acceptodds
Under review as a conference paper at ICLR 2027

STEPS: Spectral–Temporal Error Propagation for Test-Time Adaptation in Time-Series Forecasting

Abstract

Test-time adaptation (TTA) for time-series forecasting must revise a forecast from a short prefix of delayed targets that may contain transient changes or corrupted measurements. Existing methods often gain more at long horizons than at short horizons and are sensitive to unreliable observations, as noisy or transient errors can be propagated into the unseen suffix. We observe that forecast errors exhibit two complementary structures: persistent cross-window bias is concentrated in low-frequency modes, while newly revealed residuals contain useful short-term structure whose influence on future errors must be selectively propagated rather than indiscriminately amplified. Based on these observations, we propose STEPS, a Spectral–Temporal Error Propagation Solver that adapts a frozen forecaster in output space. Multiscale Spectral Bias Reconstruction (MSBR) recovers persistent bias from historical residual spectra; Short-Term Temporal Error Propagation (ST-TEP) learns how the revealed prefix should influence the unseen suffix through an empirical temporal kernel; and Orthogonal Complementary Correction (OCC) removes their overlapping component before fusion. Across six datasets, seven backbones, and four horizons, STEPS delivers a 24.85% average MSE reduction, which is more than twice the gain achieved by the strongest TTA baseline. Under 1–30% Gaussian prefix corruption, STEPS significantly outperforms competing methods. These results demonstrate accurate and robust forecast revision across temporal scales without online backbone optimization.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Test Time Adaptation, Robustness, Residual Correction, Spectral and Temporal Modeling

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_3vu6OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STEPS revises a frozen forecaster’s outputs by combining historical spectral bias estimation, prefix-conditioned error propagation, and orthogonal fusion, with promising reported gains but unresolved concerns about experimental consistency and comparison fairness.

Critical strengths
  1. STEPS provides a practical way to revise forecasts from delayed feedback without updating the backbone or performing online gradient optimization.
  2. The evaluation spans six datasets, seven backbones, and four horizons, reporting a substantial average MSE reduction of 24.85%.
  3. The spectral diagnostics, shuffled-prefix controls, and recent-change analysis provide evidence that forecast residuals contain structure useful for correction.
Critical weaknesses
  1. The comparison does not clearly establish equal feedback availability, revision timing, and access to training residuals across methods.
  2. The ablations suggest that local error propagation accounts for most gains, leaving the necessity of spectral reconstruction and orthogonal fusion insufficiently demonstrated.
  3. Robustness testing is limited to Gaussian prefix corruption on one backbone and horizon, leaving sensitivity to recent outliers, missing observations, and persistent disturbances unclear.
Questions for authors
  1. Do the improvements persist when all methods receive the same feedback and training information and are evaluated at the same revision time?
  2. How does STEPS compare with a direct ridge predictor mapping observed prefix residuals to future residuals under matched regularization?
  3. Under what conditions do MSBR and OCC consistently improve over ST-TEP alone and validation-tuned direct addition?
  4. How does STEPS behave when prefix corruption is temporally concentrated or consists of outliers, missing observations, or persistent shifts?
  5. How does STEPS relate to TameR’s spectral approach to recent-data robustness (Xu et al., 2026), and does it provide additional gains when applied to a TameR backbone?

Xu, L., et al. (2026). Taming the Recent-Data Bias: Towards Robust Time Series Forecasting with Global Context. ICML 2026.

2 comments
AuthorsOpenReview ↗

Response to 3vu6: conditions for complementary gains and controlled evidence

Thank you for the constructive questions. We distinguish the effects that are directly supported by the submitted experiments from comparisons that require additional evidence.

Q3: When do the global branch and OCC help? The clearest reported setting is Table 2: DLinear, six datasets, H=720, a=320, and MSE on the remaining 400 steps under 1-30% Gaussian prefix corruption. At every tested level, full STEPS has lower mean MSE than both local-only ST-TEP and Additive. At 1% corruption, Additive/ST-TEP/OCC are 0.4160/0.3601/0.3584; at 30%, 0.7644/0.7066/0.6938. OCC lowers MSE relative to Additive by 9.24%-13.85% across the five levels. Its additional improvement over local-only is smaller, 0.47%-1.81%, but positive at every level.

