acceptodds
Under review as a conference paper at ICLR 2027

SWiT: Spectral Wiener Correction for Test-Time Adaptation in Time Series Forecasting

Abstract

Time-series forecasting models degrade after deployment because the statistical properties of the data shift over time in real-world environments. Test-time adaptation (TTA) keeps the deployed backbone frozen and updates a small adapter from the ground truth that arrives after each forecast. Because that ground truth arrives one step at a time, the freshest windows are only partially observed when adaptation runs, so an adapter must learn from incomplete labels without admitting any unobserved value. Existing methods answer this with modules updated by stochastic gradient descent, so incomplete labels can be used only through a gradient step; the correction depends on the learning rate and the number of steps taken, and the cost grows quadratically with the horizon. We propose Spectral Wiener correction for Test-time adaptation (simply, SWiT), which corrects a frozen backbone's forecast in the frequency domain and obtains that correction in closed form rather than by gradient descent. The prediction and its input history are transformed by the Fast Fourier Transform; every bin receives an affine correction (i.e., a scale, an input term, and a rotational bias phase-locked to absolute time), and the result is transformed back. By Parseval's theorem, the time-domain mean squared error (MSE) separates into independent per-bin quadratics, so the per-batch optimum is a closed-form Wiener-type solution over running sums, with no optimizer state or learning rate, at cost in the forecasting horizon . A causal protocol splices only observed steps, the prefix of each in-progress window, into those same running sums, so the freshest evidence enters the closed-form solve itself, and an evidence gate scales each bin's correction by the benefit it has delivered on past windows. Across 216 settings spanning six backbones, nine benchmarks, and four horizons, SWiT attains the lowest average MSE, below the frozen backbone and below the state-of-the-art TTA methods, with fewer parameters and the shortest adaptation time. The code is available at https://anonymous.4open.science/r/SWiT-808C.

Primary area:
learning on time series and dynamical systems
Keywords:
Test-Time Adaptation, Time-Series Forecasting, Non-stationarity

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_pybzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SWiT, a lightweight test-time adapter that corrects the predictions of a frozen time-series forecasting model in the frequency domain. Each frequency bin is corrected using a scaled prediction, a contribution from the input history, and a bias aligned with absolute time. The correction coefficients are computed in closed form from accumulated statistics, avoiding gradient-based optimization. The method also incorporates partially observed ground truth before the full forecast horizon has elapsed, while an evidence gate controls how much correction to apply based on its effectiveness on past forecasts. Experiments across six backbones and nine datasets show improved average forecasting accuracy with fewer adapter parameters and lower adaptation cost than the compared TTA methods.

Critical strengths
  1. Reduces parameter count and adaptation cost while improving forecasting accuracy through frequency-wise correction and closed-form updates.
  2. Evaluated across multiple backbones and datasets, with comparisons against time-domain and dense correctors that support the design choices.
Critical weaknesses
  1. The conceptual novelty appears limited, mainly combining frequency-wise linear correction with least-squares estimation. The new insight beyond an efficient adaptation procedure is not sufficiently clear.
  2. The conditions under which independent per-frequency correction is sufficient remain underexplored. Dense-corrector comparisons are provided, but the limits under changing frequencies, phase shifts, and cross-frequency dependencies are unclear. Benefits on strong foundation models are also untested.
Questions for authors
  1. Under what conditions can forecast errors be adequately corrected by independent linear maps per frequency?
  • Controlled experiments varying frequency, phase, and cross-frequency dependencies would help establish when this restriction works and when it fails.
  1. Do the gains mainly come from correcting persistent backbone errors or from adapting to distribution shifts?
  • Separating these effects and testing stronger backbones, including foundation models, would help clarify the scope of the contribution.
3 comments

W1 - Novelty beyond Efficiency

SWiT rests on three findings, now stated more clearly in Section 1. Other reviews note the same points as the reviewer's second strength: the diagonal frequency-domain corrector beats its time-domain counterpart and ties a dense one (ugA6), and the spectral formulation connects representation, optimization, and efficiency (yMsM).

  1. The error after deployment is almost diagonal in the Fourier basis. Existing TS-TTA methods correct forecasts with dense time-domain modules, and frequency-domain forecasters fix their spectral structure during training (Sec. 2). SWiT corrects a frozen backbone in each frequency bin at test time. A diagonal corrector with O(H) parameters does not differ from a dense one with O(H²) parameters (p = 0.13); it is lower on 41 of 72 cells and on average (diagonal 0.3484, dense 0.3505), and significantly better than a time-domain diagonal corrector (p = 9.6 × 10⁻⁷, Appendix E.3).
  2. The basis gives much of the accuracy. One gradient step on the per-bin objective (SWiT-GD 0.3507) is already almost equal to COSA (0.3522); the closed form adds more gain (0.3341, Table 14).
  3. Delayed partial labels go into the closed form. Previous methods use partially observed ground truth (POGT) only through a gradient step (Sec. 2.3). SWiT puts POGT directly into its statistics, and the evidence gate scores each correction out of sample. Together they improve the MSE by 5.16% (Table 3).

