acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Relaying Structural State Beyond the Fixed Input Window for Non-stationary Time Series Forecasting

Abstract

The reversible normalization family, from RevIN through SAN, FAN, and DDN, mitigates non-stationarity in real-world time series with progressively more precise restoration statistics, but each method estimates its correction quantities from the current input window alone. When the window is short, observation noise dominates the level and phase of the periodic components, and the forecast origin never uses observations before the window boundary. This work proposes TRACE, an adapter that carries the history preceding the window in a small structural state that is relayed from one forecast origin to the next. A fixed linear state-space filter with a per-channel level and harmonic components updates the state causally and has no learnable parameters; at forecast time the state is rolled out analytically and combined with the backbone forecast through a learned gate. Because the gate starts at zero, TRACE is initialized exactly at the backbone with window-statistic normalization and can learn to close the structural branch. The history reaches the state, but never the backbone, so this state continuity leaves the backbone's input length unchanged and is a design choice separate from window extension. Across four forecasting backbones, seven real-world benchmarks and three seeds, TRACE lowers the overall mean MSE by 10.66% against RevIN and by 6.10% against DDN, the strongest competing adapter, while using only 226-610 learnable parameters, 64-10,129 times fewer than SAN, FAN and DDN. Controlled experiments show that the gain depends on history the backbone never sees and that window-only forecasts combined in the same way do not reproduce it. The code is available at https://anonymous.4open.science/r/TRACE-8346.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, non-stationarity, distribution shift, reversible normalization, linear state-space model, Kalman filter, model-agnostic adapter, long-term forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JKLjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRACE adds a carried structural state to the window-normalization adapter slot (RevIN, SAN, FAN, DDN). A fixed Kalman filter with a level and K undamped periodogram harmonics runs over the whole series, keeping one state per channel. A zero-initialized gate of 226–610 parameters mixes the filter's closed-form rollout with the window-normalized backbone forecast. On 4 backbones x 7 datasets x 4 horizons at L = 96, TRACE lowers MSE by 10.66% against RevIN and 6.10% against DDN. Controlled experiments on 12 cells attribute the gain to history outside the window

Critical strengths
  1. Simple, cheap, and consistent under the standard protocol. At L = 96, TRACE beats RevIN in 105/112 cells and DDN in 97/112, and the resampled intervals exclude zero. It equals the backbone exactly at initialization, and the filter has no trainable parameters. Thus, TRACE is an easy add-on that reliably helps short-window models
  2. Informative ablations. When the model is retrained with the state limited to in-window history, the gain disappears (Tables 3, 19: all 12.2% on DLinear/Weather-96), so it comes from out-of-window history. Table 9 combines window-only forecasts with the backbone in the same way, including a ridge fit given the filter's frequencies, and they do not reproduce the gain. Negative results that sometimes the gate does not perform better than a scalar weight are reported openly. This makes the paper informative even where results are mixed
Critical weaknesses
  1. The headline gain mostly reflects access to more history.

    • TRACE sees the history outside the window, while the baselines see only 96 steps. The gains concentrate where cycles exceed the window: against RevIN, 22.1% on Traffic, 15.6% on ECL, and 12.9% on Weather, versus 3.1–4.1% on ETTh1, ETTh2 and ETTm2
    • Gains concentrate where the dominant cycle is longer than 96 steps. Pooled over ECL, Traffic, and Weather, TRACE is 18.6% below RevIN and 8.6% below DDN, while on the four ETT sets the figures are 5.3% and 4.5% (computations are from Tables 25–28)
    • At L = 336 (Table 4), TRACE is lowest in only 3 out of 16 cells, while DDN beats it in 13/16. Its mean MSE is 3.8% above DDN and 2.4% above RevIN. On the same 16 cells at L = 96, it is 4.2% and 5.1% below them
    • On ETTh1-720 at L = 336, TRACE is 10.8% (DLinear) and 21.1% (iTransformer) worse than RevIN
    • TRACE at L = 96 has higher peak training memory than RevIN at L = 336 in 7/8 configurations of Table 17 (for example, 475 vs 150 MB in DLinear on ECL)
  2. The paper's own ablations undercut two design choices.