These two comparisons answer different questions. Removing OCC while keeping both branches gives Additive and tests overlap removal. Keeping only ST-TEP removes the global information as well and tests complementarity. ST-TEP propagates the current observed residual pattern; MSBR reconstructs persistent bias from multiscale residual history and current descriptors. Their corrections can overlap, so direct addition can apply the same direction twice. OCC retains the local response and adds only the global complement (up to the numerical stabilizer). The Table 2 ablation supports this fusion design: direct addition is worse than local-only, whereas the orthogonal complement yields a combined correction that improves over local-only.

The clean H=192 results in Table A10 demonstrate the local branch's near-term strength: 50.7% near-suffix reduction. They do not establish that persistent-bias correction or overlap removal is redundant at H=720 under perturbed feedback. We also do not attribute the entire cross-table difference to horizon length, since the protocols differ in other respects. These are observed conditional benefits, not a claim that every component wins in every cell.

Q1: Same information and revision time. The existing evidence separates two points. First, recomputation of the reported Table A14 paired results gives 25.05% mean full-horizon reduction over the corresponding frozen forecasts, with improvements in all 168 pairs. Second, Table A10 controls how the same revealed prefix is used: temporal permutation reduces near-suffix gain from 50.7% to 27.9%, and last-residual continuation gives 1.0% full-suffix improvement versus 37.1% for ST-TEP. Thus, useful residual structure matters beyond simply receiving observations. These findings do not establish matched training access, feedback schedules and revision times for all competing methods; the published cross-method entries must not be treated as resolving that separate question.

Q2: Direct prefix-to-suffix ridge. Appendix B derives ST-TEP's exact latent-ridge representation: with F=Phi_m A^(1/2), minimizing ||F_B u-r_B||^2 + eta||u||^2 yields F_U u=K_UB(K_BB+eta I)^(-1)r_B. This explains the regularized inference in the empirical residual subspace. It does not show an advantage over a direct unrestricted ridge predictor trained on prefix/suffix pairs. The submission does not report that matched comparison. Such a comparison must align residual normalization, centering/intercepts, channel pooling, validation budget and scored targets, rather than equating penalty values across different scalings.

Q4: Other disturbances. Table 2 supports the stated Gaussian-corruption result. Table A11 additionally examines naturally occurring large recent changes: on that subset, ST-TEP and STEPS achieve 41.1% and 40.4% near-suffix improvement, respectively. This is evidence of responding after a change becomes observable, not a controlled persistent-shift experiment. Temporally concentrated corruption, missing observations and other outlier processes are not established by these tests. In particular, finite linear sensitivity is not an arbitrary-outlier robustness guarantee.

Q5: Relation to TameR. TameR addresses recent-data bias within the forecasting model using global context and spectral/periodicity modeling. STEPS intervenes at a different stage: it revises an already-issued forecast after partial targets become available, without updating the backbone. A frozen TameR model could supply the forecasts and calibration residuals needed by STEPS, but compatibility alone does not establish extra accuracy; TameR may change the residual structure available to the corrector. No TameR+STEPS experiment is reported in the submission.

We hope the distinction between global complementarity and overlap removal, together with the explicit long-horizon ablation numbers, clarifies the role of the complete framework.

Thanks for the response, which address most of my comments. I will raise my score.

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

The paper proposes STEPS, an output-space test-time adaptation method that corrects predictions from a frozen time series forecaster using historical residuals and partially observed targets. It combines Multiscale Spectral Bias Reconstruction (MSBR) to estimate persistent forecasting bias, Short-Term Temporal Error Propagation (ST-TEP) to propagate newly observed residuals into the unobserved future, and Orthogonal Complementary Correction (OCC) to reduce overlap between the two corrections. STEPS demonstrates substantial improvements over existing TTA baselines across diverse datasets, forecasting backbones, and horizons, with additional experiments evaluating component contributions, prefix sensitivity, and robustness to observation noise.