New findings from the rebuttal experiments.

  1. Part of the post-deployment error is visible neither on validation nor at one point of the stream. A correction fitted on validation or on the first 20% of the test stream gets 6.1% or 3.9%, and SWiT gets 10.8% (General Response, Section 1).
  2. The diagonal structure has a measurable limit. SWiT removes 55% to 66% of an added error inside one bin but only 12% to 24% across bins (W2, Q1).
  3. The per-bin closed form extends to cross-channel coupling, because Parseval's separation holds for any number of channels (MV-CF 0.3314 vs. SWiT 0.3341, Appendix L).

W2, Q1 - When Per-Frequency Correction Suffices

We added controlled changes to the ETTh1 and ETTh2 test splits at two strengths (3 backbones × 4 horizons, seed 0): harmonic (H) and weekly (L) amplitude, phase drift (P), a one-time phase step (PS), a frequency chirp (F), and cross-frequency modulation (C).

No TTASWiTRemoved share (%)
Original0.44070.4069-
H (g = 1.5)0.44580.409061
L (g = 1.5)0.44380.409617
P (dphi = π/4)0.50460.462813
P (dphi = π/2)0.54150.483624
PS (dphi = π/4)0.44520.409055
PS (dphi = π/2)0.45550.412066
F (f_end = 1/22)0.52600.474421
F (f_end = 1/20)0.56490.516412
C (m = 0.5)0.44760.412914
C (m = 1.0)0.47720.435622

(Removed share: the fraction of the added No TTA error that SWiT removes. The g = 0.75 rows add no error; Appendix K.2 gives all rows.)

SWiT beats No TTA in all 24 settings.

  • Changes in one bin. For amplitude changes and phase steps, SWiT removes 55% to 66% of the added error. The exception, L at g = 1.5 (17%), has its weekly period on one bin only at H = 336.
  • Changes between bins. For frequency drift and coupling (P, F, C), SWiT still decreases the MSE by 7.7% to 10.7% but removes only 12% to 24% of the added error. This is the limit of independent per-bin correction.

On the real benchmarks, a dense frequency-domain corrector is not better than the diagonal one (Appendix E.3), so the error there comes mostly from changes in one bin.

W2, Q2 - Source of the Gains and Foundation Models

Source of the gains. We froze the corrector in two ways (72 cells, 5 seeds; General Response, Section 1): SWiT-val fits on validation, and SWiT-prefix adapts on the first 20% of the test stream. They decrease the MSE from No TTA by 6.1% and 3.9%, and SWiT by 10.8%. Persistent error thus gives more than half of the gain; SWiT-prefix is worse than SWiT-val because it uses fewer windows; and continued adaptation adds a consistent gain (SWiT beats SWiT-prefix on 71 of 72 cells, p = 5.3 × 10⁻¹³).

Foundation models. We applied SWiT without training (context 96, 5 datasets, 20 cells for each model, seed 0).

ModelNo TTATAFASPETSACOSASWiT
Chronos-Bolt-Base0.64380.46560.47400.47300.4183
Chronos-20.43690.40680.39800.40860.3630
Moirai-1.1-R-Base0.73970.55870.55430.56840.4450

SWiT decreases the MSE by 35.0%, 16.9%, and 39.8%; the existing methods get 26.4% to 27.7%, 6.5% to 8.9%, and 23.2% to 25.1%. SWiT has the lowest MSE in 54 of 60 cells and the lowest MAE on all models (Appendix M).

Thank you for the substantial additional experiments. The controlled frequency and phase perturbations clarify the limits of independent per-bin correction, while the frozen-corrector comparisons and foundation-model results substantially address my questions about the source and scope of the gains.

Some limitations remain: the diagonal–dense comparison does not establish that forecast errors are inherently diagonal, and accumulated statistics can slow adaptation after abrupt shifts. Nevertheless, the additional evidence strengthens the contribution sufficiently for me to support raising my rating to weak acceptance.

We sincerely thank the reviewer for the positive assessment and for supporting a higher rating. We will reflect both limitations in the revision.

Diagonal versus dense correction. As the reviewer notes, the comparison shows that diagonal correction suffices on the evaluated benchmarks, and it does not establish that forecast errors are inherently diagonal. The perturbations also show its limit when energy moves across bins (Appendix K.2). We will state the diagonal structure as an empirical finding and list shifts across bins as a limitation.

Adaptation after abrupt shifts. We agree. Forgetting did not help on the benchmark streams (Appendix E.7), but after a regime switch it lowers the MSE from 0.4036 to 0.3852 (Appendix K.3). We provide it as an option that keeps the closed form, and we will state this trade-off as a limitation.

We thank the reviewer again for the time and effort devoted to our work and for the constructive discussion, which has substantially improved the paper.

