- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
PeriCal learns a reusable periodic template and adapts it to the current input using position-dependent temporal offsets and variable-specific amplitude calibration before combining the calibrated periodic signal with local observations.
Critical strengths
The motivation is clear and directly connected to the proposed mechanism. The separation between temporal calibration and value calibration is intuitive, and the corresponding ablations consistently support both components.
Critical weaknesses
- The improvement over strong forecasting models is not sufficiently comprehensive. Although the paper includes several reasonably recent baselines, including TimeMosaic, TimeBridge, TQNet, and RAFT, the overall advantage is mixed rather than dominant. On several datasets PeriCal is close to or worse than competing methods, and the evidence is not sufficient to establish a clear state-of-the-art forecasting contribution.
- Traffic evaluation remains incomplete despite the strong traffic motivation. The paper evaluates PEMS03/04/07/08, which is useful, but does not include METR-LA, PEMS-BAY, or LargeST. These datasets provide stronger tests of large-scale cross-sensor structure and spatial heterogeneity. Reporting WAPE in addition to MSE/MAE would also facilitate comparison with traffic forecasting literature.
- The qualitative evidence does not directly validate the claimed calibration mechanism. Figures 3–4 visualize learned offsets, scaling factors, and output forecasts, but do not establish that the learned offsets correspond to actual timing shifts or that the learned value factors recover meaningful amplitude changes. The authors explicitly acknowledge this limitation. A controlled synthetic experiment with known temporal deformation and magnitude perturbation would provide much stronger mechanism validation.
- Representation-level evidence is limited. Existing MTS analyses, e.g. CPiRi and BasicTS, have shown that representation geometry and spatial/temporal heterogeneity can reveal what a forecasting model actually learns. A representation-space analysis, e.g., channel embeddings or clustering under different regimes, would make the claimed adaptive-periodic mechanism substantially more convincing than forecast examples alone. Prior benchmarking work has explicitly shown that temporal and spatial heterogeneity strongly affect MTS model behavior.
- Efficiency reporting could be expanded. The paper appropriately reports parameters, MACs, and epoch time, but inference latency, peak memory, and scaling with the number of variables would be useful, particularly because the method is positioned as lightweight.
Questions for authors
Can the authors evaluate on METR-LA, PEMS-BAY, or LargeST and report WAPE where appropriate? Can the learned temporal offsets be validated against synthetic or controlled ground-truth shifts? Can the authors provide representation-level analysis showing how calibration changes channel or temporal representations? How does inference cost scale with the number of variables?
4 comments
Thank you for the response and the additional experiments. The GTR comparison, repeated-run results, and controlled perturbation experiments address part of my concerns. However, my main concerns about the empirical contribution, traffic evaluation, and practical efficiency remain unresolved.
1. Traffic evaluation and metrics. The suggested evaluations on METR-LA, PEMS-BAY, or LargeST with 8,600 nodes, together with WAPE where appropriate, have been deferred to future work. Consequently, the concerns about evaluation breadth and applicability to spatially heterogeneous and larger-scale traffic settings remain unresolved. Restricting the conclusions to the existing PEMS benchmarks clarifies the scope, but does not supply the missing empirical evidence.
2. Practical efficiency and scaling. The explanation that forward computation scales linearly with the number of variables under fixed settings is useful. However, it does not establish actual inference latency, peak memory, or computational cost across different node counts. Reporting efficiency on Electricity and forecasting accuracy on PEMS07 does not constitute an empirical scaling study. Performance at substantially larger sensor counts also remains unverified. The practical accuracy–efficiency trade-off and scalability therefore remain insufficiently supported.
I acknowledge the additional evidence for the calibration mechanism, but these broader empirical limitations remain significant. I therefore maintain my rating of 2.
Please also clearly mark the substantive changes in the revised manuscript, for example using colored text, and identify the added or modified sections and tables. The revisions are currently difficult to distinguish from the original submission, which makes verification unnecessarily difficult.