Critical strengths

Well-motivated problem formulation. The paper addresses delayed forecast revision using partial ground-truth observations. Distinguishing historical systematic bias from recent prediction errors provides an intuitive framework for exploiting complementary information.

Strong empirical performance. STEPS improves upon frozen forecasts across all 168 configurations and outperforms COSA-F in 146, demonstrating effectiveness across Transformer, linear, and MLP-based forecasting architectures.

Simple and interpretable correction mechanism. The approach uses spectral representations, ridge regression, and covariance conditioning without online backbone optimization, providing a relatively transparent alternative to gradient-based adaptation.

Useful diagnostic analysis. The paper includes spectral diagnostics, shuffled-prefix controls, component ablations, and prefix sensitivity experiments. The appendices also acknowledge configurations where STEPS underperforms and where individual components contribute little.

Critical weaknesses
  1. ST-TEP essentially performs regularized covariance conditioning, $KUB(KBB+ηI)−1RBK_{UB}(K_{BB}+\eta I)^{-1}R_B $, while MSBR uses DCT-based ridge regression and OCC applies orthogonal projection. Similar principles appear in earlier dynamic forecast-updating methods. The claimed manifold formulation also relies on a fixed path-graph Laplacian rather than learned geometry, making its distinction from classical spectral and kernel-based estimation unclear.

  2. ST-TEP alone achieves nearly the same improvement as the complete STEPS framework (37.1% versus 37.4% full-suffix MSE reduction), while direct addition performs similarly. Although OCC improves robustness under corrupted observations, its advantage over ST-TEP remains small. This raises questions about whether MSBR and OCC provide sufficient complementary information to justify the additional complexity.

  3. ST-TEP conditions on raw residuals using uncentered second moments, potentially capturing the same persistent bias estimated by MSBR. OCC removes aligned correction directions, but orthogonality does not guarantee statistical independence or lower prediction error. A more principled alternative would use MSBR as a prior mean and perform covariance conditioning on the remaining residual, rather than applying a post-hoc projection.

  4. Despite emphasizing adaptation from limited observations, STEPS requires half the forecasting horizon �1��1� to be revealed in its main evaluation. Performance declines with shorter prefixes, and competing methods use different feedback schedules. This makes it difficult to separate the benefits of the proposed propagation mechanism from those of having substantial ground-truth information available.

  5. Baseline MSE values differ between Tables 1 and A14 despite claims of identical frozen checkpoints (e.g., PatchTST/Weather/96: 0.1742 versus 0.1976). Table A20 also reports unusually large seed variability for an offline-calibrated, closed-form method (e.g., 0.0471±0.07080.0471 \pm 0.0708). These discrepancies require clarification to establish the reliability and reproducibility of the reported improvements.

Questions for authors
  1. ST-TEP closely resembles GP conditioning and kernel ridge regression, as acknowledged in Appendix B. Related work on dynamic forecast updating (Shang & Hyndman, 2011; Shen & Huang, 2008) is also not discussed. Could the authors compare STEPS against GP conditioning with a fitted stationary kernel and FPCA/SVD-based ridge updating to better establish its methodological contribution?

  2. Table A10 shows that ST-TEP alone nearly matches STEPS (37.1% versus 37.4% MSE reduction), but this analysis covers only 12 configurations. Could the authors report ST-TEP alone and direct addition across all 168 configurations to establish whether MSBR and OCC consistently provide meaningful complementary gains?

  3. ST-TEP conditions on raw residuals using an uncentered kernel, potentially capturing the same persistent bias estimated by MSBR. Why is orthogonal projection preferred over treating MSBR as a prior mean and conditioning on the remaining residual $ \hat r_U=G_U+K^c_{UB}(K^c_{BB}+\eta I)^{-1}(r_B-G_B) $ where KcK^c is estimated from centered residuals? Comparing these formulations would help justify the proposed fusion mechanism.

  4. Do the competing methods revise the current forecast's unobserved suffix after observing its prefix, as STEPS does? Could the authors clarify how COSA-F's B=48B=48 relates to the available feedback and compare methods under matched prefix lengths and revision times? This would help distinguish methodological improvements from differences in observation availability.