Reviewer_ugA6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies test-time adaptation (TTA) for time-series forecasting, in which the backbone stays frozen and ground truth arrives one step at a time after each forecast. SWiT applies the real FFT to the backbone forecast and to the preceding input history. It then corrects each frequency bin with an affine map, with a real scale on the forecast, a real coefficient on the history, and a complex bias that rotates with absolute time. By Parseval's theorem the error splits into per-bin quadratics, so the coefficients are obtained in closed form from running sums, with no gradient descent. A splicing rule adds partially observed windows (POGT) to the solve, and an evidence gate shrinks each bin's correction toward the backbone, using out-of-sample statistics from past windows. The method is evaluated on different settings composed of six backbones, nine benchmarks, four horizons, five backbone seeds, against TAFAS, PETSA, COSA and a one-step Adam variant (SWiT-GD). It reports an average MSE 10.3% below the frozen backbone and 5.3% below COSA, with 674 parameters per channel and the shortest adaptation time.

Critical strengths
  1. SWiT has the lowest average MSE on all six backbones and at every horizon (Tables 1 and 17), and it improves on the frozen backbone on every dataset (Table 16). The results carry mean ± std over five backbone seeds (Tables 18 and 19). The margin over COSA (0.3366 vs 0.3556) is larger than the seed spread in most rows. This is the main reason the paper deserves serious consideration.

  2. The method is very cheap, it needs no backward pass, no optimizer state and no learning rate. Furthermore, it uses only 674 parameters per channel against about 151K for TAFAS and 173K for COSA, and adapts in 25.67 ms per batch against 98 to 294 ms (Table 2). The basis ablation (Appendix E.3, Table 12) shows that the diagonal frequency-domain corrector beats its time-domain counterpart and ties a dense corrector.

Critical weaknesses
  1. The central claim, that SWiT adapts to distribution drift at test time, is not well supported. The evidence fits well with SWiT being a static output calibration fitted on the test stream, because the running sums keep every matured window with no decay or sliding window (Sec. 3.5, "no forgetting rule"). The coefficients therefore converge to one least-squares fit over the whole stream. Such a fit corrects a fixed train/test gap well but follows drift only slowly, and drift is the problem the paper sets out to solve (Sec. 1). The experiments to verify this would be a correction fitted once on the validation split and then frozen, and a variant that forgets old windows. The paper's own long-horizon results also point toward that calibration, for example, at H = 720 on Exchange, in Table 11 gives exactly the same value (0.8540, 0.8873, 0.8368), so the matured-sum correction never takes a large effect there. The large gain in the main table (0.7382 vs 0.8842, Table 18) must then come from the partial-label path, which mostly corrects the level. Without these controls, the comparison with SGD-based adapters, which do track recent data, does not show what the paper claims.

  2. The gains depend on several hand-set choices and on a comparison that may not be completely fair, for example, the size of the improvement depends strongly on the clipping ranges (Table 22). The paper does not say whether the baselines get the same history, so part of the margin may come from extra information rather than from the method. On Solar, SWiT is worse than the frozen backbone in MAE for most backbones (for example iTransformer 0.303 vs 0.289, FreTS 0.312 vs 0.302, MICN 0.284 vs 0.276; Table 20a). This suggests that part of the MSE gain comes from shrinking large errors, not from better forecasts overall. Together, these points make it hard to tell how much of the reported margin comes from the method itself.

Questions for authors
  1. What MSE does the same spectral corrector reach if it is fitted once on the validation split (or on the first part of the test stream) and then frozen? And how does SWiT change with a forgetting factor on the running sums? This question would answer whether the gains come from test-time adaptation or from static calibration.

  2. The input history has length H (up to 720), while the backbone sees only W = 96 steps. The uX term therefore uses up to 624 past steps the backbone never sees. Do TAFAS, PETSA and COSA get equal access to that longer history? If not, how much of the margin is left when SWiT's history is limited to the last W steps?

  3. Why does SWiT get worse than the frozen backbone in MAE on Solar for most backbones (Table 20a)? Is this a heavy-tail or zero-inflation effect, and does a robust or MAE-matched objective remove it?

  4. On Exchange at H = 720, FD diag (SWiT), FD dense, TD diag and TD dense in Table 11 all give exactly the same MSE as the frozen backbone (0.8540 for iTransformer, 0.8873 for DLinear, 0.8368 for FreTS), so no correction is applied. With about 798 test windows and labels that complete only 720 steps later, almost no window is fully observed before the stream ends. Is this the reason? How many fully observed windows does each test stream have at H = 720, especially on the shorter datasets (Exchange, ETTh1, ETTh2)?

  5. Are the TAFAS, PETSA and COSA gains in this setup in line with what those papers originally report? Was any baseline learning rate re-tuned for this protocol?

2 comments

W1, Q1 - Static Calibration and Forgetting

We ran both experiments the reviewer suggested, plus two direct tests of the static-calibration hypothesis (tables in Sections 1 and 2 of the General Response).

