acceptodds
Under review as a conference paper at ICLR 2027

PeriCal: Adaptively Calibrating Periodic Patterns for Time-Series Forecasting

Abstract

Periodic patterns provide important temporal information in multivariate time-series forecasting by capturing regular variations in real-world systems. However, periodic patterns do not recur identically across contexts: their timing and magnitude may vary with evolving system states and external conditions. Existing methods typically model periodic patterns with fixed representations, which may become inconsistent with current observations when such variations occur. To address this limitation, we propose PeriCali, a lightweight forecasting framework that adapts periodic patterns through Adaptive Periodic Calibration. PeriCal learns input-conditioned calibration factors to adjust periodic patterns along both temporal and value dimensions, enabling the periodic representation to better match the current temporal context before being integrated with local observations through two-dimensional convolution. Experiments on 11 real-world benchmarks demonstrate that PeriCal achieves strong forecasting performance across datasets from multiple application domains. Further analysis verifies the effectiveness of the proposed adaptive periodic modeling approach.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, multivariate time series, periodicity modeling, adaptive periodic calibration, temporal alignment, adaptive representation learning

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_dkBKOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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.
  4. 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.
  5. 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

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.

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.

DatasetMSE: PeriCal / GTRMAE: PeriCal / GTRWAPE (%): PeriCal / GTR
PEMS040.0888±0.0013 / 0.0973±0.00160.1928±0.0021 / 0.2000±0.00079.731±0.105 / 10.085±0.016
PEMS080.1099±0.0039 / 0.1456±0.00120.2007±0.0015 / 0.2258±0.00118.172±0.103 / 9.208±0.095
METR-LA0.7967±0.0043 / 0.8193±0.00370.4724±0.0034 / 0.4807±0.005617.090±0.126 / 17.382±0.196
PEMS-BAY0.4858±0.0012 / 0.5068±0.00030.3287±0.0006 / 0.3331±0.00064.068±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).

ScaleVariablesGTR / PeriCal (ms)Extra (ms; %)GTR / PeriCal peak MiB
PEMS081701.093 / 1.1920.099; 9.1%53.63 / 53.68
METR-LA2071.199 / 1.3900.191; 15.9%62.79 / 62.84
PEMS043071.547 / 1.9360.389; 25.1%89.97 / 90.02
PEMS-BAY3251.589 / 1.8640.275; 17.3%92.43 / 92.48
Stress test8833.822 / 4.1690.347; 9.1%232.49 / 232.54
Stress test2,0488.139 / 8.5230.384; 4.7%518.32 / 518.37
Stress test8,60033.398 / 35.1141.716; 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.

Reviewer_39xHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PeriCal, a lightweight long-term forecaster built on a learnable global periodic template indexed by the absolute cycle position of each input step. Adaptive Periodic Calibration adapts the retrieved template in two ways: a temporal calibration that predicts a bounded, smoothed, variable-shared continuous offset and resamples the template by cyclic interpolation (Eqs. 3-7), and a value calibration that scales the cue by an input-conditioned, variable-specific factor in (0,2) (Eqs. 8-9). The calibrated prior is linearly mapped, fused with the input by a 2D convolution and passed to a residual MLP (Eqs. 10-13). Experiments cover 11 benchmarks with L = 96 (Tables 1/7), an ablation on ETTm1/Weather/PEMS08 (Table 3), efficiency (Table 2), visualizations (Figs. 3-4) and a 5% few-shot setting (Tables 5-6).

Critical strengths
  1. Resampling a learned periodic prior with a smooth, input-conditioned continuous offset (cyclic interpolation, Eqs. 6-7) is simple, differentiable and cheap (~40K extra parameters). If its benefit were demonstrated, it would be an easy plug-in for template/retrieval-based forecasters.
  2. The paper is unusually careful about what its visualizations do and do not show (Sec. 4.5, Sec. 5), which makes the claims easy to assess.
Critical weaknesses

The motivation is not empirically established and the ablation does not isolate the key idea. There is no stratification of errors by phase shift or amplitude change; learned offsets are tiny (mean |δ| = 0.28 steps on ETTm1, Fig. 3a) and value factors in Fig. 3b look saturated near 0 and 2, i.e., gating rather than calibration.