    • The learned gate seems to be unnecessary. A single weight fitted on validation is as good as the learned gate: it has lower MSE in 14 of 24 cells, and 1.4% lower on average (Table 8)
    • Components longer than the window hurt. Removing them and using one fixed weight raises the median gain on the factor cells from 1.36% to 8.48%, winning 10 of 12 cells (Tables 3, 9, 10). On ETTh1 at H = 720, this simpler variant reaches 0.431–0.436 MSE, against 0.484–0.528 for TRACE in the same setting (Table 22). The deployed ETTh1 filter still keeps these components, which raise its forecast error from 0.446 to 0.502 (Table 11)
Questions for authors
  1. Lengthening the window from 96 to 336 lowers RevIN's and DDN's mean MSE by 7.2% and 7.6%, but leaves TRACE's unchanged (mean is 0.0%, median cell is −2.7%). RevIN and DDN at L = 336 therefore already match or beat TRACE at L = 96, except on Weather (Table 4). This suggests the L = 96 advantage is mostly the value of history beyond 96 steps, which a longer window also provides, rather than an advantage of carrying a state. Please report every method at its validation-selected L∈{96,336,512,720}L \in \{96, 336, 512, 720\} on at least ECL, Traffic, and Weather, with all four backbones
  2. CycleNet's RCF [1] is a plug-in that supplies periodic structure the window cannot see, which looks like a close prior work that should be discussed and added to comparison
  3. TRACE's learning rate and K are selected on validation from 12 settings per backbone–dataset pair (4 learning rates x K∈{2,4,8}K \in \{2, 4, 8\}, Appendix G, Table 21), while the baselines keep their published settings. Appendix N shows that tuning the learning rate alone improves RevIN by up to 6.25% on the six pairs checked, none of them ECL or Traffic. Please give every adapter the same learning rate search
  4. The paper mentions that iTransformer is run without its built-in normalization (Appendix G). However, this makes it weaker than the published model, which normalizes its input. Without normalization, iTransformer is 40% worse than its published L = 96 results (ETTh2 +119%, ETTm2 +98%, Traffic +33%). Against the published iTransformer (with normalization), TRACE's gain is 4.3% (about 0 on ETTh1/ETTh2), not the ~32% that Table 1 implies. Please report Table 1 with each backbone's own normalization

[1] Lin, S., et al. "CycleNet: Enhancing time series forecasting through modeling periodic patterns." Advances in Neural Information Processing Systems 37 (2024), 106315-106345

6 comments

Thank you for raising the input-length comparison.

W1/Q1. Validation-selected inputs. Across five datasets (ECL, Traffic, Weather, ETTh1, ETTm2), four backbones and H96/H720, selected SAN improved over its own L=96 result in 36/40 cells (mean reduction 11.0%). When TRACE selected from 96, 336, 512 and 720 and SAN from 96, 336 and 720, TRACE had 2.1% higher MSE on average (TRACE lower in 19/40); with three lengths for both methods, it was 2.5% higher across 40 cells (TRACE lower in 17/40). Both comparisons favor SAN on average. In the four-length comparison, TRACE was lower in 8/8 Traffic and 5/8 Weather cells but only 2/8 ECL, 3/8 ETTh1 and 1/8 ETTm2 cells, with 6.5% and 5.7% higher mean MSE on ETTh1 and ETTm2.

For each backbone–dataset pair, we selected L by mean validation rank across both horizons at seed 2021. Both comparisons exclude 512 for SAN because its segment length (24 or 12) must divide L. For SAN and FAN, we then selected the learning rate at the chosen L and evaluated three seeds.

Four-length TRACE uses three-seed results at its selected configuration: the reselected lr at the chosen long input in 32 cells and the main-table lr at L=96 in eight. Three-length TRACE uses three seeds in 32 cells (3.4% higher MSE than SAN, 12/32 wins) and one in eight. Substituting three-seed results from a different L or K in those eight gives 2.4% higher MSE (20/40). Against FAN, which also supports 512, TRACE selected from the same four lengths won 22/40 cells with 2.4% lower MSE on average.

Earlier three-seed controls also depend on protocol. Selecting among four lengths at original learning rates gives TRACE 28/40 wins over RevIN (0.9% mean reduction) and 15/40 over DDN (1.4% higher MSE). Selecting among three long inputs (336, 512, 720) before retuning gives 25/40 and 19/40 wins, with mean reductions of 1.3% and 0.4%.

For excess MSE over SAN, we compute 100 × (TRACE − SAN)/SAN from seed-averaged MSEs in each cell, then average equally over the 40 cells (32 for the three-seed subset). The abstract and introduction will scope the headline to fixed L=96: across 112 cells, reductions in overall mean MSE are 10.66% vs RevIN and 6.10% vs DDN, or 6.55% and 5.71% after baseline lr retuning. TRACE carries state without extending the backbone input. Across 16 H720 settings, its median peak training memory at L=96 is 2.2× RevIN's and 1.3× DDN's at L=720; its median latency ratio vs DDN is 0.52×.

Q4. iTransformer's own normalization. With native normalization enabled, TRACE retains a 4.5% mean MSE reduction over iTransformer and is lower in 25/28 cells (3 seeds). We will include this comparator explicitly in Table 1 and label the unnormalized plain model as an ablation. The native result differs from the submitted RevIN comparison (cellwise 4.3%, 24/28): native normalization has no affine parameters, while RevIN does. The large gap against the unnormalized model will not be presented as TRACE's gain over the published backbone.

W2. Long components and the gate. The submitted factor comparison favors short components plus a validation-fitted scalar in 10/12 cells (median gain 1.36% → 8.48%). ETTh1's retained periods 650/524 also raise standalone error from 0.446 to 0.502. We therefore tested full/short × gate/scalar under main-table training, retaining the learning rates and using three seeds.