The reviewer is correct on one point. On the evaluated benchmarks, much of the gain comes from correcting persistent backbone error: SWiT-val, fitted on validation and frozen, decreases the MSE by 6.1% from No TTA, more than half of SWiT's 10.8% (last 80% of the test stream, 72 cells, 5 seeds).

Static calibration does not explain the full gain.

  • A correction frozen at one point is not sufficient. SWiT-prefix adapts like SWiT on the first 20% of the test stream and then freezes. SWiT beats it on 71 of 72 cells (p = 5.3 × 10⁻¹³) and SWiT-val on 59 (p = 1.9 × 10⁻⁶). On Exchange, SWiT-val is worse than No TTA (0.3882 vs. 0.3788).
  • The running sums follow a new regime. In a stream joining the test splits of two datasets (48 cells, 5 seeds), SWiT beats SWiT-val after the switch on all 48 cells (0.4036 vs. 0.4639, p = 1.6 × 10⁻⁹) and stays better than TAFAS, PETSA, and COSA (0.4226, 0.4328, 0.4185; p ≤ 4.9 × 10⁻⁵).
  • Forgetting helps only after a switch. A sliding window and an exponential decay with a memory of 2000 windows (SW 2000, ED 2000) give a lower MSE than SWiT after the switch (0.3852 and 0.3879), but no gain before it (SW 2000 p = 0.25, ED 2000 significantly worse), and they are significantly worse at all memory lengths on the benchmark streams (108 cells, p < 0.01).

Thus SWiT estimates the backbone error online from the test stream. After an abrupt switch it converges slowly, and forgetting, given as an option, makes it faster (Appendix E.6, Appendix E.7, Appendix K.3).

W1 - Exchange at H = 720 and the Partial-Label Path

The reviewer is correct for this setting. Without POGT (Table 11), corrections start only at t = 2H (Q4), and Exchange at H = 720 has fewer test windows than 2H, so all correctors equal the frozen backbone and the gain in the main table comes from POGT. This does not hold in general: without POGT, SWiT keeps about 70% of its gain (0.3690 → 0.3440, Table 3), and the scale term alone decreases the MSE by 5.6% (Appendix E.1, Appendix E.3).

W2 - Sensitivity to Hand-Set Choices

The hyperparameters were selected once on the validation splits of ETTh1 and ETTm2 and applied to all 216 settings without further tuning (Appendix H.1); the closed-form update has no learning rate or step count.

The gain changes with the clipping range, but SWiT stays better than the strongest baseline over a wide range: COSA gets 0.3522 in the same ablation setting (Table 14), and SWiT is lower at all tested values of δu, α_u, and α_c (Table 32 and Table 33). It is worse than COSA only when δs is too narrow (1/4) or removed (∞), and even then better than TAFAS and PETSA. Clipping removes the few bins where outliers inflate the closed-form solution; without it, the loss falls mostly on ETTh2 and Traffic (Figure 4). The exact range therefore matters little once these bins are removed.

W2, Q2 - Input History Length

TAFAS and PETSA use no data before the lookback window; COSA, like SWiT, does, through a context vector that summarizes past ground truth. We limited SWiT's input history to the backbone's lookback of W = 96 steps (SWiT(W); 108 cells, seed 0).

No TTATAFASPETSACOSASWiT(W)SWiT
Avg. MSE0.36900.36070.35910.35220.33800.3341

SWiT(W) decreases the MSE by 8.41% from No TTA, about 89% of full SWiT's 9.45% and much more than COSA (4.57%). At H = 96 the history equals the lookback, so the two variants coincide. A longer history adds a significant gain (p = 1.3 × 10⁻⁷), but most of the margin over the baselines comes from the method (Appendix E.8).

W2, Q3 - MAE on Solar

No TTATAFASPETSACOSASWiT
MAE0.32230.32090.32070.32450.3302
Night MAE0.05600.06250.06630.06880.1089
Day MAE0.53500.52720.52390.52870.5069
MAE, night-zero fill0.29750.29310.29130.29390.2819

(12 cells, seed 0; night and day MAE before post-processing.)

The cause is zero-inflation. Solar power is exactly zero at night, but the SWiT correction spans the whole horizon in frequency, so correcting the day error adds a small offset at night. The offset barely changes the MSE but adds linearly to the MAE: SWiT has a higher MAE than No TTA at night and a lower one during the day.

A robust or MAE objective does not remove it. Replacing only the (s, u) objective with a Huber loss (c = 1, 2, 4) or an L1 loss (exact IRLS) leaves the MAE at the SWiT level (Huber 0.3285 to 0.3302, L1 0.3290) and the night MAE at 0.1054 to 0.1088. These objectives act per frequency bin, so they cannot remove an error that occurs only at night in the time domain.