  5. Table A9 primarily characterizes within-window residual structure, whereas MSBR assumes that low-frequency bias persists across forecasting windows. Could the authors report the autocorrelation or temporal stability of low-order DCT coefficients across successive windows to directly validate this assumption?

  6. Why do frozen-backbone MSE values differ between Tables 1 and A14 (e.g., PatchTST/Weather/96: 0.1742 versus 0.1976) despite the stated use of identical checkpoints? Additionally, what varies across TTA seeds in an otherwise deterministic, offline-calibrated method, and what explains the large standard deviations reported in Table A20?

  7. Appendix B acknowledges that the proposed operator is not a classical harmonic extension, while the temporal geometry is constructed from a fixed path-graph Laplacian. What specific theoretical or computational advantage does the manifold formulation provide beyond conventional spectral representations and covariance-based conditioning?

1 comment
AuthorsOpenReview ↗

Response to zPBK: complementary corrections, OCC, and evidence scope

Thank you for the review. We clarify the component roles and evidence in the submitted paper.

Q2: Distinct component roles and the appropriate ablations. ST-TEP exploits the currently observed residual pattern, which is particularly informative near the prefix boundary. MSBR reconstructs persistent bias using multiscale historical residual states and current descriptors, supplying information that current-prefix propagation may not fully capture farther into the suffix. OCC controls their interaction. Removing OCC while retaining both branches produces Additive; removing the global contribution produces ST-TEP alone. These are different ablations.

Table A10 establishes strong local performance on clean H=192 diagnostics: ST-TEP gives 50.7% near-suffix improvement and 37.1% full-suffix improvement; the full method gives 37.4% on the full suffix. Table 2 establishes a clearer fusion benefit in the H=720 corruption setting: OCC reduces MSE relative to Additive by 13.85%, 13.14%, 12.13%, 10.50% and 9.24% across the five corruption levels. Full STEPS also has lower mean MSE than local-only at every level. At 30%, Additive/ST-TEP/OCC are 0.7644/0.7066/0.6938. Thus, direct addition can lose the benefit of complementary global information, whereas overlap removal enables the combined correction to improve over the local branch. The H=192 average does not establish redundancy of these components in this longer-horizon setting. These results do not constitute the requested all-168-cell ablation, nor isolate horizon length from the other protocol differences.

Q3: Why remove the overlapping direction? The possibility that the two branches represent some of the same bias is precisely the motivation for OCC. With local response l and global response g, it uses l + beta*(g - alpha*l), where alpha=<l,g>/(||l||^2+epsilon). This removes the global component aligned with the local correction while retaining the remaining global information; preservation of the local direction is approximate because of epsilon. The mechanism is directional de-duplication, not statistical independence or a universal error guarantee. Table 2 supplies empirical evidence for its benefit over direct fusion.

The suggested prior-mean formulation makes a different modeling choice: G_U is added and G_B is subtracted before conditional propagation. For a fixed conditional operator T_c, a mean-estimation error delta G changes that prediction by delta G_U - T_c delta G_B. OCC instead preserves the local response direction while adding a complement. This explains the design choice, but is not a proof of superiority to prior-mean conditioning; the submission does not compare that alternative.

Q1 and Q7: Classical conditioning and the spectral interpretation. Appendix B explicitly derives the latent ridge form with F=Phi_m A^(1/2): minimizing ||F_B u-r_B||^2 + eta||u||^2 yields F_U u = K_UB(K_BB+eta I)^(-1)r_B. The contribution is the residual-modeling construction and its combination with historical bias correction, not a new conditioning identity. The empirical K models horizon-position-dependent residual relationships and is not constrained to be stationary. The path-graph interpretation specifies a temporal basis and a roughness bound; it does not establish learned geometry or superiority over conventional spectral estimators. The GP and FPCA/SVD comparators and dynamic-updating literature you identify are relevant; the identity is not a substitute for those experiments.