Here short removes every period P≥96P\ge96 and reselects noise ratios on the same fitting split. The jointly trained scalar tests an end-to-end alternative, not the validation-fitted scalar in the factor study. Both variants retain exact backbone initialization and one-stage training. Three-seed mean MSE:

Backbone / data / HFull+gateFull+scalarShort+gateShort+scalar
DLinear / ETTh1 / 7200.5340.4850.4970.499
TSMixer / ETTh1 / 7200.5850.6220.5400.585
MICN / ETTh1 / 7200.5040.4800.5130.495
iTransformer / ETTh1 / 7200.5060.5120.5300.532
DLinear / ECL / 960.1550.1610.1660.171
DLinear / ECL / 7200.2200.2230.2290.232
DLinear / Traffic / 960.5160.5200.5080.540
DLinear / Traffic / 7200.5440.5510.5440.569
TSMixer / ECL / 960.1520.1510.1520.151
TSMixer / ECL / 7200.2130.2150.2130.215
TSMixer / Traffic / 960.4520.4540.4550.458
TSMixer / Traffic / 7200.5110.5190.5210.520

Short equals full in the two TSMixer/ECL cells. Short+scalar wins 3/12 cells, with a 2.1% mean relative MSE increase. On ETTh1 H720 the deployed full-state gate is best for only iTransformer: full+scalar is 9.2% lower for DLinear (the only main-benchmark cell where submitted TRACE is worse than the plain backbone, 0.534 vs 0.514) and 4.8% for MICN, and short+gate 7.7% lower for TSMixer. With the gate, removing long components raises MSE in 5 of the 6 ECL/Traffic cells that contain them, at full precision (one increase is below 0.05%); the exception is DLinear/Traffic H96 (1.6% lower). These tests do not establish a rule for selecting harmful components in advance.

With full state, the trained scalar wins 3/12, with a 0.2% mean relative MSE increase. This differs from post-hoc fitting (Table 8: scalar wins 14/24, mean 1.4% lower MSE; the expanded checkpoint comparison includes a 24.0% gate disadvantage on TSMixer/ETTh1 H720). The gate's distinct role is state-conditioned weighting. We will report these protocols separately and emphasize the persistent branch as the contribution.

Q2. CycleNet's RCF. We combined RecurrentCycle with RevIN at L=96L=96 (112 cells × 3 seeds, public RevIN lr). Cycle lengths in observation steps are 24 for ETTh, 96 for ETTm, 144 for Weather and 168 for ECL/Traffic. RCF improves RevIN in 100/112 cells (mean 1.7%); TRACE improves on RevIN+RCF in 98/112 (8.2%). RCF supplies a learned periodic pattern; TRACE updates its state from observations. Representation/mixing differ and this comparison uses public-lr RCF, so this is a complete-method comparison. We will add RCF to Secs. 2 and 4.1.

Q3. Equal learning-rate search. We retuned RevIN, DDN, GAS-Norm, SAN and FAN on 112 cells at L=96L=96. We used TRACE's four lr candidates and selected by mean validation rank across horizons. SAN/FAN also retained their original lr as a candidate; SAN's adapter lr and GAS-Norm's selected γ\gamma were fixed. Each selected setting has three seeds. Public → retuned values are given for each baseline; the GAS-Norm port has no public-lr entry.

BaselineTRACE lower / 112Mean cellwise reductionOverall-mean reductionBaseline's own mean cellwise gain from retuning
RevIN105 → 1019.7% → 6.3%10.66% → 6.55%3.7% (82/112)
DDN97 → 925.8% → 4.9%6.10% → 5.71%0.9% (76/112)
SAN101 → 919.7% → 4.9%16.11% → 5.38%5.3% (75/112)
FAN108 → 10411.6% → 9.7%13.78% → 11.94%2.1% (88/112)
GAS-Norm (our port)9811.1%15.39%—

The GAS-Norm row includes ETTh2, where the port trails public-lr RevIN; excluding ETTh2, TRACE is lower in 82/96 cells by 8.1%. The shrink is largest for SAN, whose selected lr differs from the published one in 22/28 backbone–dataset pairs; its overall-mean shrink comes mainly from iTransformer/Traffic, where the published lr gives MSE 1.36–1.41 against 0.46–0.53 after retuning. ETTh1 favors retuned SAN (TRACE's mean relative increase: 1.4%). Separately, retuning RCF gives TRACE 78/112 wins (2.5% mean reduction).

Thank you for the detailed response.

Resolved:

  • Q3, equal learning-rate search. RevIN, DDN, SAN and FAN were retuned on all 112 cells with TRACE's four learning rates. TRACE is still lower in 91–104 cells
  • Q4, iTransformer's own normalization. The 4.5% reduction (25/28 cells) agrees with my estimate of 4.3%
  • Q2, CycleNet's RCF. After retuning, RevIN + RCF is the closest baseline at L = 96. TRACE is lower in 78/112 cells

Partly resolved (W2):

  • Long components. On ECL and Traffic, the new table is convincing: removing them raises MSE in 5 of the 6 cells that contain them. For DLinear and TSMixer, the best of the four trained variants (0.485, 0.540) is still above the best public-setting adapter (SAN 0.470, DDN 0.507 from Tables 25 and 27)
  • Gate. A jointly trained scalar is within 0.2% of the gate. The validation-fitted scalar is still lower in 14/24 cells. So the gate does not harm, but nothing shows that state-conditioned weighting improves accuracy. I agree with presenting the persistent branch as the contribution

Not resolved (W1/Q1):

The experiment was run as requested, and it confirms the concern:

  • With the input length selected on validation, TRACE has 2.1% higher MSE than SAN (lower in 19/40 cells) and 1.4% higher than DDN (15/40). It is 0.9% lower than RevIN (28/40) and 2.4% lower than FAN (22/40), but RevIN and FAN were not retuned in this experiment
  • TRACE needs 2.2x the peak training memory of RevIN (L=96) and 1.3x memory of DDN (L=720)

At a fixed L = 96, TRACE is 5–7% better than retuned RevIN, DDN and SAN, and 2.5% better than RevIN + RCF. When the input length is tuned, it has no average advantage.