The closest prior work, GTR (Cao et al., ICLR 2026), is missing from Tables 1, 2 and 7, although PeriCal shares almost its entire pipeline (learnable global template indexed by cycle position, linear mapping, 2D-convolution fusion, residual MLP, same cycle lengths) and has nearly the same cost (1.02M / 309.89M vs. 0.98M / 306.91M parameters/MACs).

Questions for authors
  1. Please compare with GTR (Cao et al., ICLR 2026) under identical settings, with mean ± std over multiple seeds.
  2. Please stratify test windows by phase/amplitude mismatch to show that the gains come from calibration.
3 comments

We thank the reviewer for recognizing the differentiable, low-cost temporal calibration mechanism and for identifying the two missing evaluations.

Q1: Comparison with GTR. We agree that GTR is the closest baseline and that its omission obscured the contribution of calibration. We reran its released implementation using matched benchmark settings: the same data splits, look-back length (96), cycle lengths, prediction horizons, and seeds. Revised Table 1 reports seed-2024 averages on all 11 datasets, and Appendix Table 9 gives the horizon-wise results. Revised Table 2 reports mean ± sample standard deviation over seeds 2024, 2025, and 2026 on ETTh1, ETTm1, PEMS04, PEMS08, and Solar; Appendix Table 10 provides the individual runs. PeriCal has lower mean MSE and MAE on each of these five repeated-run datasets. For example, PEMS08 MSE is 0.1099 ± 0.0039 for PeriCal versus 0.1456 ± 0.0012 for GTR; on ETTm1, it is 0.3653 ± 0.0008 versus 0.3791 ± 0.0013. PeriCal retains GTR’s global template, mapping, 2D-convolution fusion, and residual MLP. The added operations are input-conditioned continuous temporal resampling and variable-specific value modulation. Its parameter and MAC counts are slightly higher than GTR’s; we do not claim a complexity advantage over GTR.

Q2: Behavior under increasing mismatch. We agree that Figure 3’s factor visualizations and the mean absolute offset of 0.28 steps do not, by themselves, establish a mismatch-dependent benefit. Table 4 removes each calibration branch separately while keeping the remaining forecaster fixed, but forecasting ablations alone cannot establish whether the factors respond to actual timing and value changes. Real benchmarks do not supply ground-truth mismatch for each test window, and grouping windows by PeriCal’s own predicted factors would make the analysis circular. We therefore added a held-out, paired benchmark with known timing-only, value-only, and joint perturbations at three predefined strengths (Table 5; generation details in the appendix). Neither model receives factor supervision. As perturbation strength increases from 0.25 to 0.50 to 1.00, 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, though moderately, with the injected changes (temporal: 0.24–0.51; value: 0.29–0.44). Table 5 provides a quantitative check: 22.7–33.6% of value factors lie near the bounds, while most remain within the interior of the allowed range. Changes in these factors are also positively correlated with the known injected amplitude changes (0.29–0.44), and PeriCal’s gain over GTR increases as the value perturbation becomes stronger. These results suggest a graded response to amplitude variation rather than a purely saturated on/off gate.

Thanks. The results compared with GTR address some of my concerns, but it also makes me more doubtful that the novelty is enough for an ICLR paper? I agree with your points on Q2. But, I still have serious doubts about how this work differs from GTR, especially in novelty, so I’d like to see the conceptual and technical novelty more clearly. Now, I will maintain the score.