Q4: Half a horizon is a choice, not a requirement. Table A12 reports 14.48% and 20.22% full-horizon improvements with a=24 and a=48 at H=192. On the identical final 48 targets, the respective gains are 9.84% and 16.28%. This demonstrates useful shorter-prefix revision, but does not prove matched feedback schedules across methods; a=48 and B=48 are not equivalent by numerical equality alone.

Q5: What the spectral diagnostics establish. Table A9 supports within-window spectral concentration. Held-out MSBR branch performance supports predictive value in some settings. Neither directly measures cross-window autocorrelation of individual coefficients. That requested persistence diagnostic is distinct from the existing concentration result.

Q6: Denominator and seed interpretation. Recomputing the reported pairs in Table A14 gives 25.05% mean full-horizon reduction versus 24.85% with Table 1's Backbone entries, with improvement in all 168 pairs. This preserves the paired-backbone result, but does not validate the blanket identical-checkpoint statement. Table A20's randomization source cannot be established from its mean/std summaries; that requires seed-level provenance, separately from the within-series bootstrap in A19. We do not treat this comment as resolving that provenance question.

We hope this clarifies the demonstrated role of complementary correction.

Reviewer_vrbzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies test-time adaptation (TTA) for time series forecasting. After a forecast is issued, the first part of its targets (the prefix) becomes known, and the rest of the forecast (the suffix) can be revised. It proposes STEPS, which keeps the backbone frozen and corrects the suffix in output space using two closed-form branches: First branch is the Multiscale Spectral Bias Reconstruction (MSBR), a channel-wise ridge regression which predicts the low-frequency DCT spectrum of the residual from multiscale averages of past residual spectra and descriptors of the current input. The second branch is the Short Term Temporal Error Propagation (ST-TEP) fits a second-moment kernel to training residuals in a 64-mode DCT basis. It then extends the revealed prefix residual into the suffix with a regularized conditional-mean operator. Finally, the Orthogonal Complementary Correction (OCC) removes from the global response the part that points along the local response, then adds the rest with a validated weight. The method is tested on six datasets (ETTh1/h2/m1/m2, Exchange, Weather), seven frozen backbones and four horizons, with half the horizon revealed. It reports a 24.85% mean full-horizon MSE reduction, against 12.00% for COSA-F, the best TTA baseline.

Critical strengths
  1. The method is simple, needs no gradient steps at test time, and helps every backbone. Additionally, both branches are closed-form fits computed offline. STEPS beats its own frozen backbone in all 168 cells (Table A6). If the comparison with other TTA methods holds up, a cheap plug-in corrector of this kind would be useful in practice.

  2. The work is fair, as it shows where the method falls, which the shuffled-prefix control shows that the order of the prefix matters: near-suffix gain falls from 50.7% to 27.9% (Table A10). The paper also says plainly that MSBR alone is weak (3.1%), that OCC adds almost nothing on clean data (37.4% for both direct addition and STEPS), and lists the cells where STEPS loses (Appendix I). This helps place the contribution: most of the gain comes from conditioning on the prefix with an empirical kernel.

Critical weaknesses
  1. The main comparison does not use matched backbones. Section 4.1 states that "published and local baseline results use the same backbone weights, splits, and forecast origins", and Table 1 takes the TAFAS/PETSA/COSA entries (†) from the published COSA comparison. However, the STEPS rows are computed on local checkpoints whose own MSE differs from the published Baseline column. For example, PatchTST/Weather/96 is .1976 in Table A14 against .1742 in Table 1, and PatchTST/Exchange/720 is .9628 against .8648. So the STEPS numbers and the TAFAS/PETSA/COSA numbers start from different backbones, and the "more than twice the gain" claim compares results from different starting points. Rerunning all methods on shared checkpoints would fix this, but it could change the headline numbers.

  2. The gains cannot be separated from the extra information STEPS receives. STEPS observes ⌊H/2⌋ targets of the current window (360 steps at H = 720), while COSA-F uses B = 48 (Section 4.1). Two baselines are missing: (i) all TTA methods run with the same revealed-prefix budget per window, and (ii) re-running the frozen backbone on the latest L = 96 observations, revealed prefix included, to forecast the remaining suffix. At a = H/2, much of the reported gain may come from forecasting a shorter remaining horizon from fresher data rather than from the proposed correction. The experiments could add these baselines, but without them the contribution of STEPS is unclear.