A night-zero fill confirms the cause. We define night as the times when at least 99% of the training values are zero (18:30 to 05:10) and set every method's forecasts there to zero. SWiT then has the lowest MAE, 5.25% below No TTA, and its MSE gain rises from 12.49% to 14.21%. Over all datasets, SWiT has the lowest average MAE for all six backbones and horizons (Table 28a, Table 20a in the submission; Appendix G.3).

Q4 - Fully Observed Windows at H = 720

Yes. With POGT disabled, the first coefficients appear at t = H, and the evidence gate needs H more steps to score them, so corrections start at t = 2H.

Dataset (H = 720)Test windowsMatured windowsWindows with a correction
Exchange798780
ETTh12,7652,0451,325
ETTh22,7652,0451,325

Q5 - Baseline Configuration

All methods use the same checkpoints (6 backbones × 5 seeds), splits, and test streams, and TAFAS, PETSA, and COSA run from their public code. The public TAFAS and PETSA code sets a learning rate per setting; for a fair comparison, both use one setting everywhere (learning rate 10⁻³, weight decay 10⁻⁴, gate initialization 0.01; Appendix F.3).

MethodShared cellsOriginal paper, No TTA → methodOurs (public code, 5 seeds), No TTA → method
TAFAS720.3663 → 0.3481 (4.99%)0.3717 → 0.3610 (2.89%)
PETSA12 (ETTh1)0.5595 → 0.5435 (2.86%)0.5625 → 0.5443 (3.24%)
COSA720.3704 → 0.3100 (16.31%)0.3717 → 0.3478 (6.42%)

Our No TTA values are within 0.4% to 1.5% of the original papers. The gains of TAFAS and PETSA are near the reported gains; the gain of COSA is smaller. All values in the last column were measured with the public code of each method on the same checkpoints, splits, and test streams (COSA: batch of 48, 3 steps, context of 10; TAFAS and PETSA: the single learning-rate setting above).

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

SWiT is a lightweight module for the time series forecasting online adaptation. It calibrates on the frozen backbone prediction to reduce estimation error and counter the distribution shift effects. It ensures the ground-truth label reveals after the calibration being made. SWiT leverage spectral correction and closed-form adaptation based Parseval's theorem. The experiments shows the largest error reduction across six backbones, nine datasets and three baselines in average.

Critical strengths

The paper’s key strength is that its spectral formulation makes accurate adaptation both compact and analytically tractable. Correcting each frequency independently replaces a dense time-domain adapter with high efficiency. The Parseval's theorem turns the objective into small per-frequency problems that yields closed-form solution. These established the connection established among representation, optimization and efficiency. The lightweight design and practical deployment settings allows SWiT potential to be applied on various time series forecasting backbones.

Critical weaknesses
  1. Insufficient empirical evidence for robustness to distribution shift. The paper claims that SWiT counters distribution shift, but the experiments do not establish whether meaningful shifts occur in the evaluated datasets or quantify their impact. Training and testing on splits of the same dataset provides limited evidence for this claim. Cross-dataset evaluation, such as training on ETTh1 and testing on ETTh2, accompanied by an explicit characterization of the shift, would provide stronger support. Otherwise, the paper should weaken its robustness claim on counter distribution shift.

  2. Limited theoretical justification and potentially diminishing adaptability. The Wiener interpretation establishes an analytical fitting rule, but it does not demonstrate robustness to nonstationarity or quantify how effectively SWiT tracks changing distributions. Moreover, accumulating historical statistics without decay or a reset mechanism may allow earlier regimes to dominate, making updates increasingly unresponsive to recent changes. The paper therefore needs theoretical analysis or targeted experiments showing whether adaptation remains effective after prolonged operation and subsequent regime shifts.

Questions for authors
  • How does the method perform when ground-truth feedback arrives with delays longer than 1 time step, which is a more realistic consideration since the ground-truth feedback may not be obtained immediately after the prediction.

  • Does the method also improve recent forecasting backbones, including foundation models? To what extent do the reported gains reflect limitations of the chosen backbones and their greater room for improvement?

  • As mentioned in weakness, if the paper attempts to verify its robustness on distribution shift, what about its performance with backbone train on ETTh1 and test on ETTh2 as a typical covariate shift case.

  • What prevents historically beneficial corrections from becoming harmful after the distribution changes?

  • Under partial label availability, what assumptions justify the m/Hm/H scaling and cancellation of spectral leakage across temporally correlated windows?

6 comments

W1, Q3 - Cross Dataset

Each backbone is trained on one dataset and tested on another, in both directions of ETTh1 ↔ ETTh2 and ETTm1 ↔ ETTm2 (3 backbones × 4 directions × 4 horizons = 48 settings, 5 seeds).

No TTATAFASPETSACOSASWiT
Avg. MSE0.55350.45870.46170.45020.4067

The shift is large. For iTransformer trained on ETTm2 and tested on ETTm1, the No TTA error is 2.64, 3.04, 2.72, and 2.82 times the in-domain error at H = 96, 192, 336, and 720 (at H = 96, 0.3966 → 1.0462).

