frequently selected across experimental configurations (68.8% and 78.1%, respectively). These materials enable reproduction of the hyperparameter selection process and the final experimental settings.

The hyperparameter tuning space for other loss functions should be introduced in detail.

As described in Appendix B.4, the hyperparameter settings of the compared loss functions followed the official implementations and recommendations of their original authors. We have also revised Table 7 to include the hyperparameter search spaces of competing objective functions. This was intended to ensure a fair comparison with PRODA while maintaining the previously validated settings of the compared methods.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Answer for Weakness 6

Complexity analysis

For a more comprehensive evaluation of computational efficiency, we have extended the analysis in Table 13 of Appendix C.6 by adding GPU memory usage and theoretical complexity analysis to the existing per-epoch training time. Since the end-to-end training time depends on the number of training epochs, we used per-epoch training time as the primary metric to compare computational costs under the same training unit.

Theoretical analysis and comparison with modern baselines.

The additional time and memory complexities of PRODA are O(CBN(K+1)+CB2N)\mathcal{O}(CBN(K+1)+CB^2N) and O(CBN+CB2+CK)\mathcal{O}(CBN+CB^2+CK), respectively, where BB, CC, NN, and KK denote the batch size, number of variables, joint trajectory length, and maximum differencing lag, respectively. We also added a comparison with DistDF [1], a recent distribution alignment-based method. The publicly available dense implementation of DistDF has a time complexity of O(CBN2+CN3)\mathcal{O}(CBN^2+CN^3) and a memory complexity of O(CBN+CN2)\mathcal{O}(CBN+CN^2). Since PRODA uses a B×BB\times B pairwise distance matrix instead of an N×NN\times N covariance matrix, it may have a relatively advantageous computational structure when B<NB<N.

We have incorporated this analysis into Appendix C.6 to compare the computational costs of the two methods both theoretically and empirically.

Answer for Question

Why do you employ a weight of MSE in the final loss, given its identified limitations?

As in other time-series forecasting studies [2,3,4], the limitations of MSE identified in our study do not imply that MSE itself is unsuitable for time-series forecasting, but rather that MSE alone may not sufficiently capture dependencies across future prediction steps and trajectory-level distributional characteristics.

Therefore, as described in Equation (2) of Section 3.3, the objective of PRODA is not to completely replace MSE, but to maintain the pointwise prediction accuracy provided by MSE while introducing a Cramér–Wold regularization term to preserve the distributional characteristics of differenced trajectories and complement overly smoothed mean prediction trajectories.

Furthermore, as shown in Figure 6 of Section 4.2, using only the Cramér–Wold term (γ=1\gamma=1) degrades forecasting performance, whereas combining it with MSE yields better performance. This supports the effectiveness of the Cramér–Wold term as a regularizer that complements MSE rather than as a standalone loss function.

Reference
[2] Wang, H., Pan, L., Lu, Y., Chu, Z., Li, X., He, S., Chen, Z., Li, H., Wen, Q., & Lin, Z. (2026). DistDF: Time-Series Forecasting Needs Joint-Distribution Wasserstein Alignment. International Conference on Learning Representations.
[3] Kudrat, D., Xie, Z., Sun, Y., Jia, T., & Hu, Q. (2025). Patch-wise Structural Loss for Time Series Forecasting. International Conference on Machine Learning.
[4] Wang, H., Pan, L., Shen, Y., Chen, Z., Yang, D., Yang, Y., Zhang, S., Liu, X., Li, H., & Tao, D. (2025). FreDF: Learning to Forecast in the Frequency Domain. International Conference on Learning Representations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to Reviewers and Summary of Revisions

We sincerely thank all reviewers (RJGx , LkWF , and YeSn) and the Area Chair Ppk9 for their careful evaluation, constructive criticism, and valuable suggestions. We have taken their concerns seriously and revised the manuscript and supplementary material accordingly.

The central contribution of our paper is PRODA, a model-agnostic regularization objective for time-series point forecasting that directly incorporates Fractional Differencing into the geometry of Cramér–Wold distribution alignment, enabling frequency-selective trajectory alignment. Rather than replacing MSE or explicitly estimating conditional probability densities, PRODA retains pointwise supervision while regularizing the empirical distributions of joint History–Future and History–Forecast trajectories. In particular, the structured trajectory metric induced by Fractional Differencing rebalances the relative contributions of low- and high-frequency discrepancies, while remaining applicable to diverse forecasting backbones.

We summarize the main clarifications and revisions below.

1. Clarified the methodological motivation, scope, and theoretical contributions.

We explained that PRODA leverages trajectory-level distributional information to regularize deterministic point forecasts, rather than estimate conditional distributions. Fractional differencing is introduced not merely to address nonstationarity, but to mitigate the dominance of low-frequency discrepancies and rebalance frequency contributions in trajectory alignment. Proposition 1 establishes that the induced kernel remains characteristic, while Proposition 2 characterizes the frequency-dependent weighting of the resulting metric.

2. Strengthened the ablation studies and comparisons with competing objectives.

We extended the ablation study to compare MSE, CW alignment without fractional differencing (w/o FD), a fractional-difference penalty without CW alignment (w/o CW), and full PRODA (Figure 5 and Appendix C.2). We also conducted paired t-tests against the strongest competing objective for each dataset–backbone setting using the same 10 random seeds. PRODA achieved lower mean MSE in 10 of 12 settings, with statistically significant improvements in 8 settings after Holm correction (pHolm<0.05p_{\mathrm{Holm}}<0.05). These tests provide stronger statistical support for PRODA's effectiveness beyond individual experimental runs.

3. Added frequency-decomposed evidence on oversmoothing and extrapolation challenges.

In the new Appendix C.8, we decomposed overall MSE into low- and high-frequency errors (LF-MSE and HF-MSE) using the same cutoff ∣ω∣/π=1/3|\omega|/\pi=1/3 as in Figure 3. PRODA reduced both LF-MSE and HF-MSE in 13 of 16 dataset–backbone settings each. These findings show that frequency-balancing regularization improves the reconstruction of both low-frequency trends and high-frequency variations, helping alleviate oversmoothing. They also support PRODA's ability to reconstruct abrupt changes and high-frequency variations in future trajectories, reinforcing the connection to the extrapolation challenges identified in the Introduction.

4. Strengthened evidence of performance consistency and generalization.

PRODA improves average MSE in 31 of 32 settings across eight datasets and four backbones (Table 1). Furthermore, in Appendix C.7, we report the mean and standard error over five random seeds across four datasets, confirming MSE improvements in 15 of 16 settings. In zero-shot forecasting, PRODA also achieves improvements in 42 of 48 MSE comparisons and 36 of 48 MAE comparisons (Table 3), supporting its generalization capability under source–target distribution shifts. We also acknowledge that some improvements are modest and discuss the overall trends alongside the exceptions.

5. Improved reproducibility and expanded computational complexity analysis.

We clarified that validation MSE is the common model-selection criterion, detailed PRODA's hyperparameter search and officially recommended settings for competing objectives, and updated execution configurations. Appendix C.6 now reports theoretical time and memory complexity, per-epoch training time, peak GPU memory, and comparisons with DistDF, providing a fuller account of the computational costs of trajectory-level regularization.

We hope that these substantive revisions and additional analyses help clarify PRODA's methodological contributions, empirical evidence, and limitations, and address the concerns raised by the reviewers. We would greatly appreciate any further feedback or questions and would be happy to provide additional clarification on any points that may require further discussion.

Once again, we sincerely thank the reviewers and the Area Chair for their valuable feedback and thoughtful consideration.


Best regards,

Paper #56534 PRODA authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.