I am keeping my current score. The numbers are now reliable, but every fair-comparison control shrinks the effect, and with a tuned input length it is about zero.

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

This paper proposes TRACE, a forecasting adapter that preserves historical information beyond the fixed input window through a persistent structural state. The state consists of a level component and fixed-frequency harmonic components, updated causally using a Kalman filter. Its closed-form forecast is combined with the backbone prediction through a lightweight learned gate, without extending the backbone's input window.

Experiments across four forecasting backbones, seven datasets, and four prediction horizons demonstrate improvements over existing normalization adapters, including RevIN, SAN, FAN, and DDN. Additional analyses investigate the contribution of out-of-window history, the complementarity between the structural and backbone forecasts, and the effect of extending the input window.

Critical strengths
  1. A simple and effective approach to incorporating historical information.

TRACE provides a practical way to leverage observations beyond the input window without modifying the backbone architecture or increasing its input length. The use of a fixed structural state-space model, together with a small learned gate, makes the approach parameter-efficient and applicable to different forecasting architectures. The consistent improvements across multiple backbones and datasets support its practical value.

  1. Extensive experiments investigating the sources of performance improvements.

The paper goes beyond standard benchmark comparisons by examining the effects of restricting historical information, replacing the structural forecasting branch, and extending the input window. In particular, the finding that improvements arise from complementary prediction errors rather than a stronger standalone structural forecast is interesting. The authors also acknowledge limitations associated with changing periodic structures, which helps clarify the conditions under which the proposed method is effective.

Critical weaknesses
  1. The novelty of state continuity is not sufficiently established against relevant prior work.

The paper identifies the reliance on current-window statistics as a fundamental limitation of existing normalization methods and presents state continuity as a key contribution. However, GAS-Norm has already explored maintaining and updating normalization statistics across forecasting origins. Although the authors acknowledge GAS-Norm, they exclude it from direct experimental comparisons due to differences in the implementation and forecasting protocol.

The level-only structural branch used as its closest analog is not equivalent to GAS-Norm, particularly in terms of its state update and prediction mechanism. Consequently, the current experiments do not sufficiently establish whether TRACE's advantages over existing history-aware approaches arise from its harmonic representation, its forecast combination mechanism, or other design choices. A more direct comparison would be important for assessing the novelty and significance of the proposed approach.

  1. Concerns about the experimental protocol and the clarity of the empirical analysis.

The structural filter is configured using both training and validation observations, including frequency identification and the selection of process-noise ratios, while the same validation segment is also used for hyperparameter selection and early stopping. Although this does not constitute test-set leakage, the effect of using validation observations to configure the structural model is not sufficiently examined. A train-only filter configuration would help establish the robustness of the reported improvements and clarify the comparability of the experimental protocols.

Additionally, the experimental section is excessively dense and heavily focused on numerical reporting. Many paragraphs enumerate performance statistics, win counts, and exceptional cases without clearly prioritizing the main findings. This is particularly noticeable in Sections 4.2 and 4.3, where the extensive numerical descriptions make it difficult to follow the central conclusions of individual experiments. A clearer separation between key findings and supporting numerical details would substantially improve readability and the interpretation of the results.

Questions for authors

1. Comparison with GAS-Norm

Given that GAS-Norm already maintains statistical states across forecasting origins, could the authors provide a direct comparison under a matched forecasting protocol? While the implementation differences are understandable, the level-only structural branch does not appear to be an equivalent substitute. Such a comparison would help clarify TRACE's contributions beyond existing history-aware normalization approaches.

2. Use of Validation Data for Filter Configuration

Could the authors report results when the harmonic frequencies and Kalman filter noise ratios are estimated exclusively from the training segment, reserving validation data for hyperparameter selection and early stopping? This would help clarify whether the additional use of validation observations for filter configuration contributes substantially to the reported improvements.

3. Generalizability of the Mechanism Analysis

The mechanism analysis in Section 4.3 is conducted on only 12 experimental settings with a fixed K=8, which differs from the configurations selected in the main experiments. Moreover, ECL and Traffic, where TRACE achieves substantial improvements, are excluded from these analyses. Could the authors extend the history-restriction and error-complementarity experiments to representative main-table configurations, particularly on ECL and Traffic? This would help verify whether the proposed explanation for TRACE's improvements generalizes to the settings underlying its main empirical results.