SWiT adapts best. It decreases the MSE by 26.5% from No TTA (COSA 18.7%) and is lowest in 44 of 48 settings. The gain grows with the shift: 23.3% for ETTh2 → ETTh1, 38.7% for ETTm2 → ETTm1, and 12.4% for ETTm1 → ETTm2. On ETTh1 → ETTh2, SWiT gets 11.6% (0.3258 → 0.2880) and COSA 2.9% (0.3164).

A shift also occurs within one dataset. The mean backbone residual shifts much more on Exchange (0.550) and ETTh2 (0.096) than on Weather (0.005). On these two, a correction frozen after validation (SWiT-val) is clearly worse than SWiT on the full test stream (ETTh2 0.3097 vs. 0.2924, Exchange 0.3735 vs. 0.3233; Appendix E.6).

About the claim. SWiT estimates the post-deployment backbone error online. This error has a persistent part and a part that changes along the stream. On the standard benchmarks the persistent part gives more than half of the gain (Appendix E.6); under the larger cross-dataset shift the gain rises to 26.5%. The abstract and Section 1 now state this, and Appendix K.1 gives the full results.

W2 - Forgetting

Section 2 of the General Response gives the tables.

Benchmark streams (108 cells, seed 0). With a sliding window (SW) or an exponential decay (ED) of 500, 1000, or 2000 windows in the running sums, all six variants are significantly worse than SWiT (p < 0.01): forgetting only removes samples.

What the Wiener solution guarantees. The per-bin coefficients are the least-squares (Wiener) solution over all accumulated windows. This minimizes the error over those windows but bounds no convergence time: a new regime's weight grows only with its share of windows, so a late switch moves the coefficients slowly. We therefore measured the convergence directly.

Regime switch after a long run (48 cells, 5 seeds). The test splits of two datasets are joined into one stream.

  • Right after the switch, SWiT has the lowest mean MSE, but its difference from No TTA is not significant (p = 0.52).
  • Later, SW 2000 and ED 2000 in the running sums are better than SWiT on 40 and 42 of 48 cells: as the reviewer expected, the old statistics slow the convergence. Before the switch they give no gain (SW 2000 p = 0.25; ED 2000 significantly worse, p = 0.031).
  • SWiT still adapts: it beats the frozen SWiT-val on all 48 cells (0.4036 vs. 0.4639) and, after the switch, all existing TTA methods (TAFAS 0.4226, PETSA 0.4328, COSA 0.4185, p ≤ 4.9 × 10⁻⁵).
  • The evidence gate does not detect the switch well. It closes 8.7% of the open bins (3.2% of those with w ≥ 0.5), against 10.6% and 9.9% for SW 2000 and ED 2000. Forgetting in the gate only (SW 2000, ED 2000) closes 29.6% and 23.2% but lowers the MSE by at most 1.0% (0.4006 and 0.3996), so the slow convergence comes from the running sums.

Rate of convergence. We count the windows until the moving-average MSE of SWiT stays below No TTA for 2H windows (median over 5 seeds): at most 120 in 23 of 24 backbone-horizon pairs of ETTh2 → ETTh1 and ETTm2 → ETTm1, and up to about 7,000 for ETTm1 → ETTm2. For ETTh1 → ETTh2 at H = 720, SWiT is above No TTA in the first 2H windows on 14 of 15 runs, and no recovery starts up to window 1,325, the last start that leaves 2H windows; a later recovery cannot be observed. In the last 1,325 windows, SWiT beats No TTA, SW 2000, and ED 2000 on all 15 runs (0.2212 vs. 0.2988 for No TTA).

After a one-time phase shift (PS) within one dataset, SWiT removes 55% to 66% of the added error (Appendix K.2). Forgetting keeps the closed form, so we give it as an option for deployments with regime switches (Appendix E.7, Appendix K.3).

Q1 - Delayed Feedback

The ground truth of an H-step forecast arrives one step at a time, so the protocol already has delays of up to H = 720 steps (Sec. 3.1). SWiT uses only observed ground truth: only matured windows enter the running sums, incomplete windows add only their observed prefix (POGT) through the splicing rule (Sec. 3.6), and steps without new labels use the latest coefficients. An additional delay changes only the number of observations and the update frequency, and needs no new learning rate.

We added a delay of d = 24 and d = 96 steps, with the maturity check and all POGT terms using t − d as the current time (108 cells, seed 0).

No TTASWiT (d = 0)SWiT (d = 24)SWiT (d = 96)
Avg. MSE0.36900.33410.33440.3350

The MSE rises by at most 0.0009, and neither delay differs significantly from d = 0 (p = 0.89 and 0.25). The largest increases, on Exchange (0.3232 → 0.3287) and ETTh2 (0.2896 → 0.2967), stay below No TTA (0.3643 and 0.3231). Section 3.1 and Appendix B give the details.