Questions for authors
  1. Were STEPS and the † baselines run on the same backbone checkpoints? If so, why does the Baseline MSE in Table A14 differ from Table 1 in many cells (for example PatchTST/Weather/96: .1976 against .1742)? Are the relative gains in Eq. 9 measured against which backbone?

  2. How do the baselines use the current window's targets? Do TAFAS, PETSA and COSA also see ⌊H/2⌋ targets of the same window before scoring? Could results be given with every method on an equal prefix budget, for example a = 48 at all horizons?

  3. How does STEPS compare with re-running the frozen backbone on the updated lookback window (revealed prefix included) to forecast the remaining suffix, at the same "a"?

  4. Table 2 has no 0% corruption column and no significance test, yet the abstract says "significantly outperforms". Can both be added? Also, why does direct addition fall well behind ST-TEP under noise (0.4160 against 0.3601 at 1%) while matching it on clean data (Table A10)?

1 comment
AuthorsOpenReview ↗

Response to vrbz: paired gains, feedback controls, and OCC ablation

Thank you for the careful review. We address below the points that can be clarified from the submitted equations and tables. The recalculation is arithmetic on the reported results, not a new experiment.

Q1: Gain denominator and checkpoint provenance. The reported 24.85% is the mean cellwise full-horizon MSE reduction using the Backbone entries in Table 1. Recomputing the same statistic with each STEPS result and its paired Backbone in Table A14 gives 25.05%; all 168 reported paired entries improve over their corresponding frozen forecasts. Thus, the improvement over the forecast being corrected is not explained by choosing the external baseline denominator. This does not establish matched-checkpoint superiority over the other TTA methods. The differing Backbone entries mean that the blanket provenance statement is too strong: the paired-backbone result and the comparison against published entries must be distinguished. Numerical agreement of any subset would not itself prove checkpoint identity.

Q2: Is observing half the horizon necessary? It is the main evaluation choice, not a requirement of the operator. Table A12 already evaluates shorter prefixes at H=192 across six DLinear datasets. With a=24, full-horizon MSE improves by 14.48%; with a=48, by 20.22%. On the common final 48 target positions, the respective improvements are 9.84% and 16.28%. The common-tail comparison holds the scored positions fixed. This supports useful revision with shorter prefixes, while remaining distinct from the requested information-matched comparison against other methods. In particular, a=48 and COSA-F's B=48 cannot be treated as equivalent merely because their numerical values coincide.

Q3: What is established beyond receiving newer observations? Table A10 compares ways of using the same prefix: ordered ST-TEP achieves 50.7% near-suffix improvement versus 27.9% after temporal permutation; last-residual continuation achieves 1.0% full-suffix improvement versus 37.1% for ST-TEP. These controls show that the structure and use of the feedback matter. They are not a refreshed-lookback baseline, which is not reported in the submission. That control must re-run the same frozen checkpoint at the revision time and score exactly the original unobserved target positions; we do not infer its outcome from these controls.

Q4: Clean versus noisy fusion, and the role of OCC. Table A10 and Table 2 are different protocols: the former uses H=192, a=96 and two backbones, whereas the latter uses DLinear, H=720, a=320 and a 400-step scored suffix. Consequently, Table A10 is not Table 2's zero-corruption reference, and the two additive outcomes do not constitute an internal numerical contradiction.

It is also important to separate removing OCC from removing the global branch. Removing OCC while retaining both branches gives Additive; local-only ST-TEP removes the global contribution as well. Within Table 2, OCC lowers MSE relative to Additive by 13.85%, 13.14%, 12.13%, 10.50% and 9.24% at 1%, 5%, 10%, 20% and 30% corruption. At the endpoints, Additive/ST-TEP/OCC are 0.4160/0.3601/0.3584 and 0.7644/0.7066/0.6938. Full STEPS therefore also improves over local-only at every tested level. This is consistent with the intended mechanism: remove the global component already represented by the local response, retaining complementary information without applying the shared direction twice. A strong local branch on clean H=192 diagnostics does not imply that overlap removal is redundant in the H=720 corruption setting.