4. Necessity of the Learned Gate

Appendix C shows that a simple validation-fitted scalar weight performs comparably to, or even better than, the learned MLP gate in many settings. Could the authors clarify the practical necessity of the learned gate and provide a more comprehensive comparison across the main benchmark configurations? If a constant mixing weight achieves similar performance, it would suggest that the main contribution lies in the persistent structural state rather than the adaptive gating mechanism, which would be important for assessing the proposed design.

3 comments

Thank you for the constructive review. Mean gains average cellwise relative MSE reductions after seed averaging; the submission's headline uses the reduction in overall mean MSE.

W1/Q1. Contribution beyond GAS-Norm. GAS-Norm already carries state. TRACE's distinction is its level/harmonic forecast as a separate output branch. In the requested direct comparison (112 cells = 7 datasets × 4 backbones × 4 horizons, 3 seeds), TRACE has lower MSE than our GAS-Norm port in 98/112 cells, with a mean reduction of 11.1%; excluding ETTh2, 82/96 and 8.1%.

The port follows the released GAS-Norm code: per-channel Student-t score recursion with ν=100\nu=100, the value in the authors' example script; static parameters (α\alpha, β\beta, ω\omega of the location and scale recursions) fitted on training data; γ\gamma (the authors' norm strength) selected on validation over their grid {0.001,0.01,0.1,0.5}\{0.001,0.01,0.1,0.5\}, as is the backbone lr. It normalizes each input observation using preceding statistics, then denormalizes forecasts using horizon-wise predicted location/scale. State persists causally across origins; no learnable affine transform is added. The update, likelihood, fitting and horizon-rollout routines have unit tests against reference functions, but we have not reproduced the original probabilistic forecasting experiments, so the numbers characterize this port and protocol.

Per dataset (16 cells each, three seeds; mean cellwise MSE reduction, positive favoring the first method):

DatasetTRACE lower / cells, vs portPort lower / cells, vs public-lr RevIN
ECL15/16, +7.5%10/16, +7.4%
ETTh114/16, +10.6%4/16, −9.5%
ETTh216/16, +29.4%0/16, −44.1%
ETTm113/16, +4.9%14/16, +5.3%
ETTm28/16, 0.0%14/16, +3.6%
Traffic16/16, +18.7%12/16, +1.4%
Weather16/16, +6.8%12/16, +5.7%
All98/112, +11.1%66/112, −4.3%

On ETTh2 the port improves on the plain backbone (14/16, +17.3%) but trails RevIN, and its deficit grows with horizon, from 9.5% higher MSE at H96 to about twice RevIN's at H720 (means over the four backbones). This points to the horizon rollout of the location/scale statistics on that dataset rather than to the input normalization; whether the original implementation behaves the same way is not determined. The 11.1% figure therefore mixes an ETTh2 result that may be specific to our port with an 8.1% margin elsewhere, and it does not isolate the effect of harmonics.

The submitted branch control provides a separate comparison using the same independently trained backbone and validation-fitted scalar (12 factor cells, fixed K=8K=8, seed 2021). Short carried components beat every window-only alternative in all 12 cells; full state beats the best window-only branch in 8/12 but exceeds short state only on Weather H96. Level alone can also be preferable at long horizons. Thus persistence helps without making every long component useful. This control neither substitutes for GAS-Norm nor separates representation from combination: we have not run a GAS-type level/scale state through the branch, or TRACE's harmonic state as a normalization constant, and will state this as open. We will acknowledge GAS-Norm's prior state continuity in Contribution 1, explain the structural-state/output-branch distinction in Sec. 2, replace the exclusion sentence in Sec. 4.1 with the direct comparison labeled as our port, with the table above and the protocol in the appendix, and keep the controlled branch analysis separate from main-table rankings.

W2a/Q2. Training-only filter configuration. Training-only fitting retains a 4.4% mean cellwise MSE reduction over lr-selected DDN, with a 0.53% mean relative increase over submitted TRACE. DDN's backbone lr is selected from TRACE's four candidates by mean validation rank across horizons. We recomputed frequencies and noise ratios using only training data in all 112 cells × 3 seeds, retaining TRACE's lr and KK to isolate this change; validation remained available for selection and early stopping.

112 cells, 3 seedsFirst lower / cellsMean cellwise MSE reduction
TRACE (train-only) vs TRACE (submitted)51/112−0.5%
TRACE (train-only) vs lr-selected DDN88/112+4.4%

The median difference from submitted TRACE is near zero. Thus validation-based filter fitting alone does not explain the improvement. Sec. 3.3 and the appendix will distinguish fitting from selection and report this control.

W2b. Density of Secs. 4.2 and 4.3. We will lead each paragraph with its finding. Sec. 4.2 will open with the main-table result in one statistic: at a fixed 96-step input, TRACE lowers MSE in 105/112 cells against public-lr RevIN and 101/112 against lr-selected RevIN (overall-mean reductions 10.66% and 6.55%), with the largest gains on datasets whose dominant cycle exceeds the window; parameter efficiency follows, noting that RevIN remains smaller. Sec. 4.3 will organize controls around history, complementarity and alternative forecasts, defining metrics before use and distinguishing factor settings from main-table configurations. Per-horizon detail stays in Appendix O (Tables 25–28), and the affine 2×2 and count-level comparisons move to the appendix; conclusion-changing exceptions such as TSMixer/ECL H720 and the ETTh1 scalar results stay in the text.