Q2 - Foundation Model

We applied SWiT without training to three foundation models (context 96, 5 datasets, 4 horizons; 20 cells per model, seed 0).

ModelNo TTATAFASPETSACOSASWiT
Chronos-Bolt-Base0.64380.46560.47400.47300.4183
Chronos-20.43690.40680.39800.40860.3630
Moirai-1.1-R-Base0.73970.55870.55430.56840.4450

SWiT is the best TTA method on all three, decreasing the MSE by 35.0%, 16.9%, and 39.8% (existing methods: 26.4% to 27.7%, 6.5% to 8.9%, and 23.2% to 25.1%). It has the lowest MSE in 54 of 60 cells and the lowest MAE on all models (Appendix M).

How does SWiT address the limits of the backbone? A frozen backbone, even a foundation model, cannot follow a shift after deployment; SWiT corrects its forecasts online without changing it. The gain is larger for weaker backbones: over nine backbones, the Spearman correlation between No TTA MSE and the SWiT gain is 0.95 (p = 0.0004), and 1.00 on the five datasets shared by the trained and foundation models. A large part of the gain thus reflects the backbone's room for improvement.

Q4 - Safeguard

  1. Evidence gate (Sec. 3.4). The gate scores each correction against later ground truth and closes bins that no longer help; on the example channels of Appendix I, 45% to 65% of the opened bins close again. After the abrupt switch of W2, however, it closes only 8.7% of the open bins, and forgetting in the gate lowers the MSE by at most 1.0%, so forgetting in the running sums is the better option for abrupt switches.
  2. Clipping (Sec. 3.5). Limits on s_k and u_k bound the damage; without them, the loss reaches 29% on some datasets (Appendix H.2).
  3. Shrinkage (Sec. 3.5). SWiT halves u_k and c_k. Without α_c, the gain drops from 9.5% to 7.9%; without α_u, to 9.2% (Table 33).

Q5 - Ratio Weight and Spectral Leakage

The zero-filled target is the full target times a mask, which in the frequency domain gives the exact relation

G˘k(j)=mjH Gk(j)+Lk(j),Lk(j)=1H∑k′≠kM^k−k′(j) Gk′(j)\breve{G}_k^{(j)} = \frac{m_j}{H}\, G_k^{(j)} + L_k^{(j)}, \qquad L_k^{(j)} = \frac{1}{H} \sum_{k' \neq k} \hat{M}^{(j)}_{k-k'}\, G_{k'}^{(j)}

The factor m_j/H is the in-bin gain from the DC component of the mask, and L_k^{(j)} is the leakage from other bins. SWiT multiplies the input by the same factor (X̆_k^{(j)} = (m_j/H)·X_k^{(j)}), so all windows give one least-squares problem for u. The only assumption is that L_k^{(j)} does not correlate with X̆_k^{(j)}. This needs only local stationarity within each window, so drift between windows is permitted: different bins are then approximately uncorrelated, and the estimate of u is approximately unbiased. Window overlap changes only the variance (Appendix C).

The rotational-bias path can keep a small bias because the phase cancellation is partial; the decreasing POGT weight (O(H/n)), the shrinkage of c (α_c = 1/2), the separate DC bin (Appendix D), and the evidence gate limit it. POGT splicing improves the MSE by 2.95% (Table 3; Section 3.6 and Appendix C).

I appreciate the authors' additional experiment results, while I found the delayed experiment result rather surprising. My experiments with delayed labels and current inputs given the provided code showed substantially greater degradation. It is very abnormal to see online calibration with longer delay got almost no degradation. Please provide the delay-specific implementation and per-setting results, including confirmation that the delay applies to the coefficient updates, bias correction, evidence gate, and every POGT path. Please also clarify whether the reported average masks larger losses in particular settings and whether startup windows are included.

Reviewer_m7mEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents SWiT as a novel test-time adaptation approach for time series forecasting models. The specific problem is how to efficiently adapt the forecasts of a frozen forecaster online, even when the ground truth of the full prediction horizon is not yet fully revealed. SWiT transforms both the forecast and the last steps of history into the frequency domain with FFT. Then each frequency bin is corrected with an affine correction based on 3 components (a scale based on the forecasts, a weight based on input history and rotational bias based on absolute time). An evidence gate decides how much of the resulting correction to apply in each bin, based on historical usefulness. The parameters necessary for the method can be obtained from an efficient closed form solution, which also includes the partially observed ground truth of the most recent windows. SWiT is validated with extensive empirical testing.

Critical strengths
  1. Paper well written and structured.

  2. The proposed method is simple, well justified and efficient, making it a practically useful proposal for real-world use.

  3. The empirical analysis is complete on multiple datasets and with multiple forecasters. I particularly appreciated the presence of both the MSE and MAE metrics, the ablation study and the computational overhead analysis.

Critical weaknesses
  1. The method heavily relies on running statistics which accumulate constantly at test time. This makes the adaptation not really adaptive in the long run, stopping the possibility of forgetting non-relevant information.

