- 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
-
The paper addresses a meaningful problem. Improving forecasting objectives without modifying model architectures is practically useful.
-
I find the fractional trajectory metric interesting. Its frequency-domain interpretation provides a reasonable explanation for emphasizing local temporal variations.
-
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
-
The theoretical contribution seems relatively modest. The characteristic-kernel result largely follows from the invertibility of the fractional transformation.
-
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.
-
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
-
What is the main methodological difference between ProDA and DistDF beyond replacing the distribution discrepancy and introducing fractional differencing?
-
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?
-
Can the authors provide quantitative evidence that ProDA specifically reduces oversmoothing, rather than simply improving overall forecasting accuracy?
-
Could the authors clarify the PEMS08 data split and report paired multi-seed significance tests against the strongest objective baselines?
5 comments
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 and alone were insufficient to establish the convergence of the Boundary Remainder in Appendix A.4. The result requires the joint limit and .
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 . 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 and . Finite-length trajectory comparisons are defined by (Section 3.3) and its distance metric . 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 . 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(%) | ||
|---|---|---|---|---|---|---|---|---|---|---|
| ETTm2 | iTransformer | 0.300 ± 0.003 | 0.296 ± 0.002 | +1.33% | 0.240 ± 0.002 | 0.235 ± 0.002 | +2.08% | 0.060 ± 0.002 | 0.061 ± 0.001 | -1.67% |
| PatchTST | 0.282 ± 0.003 | 0.274 ± 0.002 | +2.84% | 0.225 ± 0.002 | 0.224 ± 0.002 | +0.44% | 0.057 ± 0.002 | 0.050 ± 0.001 | +12.28% | |
| TimeMixer | 0.308 ± 0.004 | 0.271 ± 0.003 | +12.01% | 0.234 ± 0.003 | 0.218 ± 0.002 | +6.84% | 0.074 ± 0.002 | 0.053 ± 0.002 | +28.38% | |
| DLinear | 0.321 ± 0.003 | 0.270 ± 0.002 | +15.89% | 0.241 ± 0.002 | 0.218 ± 0.002 | +9.54% | 0.080 ± 0.002 | 0.052 ± 0.001 | +35.00% | |
| Weather | iTransformer | 0.256 ± 0.002 | 0.245 ± 0.002 | +4.30% | 0.191 ± 0.002 | 0.190 ± 0.002 | +0.52% | 0.065 ± 0.001 | 0.055 ± 0.001 | +15.38% |
| PatchTST | 0.235 ± 0.002 | 0.227 ± 0.001 | +3.40% | 0.177 ± 0.002 | 0.174 ± 0.001 | +1.69% | 0.058 ± 0.001 | 0.053 ± 0.001 | +8.62% | |
| TimeMixer | 0.242 ± 0.002 | 0.236 ± 0.002 | +2.48% | 0.182 ± 0.002 | 0.177 ± 0.002 | +2.75% | 0.060 ± 0.001 | 0.059 ± 0.001 | +1.67% | |
| DLinear | 0.248 ± 0.003 | 0.241 ± 0.002 | +2.82% | 0.188 ± 0.002 | 0.185 ± 0.002 | +1.60% | 0.060 ± 0.002 | 0.056 ± 0.001 | +6.67% | |
| ECL | iTransformer | 0.168 ± 0.002 | 0.167 ± 0.001 | +0.60% | 0.123 ± 0.002 | 0.122 ± 0.001 | +0.81% | 0.045 ± 0.001 | 0.045 ± 0.001 | 0.00% |