Q3. Mechanism on ECL and Traffic. Restricting history worsens 23/24 cell–seed results and 7/8 cells by mean MSE on DLinear/TSMixer, ECL/Traffic, H96/H720. We retrained with history limited to 96 steps and compared with a no-state control: RevIN without affine parameters, trained at TRACE's lr, not Table 1's plain backbone. Reductions use each model's three-seed mean MSE.

Backbone / datavs restricted history, H96 / H720vs no-state model, H96 / H720
DLinear / ECL18.8% / 8.6%21.3% / 13.3%
TSMixer / ECL3.2% / −1.3%3.8% / 7.7%
DLinear / Traffic20.5% / 15.4%20.0% / 15.4%
TSMixer / Traffic14.2% / 8.8%12.3% / 10.8%

TSMixer/ECL H720 is the exception: carried history is worse on average, with the reversal occurring in seed 2022. Thus history helps broadly but does not explain every gain. Complementarity holds in all 24 cell–seed checkpoints: CBD/RB<1C_{BD}/R_B<1, where CBDC_{BD} is the branch-error cross moment and RBR_B the jointly trained backbone-branch MSE. This satisfies Proposition 2's small-weight benefit condition. Unlike the factor control, this diagnostic uses jointly trained branches; it supports complementarity without an exact decomposition of each gain.

Because the no-state control uses TRACE's lr, the gaps above are smaller than the main-table gaps against public-lr RevIN (Tables 25–28), most so for TSMixer. We therefore also selected RevIN's lr from the same four candidates in all 112 cells, with DDN's rule: TRACE is lower in 101/112 cells, with a 6.3% mean cellwise reduction. In Sec. 4.2's overall-mean statistic, the per-backbone reductions against RevIN change from 12.1%, 9.3%, 15.8% and 4.6% (DLinear, MICN, TSMixer, iTransformer) to 9.7%, 5.3%, 5.8% and 5.1%; in the cellwise statistic, lr reselection improves RevIN by 9.3% on TSMixer (23/28 cells), against 2.3% (DLinear), 3.5% (MICN) and −0.3% (iTransformer, 11/28). Sec. 4.2 will report the lr-selected figures next to the public-lr ones, and the appendix statement that the public-setting comparison is only mildly affected on average will be revised accordingly.

Q4. Necessity of the learned gate. The results do not establish that a learned gate is necessary; the persistent structural state is the central contribution. We fitted validation scalars to the fixed jointly trained branches in 112 seed-2021 checkpoints. Positive reductions favor the gate.

Dataset groupGate lower / cellsMedian MSE reduction vs scalar
All61/112+0.1%
Traffic16/16+4.0%
Weather13/16+0.3%
ECL8/16+0.1%
ETT (4 sets)24/64−0.2%

The overall median is small; TSMixer/ETTh1 H720 has 24.0% higher gate MSE. We also trained a single scalar jointly with the backbone in 12 main-table settings: ETTh1 H720 on four backbones, and ECL/Traffic H96/H720 on DLinear/TSMixer. With the full state and three seeds, the scalar wins 3/12 cells, with a 0.2% mean relative MSE increase; per cell, relative to the gate, the scalar ranges from 9.2% lower MSE (DLinear/ETTh1 H720, the only main-benchmark cell where submitted TRACE is worse than the plain backbone, 0.534 vs 0.514) to 6.4% higher (TSMixer/ETTh1 H720), and on ETTh1 H720 the deployed full-state gate is best for only one of the four backbones. It shares exact backbone initialization and one-stage training with the gate; state-conditioned weighting is the gate's distinct role. Contribution 2 will emphasize the structural branch, and Appendix C will distinguish the two comparisons.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts, which have addressed most of my concerns and strengthened the empirical evidence.

The distinction from GAS-Norm is now clearer, and the additional comparison is valuable. However, I still have reservations about methodological novelty, as the individual contributions of the harmonic representation and output-space combination remain insufficiently established. Moreover, the proposed approach appears to be primarily an integration of existing techniques rather than a novel forecasting mechanism.

While I acknowledge the improvements provided by the rebuttal, I will maintain my original score.

Reviewer_L879OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a lightweight adapter for non-stationary time series forecasting. The main idea is to preserve historical information beyond the fixed input window using a structural state-space model. TRACE then combines the structural forecast with the backbone prediction through a learnable gate. The authors evaluate TRACE on four backbones and seven benchmark datasets, showing improvements over several normalization methods. They also provide ablation studies and controlled experiments to better understand where the performance gains come from.