The missing same-protocol 0% condition and component-specific paired significance test remain separate evidentiary requirements. Here, “outperforms” refers to the reported mean MSE; the abstract's “significantly” should not be interpreted as a statistical result established by Table 2. We hope these clarifications resolve the denominator interpretation and apparent cross-table contradiction while making the remaining controlled-comparison questions explicit.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Supplementary six-dataset experiment on OCC, tuned addition, and clean versus noisy feedback

We thank vrbz (Q4), 3vu6 (Q3), and zPBK (Q2) for asking whether the fusion gain survives a clean control and a separately tuned additive baseline. We ran a new local DLinear fit across all six datasets at H = 720. Each revision method used the same frozen forecast and 320-step observed prefix. We scored the next 400 steps at all 42,582 test origins. At each nonzero corruption rate we averaged over five shared draws of Gaussian prefix noise (standard deviation six times each channel’s training residual RMS), then weighted the datasets equally. These supplementary results are reported separately from Table 2. The original run metadata for that table remain unverified.

How do the effects vary by dataset? The A10 column gives the paper’s ST-TEP gain over the frozen forecast on the first 24 suffix steps (clean prefix, H = 192, a = 96, DLinear and PatchTST). The six corruption columns give the MSE reduction (%) from adding the OCC projection to matched-weight Direct over all 400 suffix steps (DLinear, H = 720). Positive values favor OCC. The last row uses the ratio of the six-dataset mean MSEs for each corruption level.

DatasetA10 near (%)0%1%5%10%20%30%
ETTh138.8+0.010.00−0.02−0.04−0.06−0.07
ETTh245.1+1.43+1.38+1.18+1.02+0.86+0.74
ETTm124.3+0.34+0.49+1.01+1.51+2.23+2.69
ETTm252.0+0.32+0.57+1.24+1.78+2.21+2.35
Exchange90.1−0.39−0.28−0.03+0.07+0.13+0.13
Weather53.8−0.35+0.06+1.37+2.52+3.93+4.66
Six-dataset mean50.7+0.16+0.27+0.58+0.82+1.03+1.09

At 30% corruption, Weather, ETTm1 and ETTm2 show the largest reductions. ETTh1 slightly favors direct addition.

Mean suffix MSE (lower is better). Direct (matched) uses the weight selected for OCC, so removing the projection is the only change. Direct (tuned) and OCC each select their own weight on clean validation data from the same grid and keep it fixed across noise levels.

Prefix corruptionST-TEP aloneDirect (matched)Direct (tuned)OCC
0%0.364280.365920.361660.36535
1%0.375940.377590.373330.37658
5%0.420490.422130.417920.41968
10%0.474070.475710.471390.47183
20%0.581520.583190.578820.57718
30%0.688600.690240.685860.68271

Does removing OCC hurt? At a matched weight, removing the component of the global correction parallel to the local correction lowered mean MSE at every noise level. The clean difference was small (OCC minus matched Direct −0.00057, 95% interval [−0.00249, +0.00128]). At 5%, 10%, 20%, and 30% corruption, the paired intervals excluded zero. At 30% the difference reached −0.00753 [−0.00934, −0.00574]. This isolates the value of avoiding a repeated correction along the local direction. It also helps explain why clean results at H = 192 in Table A10 need not predict noisy results at H = 720.

Can scalar tuning recover the gain? Independently tuned Direct performed better on clean feedback. OCC had lower mean MSE at 20% and 30% corruption. At 20% the interval against tuned Direct included zero. At 30% the difference was −0.00314 [−0.00559, −0.00068]. The original frozen forecast scored 0.67061 at every rate and outperformed both at 30%. Thus OCC reduced the cost of using heavily corrupted feedback, although preserving the issued forecast did better at that level.