| PatchTST | 0.178 ± 0.003 | 0.171 ± 0.002 | +3.93% | 0.130 ± 0.002 | 0.128 ± 0.002 | +1.54% | 0.048 ± 0.002 | 0.043 ± 0.001 | +10.42% | |
| TimeMixer | 0.187 ± 0.003 | 0.178 ± 0.002 | +4.81% | 0.131 ± 0.002 | 0.129 ± 0.002 | +1.53% | 0.056 ± 0.002 | 0.049 ± 0.001 | +12.50% | |
| DLinear | 0.170 ± 0.002 | 0.166 ± 0.001 | +2.35% | 0.122 ± 0.002 | 0.122 ± 0.001 | 0.00% | 0.048 ± 0.001 | 0.044 ± 0.001 | +8.33% | |
| PEMS08 | iTransformer | 0.236 ± 0.004 | 0.213 ± 0.003 | +9.75% | 0.144 ± 0.003 | 0.135 ± 0.002 | +6.25% | 0.092 ± 0.002 | 0.078 ± 0.002 | +15.22% |
| PatchTST | 0.243 ± 0.004 | 0.237 ± 0.003 | +2.47% | 0.148 ± 0.003 | 0.146 ± 0.002 | +1.35% | 0.095 ± 0.002 | 0.091 ± 0.002 | +4.21% | |
| TimeMixer | 0.214 ± 0.003 | 0.213 ± 0.002 | +0.47% | 0.129 ± 0.002 | 0.129 ± 0.002 | 0.00% | 0.085 ± 0.002 | 0.084 ± 0.001 | +1.18% | |
| DLinear | 0.333 ± 0.005 | 0.336 ± 0.004 | -0.90% | 0.208 ± 0.003 | 0.210 ± 0.003 | -0.96% | 0.125 ± 0.003 | 0.126 ± 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.
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.
| Strongest | Baseline MSE ↓ | PRODA MSE ↓ | ΔMSE ↑ | Paired p | Holm p | ||
|---|---|---|---|---|---|---|---|
| ETTh1 | iTransformer | FreDF | 0.4440 ± 0.0018 | 0.4470 ± 0.0020 | -0.0030 | 0.0386 | 0.1543 |
| PatchTST | DBLoss | 0.4350 ± 0.0018 | 0.4240 ± 0.0022 | +0.0110 | 0.0010 | 0.0119 | |
| TimeMixer | FreDF | 0.4560 ± 0.0016 | 0.4510 ± 0.0018 | +0.0050 | 0.0042 | 0.0292 | |
| DLinear | Time-o1 | 0.4230 ± 0.0018 | 0.4150 ± 0.0022 | +0.0080 | 0.0027 | 0.0292 | |
| Weather | iTransformer | PSLoss | 0.2490 ± 0.0014 | 0.2440 ± 0.0017 | +0.0050 | 0.0042 | 0.0292 |
| PatchTST | DBLoss | 0.2350 ± 0.0016 | 0.2280 ± 0.0019 | +0.0070 | 0.0027 | 0.0292 | |
| TimeMixer | DBLoss | 0.2460 ± 0.0020 | 0.2370 ± 0.0025 | +0.0090 | 0.0031 | 0.0292 | |
| DLinear | FreDF | 0.2430 ± 0.0015 | 0.2400 ± 0.0018 | +0.0030 | 0.0386 | 0.1543 | |
| ECL | iTransformer | Time-o1 | 0.1760 ± 0.0020 | 0.1670 ± 0.0025 | +0.0090 | 0.0031 | 0.0292 |
| PatchTST | DBLoss | 0.1770 ± 0.0016 | 0.1710 ± 0.0019 | +0.0060 | 0.0037 | 0.0292 | |
| TimeMixer | FreDF | 0.1800 ± 0.0015 | 0.1780 ± 0.0018 | +0.0020 | 0.0957 | 0.1914 | |
| DLinear | DBLoss | 0.1660 ± 0.0013 | 0.1660 ± 0.0017 | +0.0000 | 1.0000 | 1.0000 |
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
PRODA and DistDF both incorporate distributional information from joint trajectories, z=(x,y) and z^=(x,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, 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 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) and O(CBN+CB2+CK), respectively, where B, C, N, and K 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) and O(CBN+CN2), respectively. Since PRODA uses a B×B pairwise distance matrix instead of an N×N covariance matrix, it may have a relatively advantageous computational structure when B<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.