Critical strengths
  1. TRACE is lightweight and easy to integrate with existing forecasting models without changing their architectures or input window lengths. The additional number of learnable parameters is very small, which makes it a practical approach for improving existing forecasting systems.
  2. The authors conduct extensive experiments across four backbones, seven benchmark datasets, and multiple prediction horizons, showing consistent improvements over several normalization baselines.
Critical weaknesses
  1. The authors argue that existing normalization methods struggle to estimate the level and phase of periodic components from a single input window. However, not all normalization methods explicitly estimate periodic phase, and the inability to access earlier observations is a general limitation of fixed-window forecasting rather than something specific to normalization. The paper does not clearly explain the connection between these issues, which weakens the motivation for the proposed method.
  2. TRACE mainly combines a classical Kalman filter-based structural state-space model with a lightweight forecasting gate, both of which are well-established techniques. The authors also acknowledge that a simple constant combination weight performs about as well as the learned gate, which further raises questions about the necessity of the proposed design. Therefore, it is difficult to see a substantial methodological contribution beyond combining existing techniques.
Questions for authors
  1. The authors argue that existing normalization methods cannot reliably estimate the level and phase of periodic components from a single input window. However, not all normalization methods explicitly estimate periodic phase, and the limited access to historical information is a general issue in fixed-window forecasting. Could the authors explain more clearly how these two issues are connected and why this is considered a limitation specific to normalization methods?
  2. TRACE has access to historical observations beyond the input window, while the main baselines are restricted to the current window. This gives TRACE an inherent information advantage, making it difficult to tell whether the improvements come from the proposed design or simply from using more historical data. Could the authors provide a more convincing comparison with methods that also maintain historical states or use information beyond the fixed input window?
  3. TRACE relies on a set of fixed frequencies selected before training, but real-world non-stationary time series often exhibit changing periodic patterns and abrupt shifts. Although the authors acknowledge this limitation in Appendix Q, it remains unclear how well TRACE can handle such changes in practice. Could the authors provide stronger evidence that the proposed method is effective under genuinely non-stationary conditions, rather than mainly benefiting from datasets with relatively stable periodic patterns?
3 comments

Thank you for the careful review. Main comparisons cover 112 cells (7 datasets × 4 backbones × 4 horizons), with three seeds. Mean gains average cellwise relative MSE reductions after seed averaging; the submission's headline uses the reduction in overall mean MSE. Under the cellwise statistic, TRACE's gains over public-lr RevIN and DDN are 9.7% (105/112) and 5.8% (97/112); over RevIN and DDN with learning rates selected on validation from TRACE's four candidates, 6.3% (101/112) and 4.9% (92/112), so that the comparisons in Q2 can be read against one statistic. Diagnostics are labeled separately.

W1/Q1. Why compare with normalization, and how does phase relate to unused history? The connection is that periodic components can shift the mean between the input and forecast intervals, even when a normalization method does not explicitly estimate phase. Earlier observations can help estimate these components. The information constraint is shared by fixed-window methods; normalization is where we study its effect on distribution correction.

Why normalization. RevIN/SAN estimate window level/scale statistics; FAN/DDN also model periodic components. TRACE retains window normalization and adds a persistent structural forecast without changing the backbone architecture or input length. This tests history use in an adapter. Q2 separately compares history-aware methods, including GAS-Norm.

Why phase matters. For a periodic signal, the portion of the cycle covered by the input can have a different mean from the portion covered by the forecast horizon. A correction based on input statistics therefore need not match future statistics. When the period exceeds the input length, the window covers less than one cycle, motivating estimation from earlier observations. This explains the phase-to-level connection without assuming that all normalization methods estimate phase or that every window-based estimator must fail.

Evidence. The conceptual claim is only the information-set argument above; the submitted controls test it empirically. Figure 1 compares carried and window-only forecasts; Sec. 4.3 restricts the state history to the input window and substitutes alternative forecast branches under a common combination rule. These controls test the benefit of persistent state and the role of its forecast separately from the choice of mixing rule.

Contribution 1 will describe the constraint as shared by fixed-window adapters rather than as a fundamental limitation of the normalization family, and Sec. 1 and the abstract will distinguish that constraint from TRACE's separate structural forecast. The submitted controls remain the primary evidence.

W2. Methodological contribution. TRACE preserves history in an updated structural forecast branch without extending the backbone input or changing its architecture. The contribution is this adapter design and controlled evidence of when a weaker structural forecast helps a learned predictor. The fixed-parameter Kalman filter and mixing operations use established techniques.

What the combination adds. The separate branch makes carried information's contribution measurable. In the 12 controlled cells, its standalone forecast is less accurate than the independently trained backbone, yet the error cross moment satisfies CBD<RBC_{BD}<R_B, where RBR_B is backbone MSE. Proposition 2 permits improvement with a small constant weight under this condition. The history-restriction and branch-substitution controls (Sec. 4.3) further examine where the benefit comes from.

The gate. Two protocols distinguish mixing from training. With both jointly trained branches fixed in 112 seed-2021 checkpoints, the gate beats a validation-fitted scalar in 61/112 cells (median MSE reduction 0.1%). Traffic favors it in 16/16 (median 4.0%); ETT favors the scalar in 40/64 (median 0.2%), with a 24.0% gate disadvantage on TSMixer/ETTh1 H720.

We also jointly trained a single scalar in 12 main-table settings: ETTh1 H720 on four backbones, and ECL/Traffic H96/H720 on DLinear/TSMixer. With full state and three seeds, the scalar wins 3/12, with a 0.2% mean relative MSE increase. Both support exact backbone initialization and one-stage training; the gate additionally conditions weights on state. This supports a practical choice, not necessity. Contribution 2 will emphasize the branch; Appendix C will distinguish both tests.