Thank you for specifying the remaining empirical questions. We have added evaluations on METR-LA and PEMS-BAY, WAPE on four traffic datasets, and measured inference cost across sensor counts. The substantive additions are marked in blue in the revised PDF: Appendix H (Table 10, traffic repeated-run comparison), Appendix I (Table 11, inference cost and scaling), and Appendix J (Table 12, per-seed horizon-wise results).
Traffic evaluation and WAPE. PeriCal and the released GTR implementation were evaluated under matched settings with seeds 2024, 2025, and 2026. Each entry below is the mean ± sample standard deviation of the four-horizon average within each seed; WAPE is a percentage. Lower is better. The full table is Appendix Table 10.
| Dataset | MSE: PeriCal / GTR | MAE: PeriCal / GTR | WAPE (%): PeriCal / GTR |
|---|---|---|---|
| PEMS04 | 0.0888±0.0013 / 0.0973±0.0016 | 0.1928±0.0021 / 0.2000±0.0007 | 9.731±0.105 / 10.085±0.016 |
| PEMS08 | 0.1099±0.0039 / 0.1456±0.0012 | 0.2007±0.0015 / 0.2258±0.0011 | 8.172±0.103 / 9.208±0.095 |
| METR-LA | 0.7967±0.0043 / 0.8193±0.0037 | 0.4724±0.0034 / 0.4807±0.0056 | 17.090±0.126 / 17.382±0.196 |
| PEMS-BAY | 0.4858±0.0012 / 0.5068±0.0003 | 0.3287±0.0006 / 0.3331±0.0006 | 4.068±0.004 / 4.118±0.006 |
Across all four traffic datasets, PeriCal has lower mean MSE, MAE, and WAPE than reproduced GTR. The same direction of improvement holds on both the original PEMS benchmarks and the newly added METR-LA and PEMS-BAY, so the benefit is not confined to a single traffic network or metric. On PEMS08, mean MSE falls by 24.5% and WAPE by 11.3% relative to GTR; the MSE reductions on METR-LA and PEMS-BAY are 2.8% and 4.1%. The latter gains are smaller but establish that the calibration mechanism remains useful on two additional sensor networks. We do not infer formal statistical significance from three seeds alone.
Practical inference cost and scaling. We profiled both model architectures on an NVIDIA RTX A6000 with input/output lengths 96/96, batch size 32, cycle length 288, hidden width 512, 20 warm-up forwards, and 250 CUDA-synchronized timed forwards. We used randomly initialized weights and preloaded random GPU inputs to isolate operator cost. Latency is the median; memory is peak PyTorch allocated memory. Dataset labels indicate matched variable counts, not timing on raw dataset batches. The last three rows are synthetic input-size stress tests (Appendix Table 11).
| Scale | Variables | GTR / PeriCal (ms) | Extra (ms; %) | GTR / PeriCal peak MiB |
|---|---|---|---|---|
| PEMS08 | 170 | 1.093 / 1.192 | 0.099; 9.1% | 53.63 / 53.68 |
| METR-LA | 207 | 1.199 / 1.390 | 0.191; 15.9% | 62.79 / 62.84 |
| PEMS04 | 307 | 1.547 / 1.936 | 0.389; 25.1% | 89.97 / 90.02 |
| PEMS-BAY | 325 | 1.589 / 1.864 | 0.275; 17.3% | 92.43 / 92.48 |
| Stress test | 883 | 3.822 / 4.169 | 0.347; 9.1% | 232.49 / 232.54 |
| Stress test | 2,048 | 8.139 / 8.523 | 0.384; 4.7% | 518.32 / 518.37 |
| Stress test | 8,600 | 33.398 / 35.114 | 1.716; 5.1% | 2141.52 / 2141.57 |
At the four traffic-sized settings, PeriCal adds 0.099–0.389 ms per batch over GTR, with a 0.05 MiB difference in peak allocated memory. Together with the consistent traffic forecasting improvements above, these measurements show that the accuracy gain is obtained with a small absolute inference and memory overhead at the evaluated sensor counts. At 8,600 variables, the additional latency is 1.716 ms (5.1%), indicating that the calibration branches remain computationally tractable as channel count grows. This last result is a forward-pass stress test, not an accuracy evaluation on LargeST.