Thank you for the follow-up. We agree that PeriCal retains GTR’s global template, time indexing, 2D-convolution fusion, and MLP predictor. We do not claim a new forecasting backbone. The distinction is what determines the periodic cue before fusion. Let rt=(τ+t) mod Pr_t=(\tau+t)\bmod P. At the retrieval stage, the two methods can be summarized as

Rc,tGTR=Qrt,cR^{\mathrm{GTR}}_{c,t}=Q_{r_t,c}

Rc,tPeriCal=ac,t(Z)⋅Interpcyc(Q⋅,c,rt+δt(Z))R^{\mathrm{PeriCal}}_{c,t}=a_{c,t}(\mathbf{Z})\cdot\mathrm{Interp}_{\mathrm{cyc}}\big(\mathbf{Q}_{\cdot,c},\quad r_t+\delta_t(\mathbf{Z})\big)

Thus, inputs with the same cycle index retrieve the same template values in GTR, whereas PeriCal conditions both the sampling coordinates and the retrieved values on the current input Z\mathbf{Z}.

This is an operator-level difference, not simply two layers appended to the predictor. The bounded, smoothed, position-dependent offset permits differentiable resampling at continuous cyclic coordinates; different positions need not undergo the same shift. The variable- and position-specific value factor then calibrates the resampled prior, before its fusion with local observations. GTR’s downstream network may learn to compensate for some mismatch—we do not claim otherwise—but its retrieval coordinates themselves are fixed by the cycle index.

To distinguish this mechanism from additional model capacity, we compare against a parameter-matched GTR control:

DatasetPeriCal MSEPeriCal MAEGTR MSEGTR MAEMatched GTR MSE*Matched GTR MAE*
ETTm10.3653 ± 0.00060.3890 ± 0.00100.3790 ± 0.00100.3983 ± 0.00150.3763 ± 0.00120.3960 ± 0.0010
PEMS040.0890 ± 0.00100.1927 ± 0.00210.0973 ± 0.00150.2003 ± 0.00060.0950 ± 0.00100.1977 ± 0.0012
PEMS080.1103 ± 0.00380.2007 ± 0.00150.1457 ± 0.00120.2260 ± 0.00100.1423 ± 0.00060.2237 ± 0.0006
Solar0.1943 ± 0.00060.2457 ± 0.00060.2040 ± 0.00100.2560 ± 0.00170.2023 ± 0.00060.2537 ± 0.0015

* Parameter-matched GTR control. Values are mean ± sample standard deviation across three seeds; each seed’s value is averaged over four forecasting horizons. Lower is better.

PeriCal has lower MSE and MAE than the matched control on all four datasets. Its relative average-MSE reductions are 2.9% on ETTm1, 6.3% on PEMS04, 22.5% on PEMS08, and 4.0% on Solar. Matching the parameter budget narrows the gap but does not remove it. Together with the branch ablations and controlled perturbations, this supports the value of input-conditioned calibration of retrieval itself, rather than attributing the gains solely to a larger predictor.

Our claim is deliberately specific: PeriCal models a recurring pattern as a reusable template whose timing and magnitude can depend on the present input. It does not replace GTR’s overall forecasting pipeline. We hope the distinction between fixed indexed retrieval followed by input-aware fusion and input-aware retrieval before fusion clarifies the conceptual and technical contribution, and we would appreciate reconsideration of the novelty assessment.

Reviewer_3wWNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attempts to address the problem that fixed periodic representations fail to adapt to dynamic variations in time-series forecasting. The authors propose PeriCal, a forecasting method that automatically calibrates the timing and magnitude of periodic patterns based on current input and fuses the calibrated patterns with local observations to improve prediction accuracy.