  2. The method is not compared with its non-adaptive version: using SWiT only as a static frequency corrector, fitted once on the validation set, and kept as a frozen component of the overall forecaster.

  3. The correction is channel-independent: multivariate series are treated as independent univariate series, so cross-variable dependencies are not used.

Questions for authors
  1. Can you please explain why you decided to avoid any forgetting mechanism in the running sums?

  2. I think that being restricted to a channel-independent correction is quite limiting. What is stopping SWiT from working with multivariate series using cross-variable dependence?

  3. This is not fundamental, but as you cited both online approaches and continual learning methods in the related work, may I suggest having a look at the recent work of Urettini et al., “Online Continual Learning for Time Series: a Natural Score-driven Approach”, which connects online continual learning and time series forecasting.

1 comment

W1, Q1 - Forgetting

We keep the running sums without forgetting for three reasons. (1) A decay rate or window length is a hyperparameter that changes with the dataset, while SWiT uses one configuration for all datasets. (2) Each per-bin coefficient is estimated from few windows, so more windows decrease the variance. (3) POGT splicing already adds recent information, and the evidence gate scores each correction against later ground truth.

Two experiments test this design (General Response, Section 2):

  • Benchmark streams (108 cells, seed 0). All six variants with forgetting in the running sums are significantly worse than SWiT (p < 0.01). Here forgetting only removes samples.
  • Abrupt regime switch (48 cells, 5 seeds). After the switch, SW 2000 and ED 2000 are better than SWiT (0.3852 and 0.3879, against 0.4036); before it, forgetting gives no gain. SWiT still adapts: it is better than the frozen SWiT-val on all 48 cells and than TAFAS, PETSA, and COSA (p ≤ 4.9 × 10⁻⁵). But the old statistics make its convergence slow, and the evidence gate closes only 8.7% of the open bins. For ETTh1 → ETTh2 at H = 720, SWiT stays above No TTA for the first 2H windows, but in the last 1,325 windows it is better than No TTA, SW 2000, and ED 2000 on all 15 runs.

Thus the reviewer is correct: forgetting helps after an abrupt switch in a long deployment. It keeps the closed form, so we give it as an option (Appendix E.7, Appendix K.3).

W2 - Static Calibration

SWiT-val fits SWiT on validation and freezes it (72 cells, 5 seeds). On the full test stream, it decreases the MSE by 6.47% from No TTA, almost equal to COSA (6.42%). SWiT gets 9.58% and is better than SWiT-val on 50 of 72 cells (p = 4.7 × 10⁻⁴), most at H = 720 (SWiT 0.4744, SWiT-val 0.4997). On the last 80% of the stream, SWiT is better than SWiT-val on 59 cells and SWiT-prefix on 71 cells (General Response, Section 1).

Full test stream:

DatasetResidual shift |test − val|SWiT-valSWiT
ETTh10.0260.51240.5259
ETTh20.0960.30970.2924
ETTm10.0340.43410.4253
ETTm20.0240.21560.2097
Exchange0.5500.37350.3233
Weather0.0050.24060.2400

(The residual shift is the absolute difference of the mean backbone residual between test and validation.)

The difference follows the residual shift. When the shift is small, the two methods are almost equal. On ETTh2 and Exchange, the shift is large and SWiT-val becomes much worse; on Exchange, it is even worse than No TTA (0.3735 vs. 0.3669), while SWiT stays better than No TTA on both. The shift after deployment is not known in advance, so continued adaptation is safer (Appendix E.6).

W3, Q2 - Multivariate

Parseval's separation holds for any number of channels, so the per-channel scale s_k can be replaced by a coupling matrix B_k ∈ ℂ^{V×V} while the closed form and all other parts of SWiT stay unchanged. We tested three estimators of B_k (108 cells, seed 0). Naive MV estimates the diagonal and off-diagonal entries together. MV-CF keeps the diagonal at s_k and adds the off-diagonal entries in closed form. MV-GD uses the MV-CF objective with one Adam step for each block.

No TTASWiTNaive MVMV-CFMV-GD
Avg. MSE0.36900.33410.33610.33140.3363

MV-CF is better than SWiT on 58 of 108 cells (p = 0.039, no correction for multiple comparisons) and is the best method on Solar (137 channels), Electricity (321), and Traffic (862). Naive MV is significantly worse than SWiT (p = 0.0047), and MV-GD is not significantly different (p = 0.34).

The benefit is not uniform: on ETTh1, ETTh2, ETTm2, Exchange, and Weather, MV-CF is almost equal to SWiT or slightly worse. The computation also grows with O(V³); on Traffic, MV-CF keeps four 862 × 862 complex128 Gram matrices for each of 361 bins (approximately 17 GB). Thus we keep the channel-independent design as the default (Appendix L).

Q3 - Additional Related Works

Thank you; the revision cites NatSR, which connects online continual learning with time-series forecasting, in Section 2.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.