Thank you for the substantial additional experiments and the clearly marked revisions.
The new evaluations on METR-LA and PEMS-BAY, together with the WAPE results and measured inference costs across different variable counts, substantially address my previous concerns about evaluation breadth and practical efficiency. I appreciate the authors' efforts to strengthen the empirical evidence.
Nevertheless, I retain some reservations about the overall forecasting capability and practical significance of the improvements. In particular, the reported WAPE of 17.09% on METR-LA raises concerns about PeriCal's competitiveness relative to strong MTSF methods such as STID and CPiRi, which have reported substantially lower errors under different evaluation settings. Although these results are not directly comparable due to differences in forecasting horizons and protocols, a performance gap remains a concern that warrants further evaluation under matched settings. Moreover, the WAPE improvements over GTR on METR-LA and PEMS-BAY are relatively modest (0.292 and 0.050 percentage points, respectively). The methodological contribution also remains closely related to GTR.
Overall, the additional experiments have meaningfully improved my assessment of the paper. I am therefore willing to raise my rating from 2 to 3, while retaining the above reservations regarding forecasting competitiveness, practical improvements, and methodological novelty.
We thank the reviewer for recognizing the motivation and the consistent component ablations. The revision adds direct evidence for the calibration mechanism and clarifies the scope of our forecasting claims.
Forecasting performance and traffic scope. We agree that PeriCal does not dominate every strong baseline. Table 1 reports the best average MSE on six of 11 datasets and the best MAE on five; Sec. 4.2 also identifies settings where TimeMosaic or CycleNet is stronger. Table 1 now includes a reproduced GTR comparison on all 11 datasets, and Table 2 reports three-seed comparisons on five. We characterize the results as strong performance in the evaluated settings, not universal state-of-the-art performance. The traffic results concern PEMS03/04/07/08 under the paper’s MSE/MAE protocol. We plan to evaluate METR-LA, PEMS-BAY, and LargeST and report WAPE in follow-up work, extending the assessment to spatially structured and larger-scale traffic settings. The present conclusions remain limited to the four evaluated PEMS benchmarks
Controlled validation of calibration. We agree that forecast examples and factor visualizations alone cannot show whether the factors respond to known timing and magnitude changes. We added a held-out paired benchmark with externally specified timing-only, value-only, and joint perturbations at three strengths (Table 5; protocol in the appendix). Both models are trained for forecasting, without direct supervision of PeriCal’s factors. As perturbation strength increases, PeriCal’s MSE gain over GTR rises from 1.41% to 3.60% to 7.24% for timing, from 1.94% to 3.83% to 6.70% for value, and from 2.08% to 5.83% to 9.59% for joint changes. Changes in the learned factors correlate positively.
Internal representations. A regime-wise embedding analysis could offer a different perspective. For the specific calibration claim, Table 5 compares the relevant internal factors with independently known perturbations. Appendix Figure 5 additionally compares GTR and PeriCal using the same-scale gradient sensitivity of forecasts to the periodic cue on ETTm1. We interpret this as complementary evidence about the cue’s influence on predictions, rather than as a channel-embedding or clustering result.
Inference scaling. At fixed look-back length, prediction horizon, cycle length, and hidden width, PeriCal forms a cross-channel temporal summary and then applies retrieval, value calibration, fusion, and the shared predictor per variable. It does not construct pairwise channel attention or a sensor graph. The template contains P × C values, and forward arithmetic scales linearly with the variable count C under these fixed settings. Table 3 reports parameters, forward MACs, and epoch time on Electricity (C=321); PEMS07 evaluates forecasting with 883 variables. This explains architectural scaling, while measured inference latency and peak memory remain separate empirical questions.