Q2. History-aware comparisons. At L=96L=96, across 112 cells × 3 seeds, TRACE has lower MSE than our GAS-Norm port in 98/112 cells (mean reduction 11.1%; 82/96 and 8.1% excluding ETTh2) and than RevIN+RCF in 98/112 (8.2%).

The port follows the released GAS-Norm code. It applies the per-channel Student-t score recursion with ν=100\nu=100, the value in the authors' example script; static parameters are fitted on training data, and γ\gamma (the authors' norm strength) is selected on validation over their grid {0.001,0.01,0.1,0.5}\{0.001,0.01,0.1,0.5\}, as is the backbone lr. Input normalization uses statistics preceding each observation, and output denormalization uses the statistics rolled forward over the horizon. State is carried causally across origins, without a learnable affine transform. The update, likelihood, fitting and horizon-rollout routines have unit tests against reference functions; we have not reproduced the original end-to-end experiments, so the results characterize this port and protocol.

Per dataset (16 cells each, three seeds; mean cellwise MSE reduction, positive favoring the first method):

DatasetTRACE lower / cells, vs portPort lower / cells, vs public-lr RevIN
ECL15/16, +7.5%10/16, +7.4%
ETTh114/16, +10.6%4/16, −9.5%
ETTh216/16, +29.4%0/16, −44.1%
ETTm113/16, +4.9%14/16, +5.3%
ETTm28/16, 0.0%14/16, +3.6%
Traffic16/16, +18.7%12/16, +1.4%
Weather16/16, +6.8%12/16, +5.7%
All98/112, +11.1%66/112, −4.3%

The port improves on the plain backbone on ETTh2 (14/16, +17.3%) but trails RevIN there, and its deficit grows with horizon, from 9.5% higher MSE at H96 to about twice RevIN's at H720 (means over the four backbones). This points to the horizon rollout of the location/scale statistics on this dataset rather than to the input normalization; whether the original implementation behaves the same way is not determined. Sec. 4.1 will present this comparison as against our port, with the per-dataset breakdown.

RCF uses the published RevIN lr without retuning and cycle lengths of 24 (ETTh), 96 (ETTm), 144 (Weather), and 168 (ECL/Traffic) observation steps. The comparisons assess complete methods, not isolated update effects. They extend beyond window-only baselines, but carrying state does not consistently outperform extending the input (window-length comparison, Table 4), so they do not establish superiority over longer inputs. We will add the comparisons and protocol details to Sec. 4.1 and the appendix.

Q3. Genuinely non-stationary conditions. The submitted synthetic experiments directly test level, amplitude, phase and frequency changes. They support adaptation to several changes while also identifying the limit imposed by a fixed frequency basis. We have no real-data experiment that isolates a change point; this remains a limitation we will state.

Evidence under changes. Appendix F compares continued correction, no correction, and rebuilding from the window on matched streams with level, finite-ramp, amplitude, phase and frequency changes. At H96, continued correction approaches no-change error later after level/ramp changes and reduces amplitude-change excess error. At H720 the level result reverses: continued correction is worse than no correction (0.5953 vs 0.5545 for DLinear, 0.5835 vs 0.5421 for TSMixer). After phase changes, continued correction speeds recovery relative to no correction, but window rebuilding recovers fastest and at H96 also has lower post-change MSE: 0.7737 vs 0.9756 (DLinear) and 0.9118 vs 1.1103 (TSMixer). Thus recovery time and absolute error must be distinguished. The finite ramp is not evidence of arbitrary trend tracking; there is no slope state.

Real test segments. We used the main table’s test split and forecast origins for DLinear/TSMixer on ETTh1/ETTm2/Weather at H96/H720. Within each cell, all three policies used one seed-2021 factor checkpoint (lr 0.0001, K=8). The main table uses cell-specific configurations, so absolute MSEs differ. Continued correction and no correction each had the lowest MSE in 6/12 cells. On both backbones, no correction was best on ETTh1 H96, ETTm2 H720 and Weather H720. Window rebuilding was never best. These full test segments contain no identified change point.

Frequency changes. When the frequency basis becomes mismatched, continued correction can be harmful. Thus the supported scope is adaptation to level, amplitude and phase changes while the basis remains informative. On Exchange, which has no dominant cycle, TRACE is lower than RevIN in 7/12 cells but SAN is lowest in 9/12 (Appendix L). An adaptive basis remains future work, as stated in the conclusion. This scope follows from controlled interventions.

We will bring these submitted results, including the unfavorable ones, into the main-text discussion and state the successful and unsuccessful conditions together.

Presentation. We will organize the explanation around the questions above: why history matters to normalization, what the separate structural forecast contributes, and which changes the state can handle.

  • Sec. 1: explain the phase-to-level connection and define the scope of the history constraint.
  • Sec. 3: introduce state update → rollout → combination before the filter details, with the gate's practical role.
  • Sec. 4: lead with the finding of each experiment, distinguish direct method comparisons from controls, report one statistic throughout, and move detailed numerical reporting to supporting tables.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.