When does the global branch add to ST-TEP? With clean feedback, local-only had lower MSE over the first 100 suffix steps (OCC minus local +0.00344 [+0.00167, +0.00537]). Over the last 100 steps the difference was −0.00033 [−0.00539, +0.00469]. The interval for the far-minus-near contrast included zero. The clearest full-suffix benefit over local-only was under 20% and 30% corruption (differences −0.00434 and −0.00588, both intervals excluding zero). At 30%, OCC beat both local-only and tuned Direct on three of the six individual datasets. We therefore qualify the claim by feedback quality and distinguish the six-dataset mean from consistency across datasets.

The intervals are descriptive paired 95% circular block-bootstrap intervals (720-step blocks, 2,000 resamples), conditional on the observed checkpoints, datasets, and corruption draws. They assess component comparisons in this DLinear experiment. Significance against external TTA methods, the requested ablation over all 168 configurations, and effects of changing total horizon remain open.

AuthorsOpenReview ↗

Author clarification on the records behind Tables 1, 2, A14, and A20

We followed up on the provenance questions from vrbz (Q1, Q4) and zPBK (Q6), which also affect the interpretation of the ablation discussed with 3vu6. Thank you for flagging the discrepancies. We correct our earlier attribution of the submitted Table 2 values to a verified five-seed, 320-step-prefix run.

Table 2. Section 4.4 specifies a 320-step prefix and five corruption seeds. We recovered dataset-level records labeled with a 360-step prefix and a summarization script that selects seed 0 for the internal methods. Averaging the seed-0 MSEs equally across the six datasets reproduces all 15 printed entries for ST-TEP, Additive, and OCC across the five noise levels to four decimals. At 30% corruption the aggregation can be checked directly:

SourceST-TEPAdditiveOCC
Submitted Table 20.70660.76440.6938
Archived rows labeled a=360, seed 00.70660.76440.6938
Same archive, seeds 0–4 averaged0.69500.75140.6825

The four external methods in the recovered archive have records for seed 0 only. Their six-dataset means reproduce none of the 20 corresponding printed entries to four decimals. For example, PETSA at 30% gives 0.6488 in that archive and 0.9504 in Table 2. We have not recovered a final run manifest tying all seven rows to the caption. The numerical match identifies a specific conflict but cannot by itself determine which metadata accompanied the final runs. Consequently, the original table does not establish the claimed matched five-seed comparison against external methods. Our separate supplementary experiment reports the 320/400 protocol with shared corruption draws and controls for the fusion weight; it evaluates the internal components only.

Table 1 versus A14. For the reviewer's PatchTST/Weather/96 example, 0.1742 is the published reference value in Table 2 of the COSA paper. The retained STEPS result records full-horizon MSEs of 0.19759644 for its own frozen forecast and 0.15233167 after revision, reproducing A14's 0.1976 and 0.1523. These records explain the two denominators. They support the within-run improvement for this example. The identity of checkpoints and forecast origins across methods is unverified.

An additional A14 consistency check. If the prefix forecast is unchanged and both methods are scored on the same windows with the same weights, the full-horizon MSE reduction equals the suffix MSE reduction multiplied by (H−a)/H. For MICN/ETTh1 at H=192 and a=96, the printed full reduction is 0.5954−0.5072 = 0.08820. The suffix columns imply (0.6616−0.4691)/2 = 0.09625. Across A14, 23 of the 24 MICN rows exceed the maximum discrepancy attributable to four-decimal rounding. All 144 rows for the other six backbones pass this check. We cannot identify which MICN columns or scoring windows produced the discrepancy from the summaries alone. The previously quoted 25.05% remains an arithmetic summary of the printed full-horizon columns; the MICN records require reconciliation before they can support a verified 168-cell paired claim.

Table A20. For iTransformer/Exchange/96, the retained result records a full-horizon MSE of 0.04712723 and identifies backbone training seed 0. This explains where a value rounding to 0.0471 can be found. We have not recovered the five individual TTA scores, the associated randomized operation, or the aggregation record needed to verify the reported standard deviation of 0.0708. We therefore cannot substantiate that standard deviation or use A20 as evidence of across-run stability. The within-series bootstrap in A19 measures a different source of uncertainty.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.