Critical strengths
  1. It proposes a two-dimensional adaptive periodic calibration mechanism across temporal and value dimensions, which alleviates the poor adaptability of fixed periodic representations.
  2. It delivers strong forecasting performance on 11 real-world benchmark datasets, with notable accuracy improvements on some traffic datasets.
  3. It adopts a dual-path fusion design of periodic priors and local observations, capturing both long-term periodic patterns and short-term fluctuations.
  4. The model features a lightweight architecture, with fewer parameters and lower computational costs in terms of MACs than Transformer-based methods.
Critical weaknesses
  1. The paper mainly adds temporal and value calibration to existing periodic modeling approaches, with limited innovation, and lacks a direct forecasting performance comparison with the similar GTR method.
  2. The cycle length must be set manually, and the model explicitly considers only a single cycle.
  3. There is no direct evidence that the learned calibration factors capture actual changes in cycle timing and magnitude.
  4. The model shows little improvement on some datasets, and even performs worse than existing methods; the results lack statistical significance tests.
Questions for authors
  1. What are the core innovations of PeriCal compared with the similar GTR method? Could you provide a direct comparison of forecasting performance?
  2. Can this method remain effective when the cycle length changes or multiple overlapping cycles exist?
  3. Could you directly demonstrate through experiments that the learned calibration factors capture actual changes in cycle timing and magnitude?
  4. Could you provide results from multiple independent runs and statistical significance tests to verify the stability and reliability of the forecasting improvements?
1 comment

We thank the reviewer for recognizing the adaptive calibration design and for identifying the evidence needed to assess it. We address the four questions with a reproduced GTR comparison (Table 1), three-seed results (Table 2), and controlled perturbations (Table 5).

Q1. Difference from GTR. PeriCal retains GTR's learnable global template, time-indexed retrieval, 2D-convolutional fusion, and MLP backbone. It adds input-conditioned calibration before fusion: a position-dependent offset changes the template coordinates through cyclic interpolation, while variable- and position-dependent factors adjust the retrieved values (Sec. 3.2). Two inputs with the same cycle index can therefore receive different periodic cues. In our rerun of the released GTR implementation, PeriCal has lower four-horizon average MSE on 11/11 datasets and lower MAE on 10/11 with seed 2024 (Table 1). The appendix reports every horizon; Table 2 compares three matched seeds.

Q2. Changing or overlapping periods. We set one nominal period PP per dataset from its sampling interval and an expected daily or weekly cycle, and use it for every horizon (Appendix, Dataset Details). A length-PP template can encode components whose periods divide PP; for example, a 168-hour template can contain both daily and weekly structure, although it does not separate them. The position-dependent offset δt\delta_t adjusts the sampling coordinate locally within ∣δt∣≤Δmax⁡|\delta_t|\leq\Delta_{\max}. This permits bounded timing drift around the nominal period, not arbitrary changes to PP. We have not tested changing fundamental periods or independently varying, noncommensurate cycles; Sec. 5 states this limit.

Q3. Factor validity. We added paired clean and perturbed multi-harmonic trajectories with known timing offsets, channel-specific value changes, or both (Table 5; Appendix, Controlled Synthetic Calibration Details). Each pair shares its trend and noise, and the factors are not supervised. On held-out pairs, changes in learned offsets correlate with injected timing shifts (Pearson rδ=0.32r_\delta=0.32–0.480.48 in timing-only conditions); changes in value factors correlate with injected value changes (rA=0.31r_A=0.31–0.420.42 in value-only conditions). PeriCal's MSE gain over GTR rises with perturbation strength, reaching 7.24% for timing, 6.70% for value, and 9.59% for both. The positive, moderate correlations support sensitivity to injected changes, not exact factor recovery.

Q4. Multiple runs and significance. We trained both models with seeds 2024, 2025, and 2026 on ETTh1, ETTm1, Solar-Energy, PEMS04, and PEMS08. Table 2 reports the mean ±\pm sample standard deviation of each seed's four-horizon average; Appendix Table 10 lists every run. PeriCal has lower MSE and MAE for every matched seed on all five datasets. Its MSE is 0.4365±0.00160.4365\pm0.0016 versus 0.4448±0.00290.4448\pm0.0029 on ETTh1 and 0.1099±0.00390.1099\pm0.0039 versus 0.1456±0.00120.1456\pm0.0012 on PEMS08. The three matched seeds show a consistent direction of improvement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.