acceptodds
Under review as a conference paper at ICLR 2027

Anchor Before You Learn: Zero-Parameter Climatology Beats Learned Normalization

Abstract

Trainable per-instance normalization is the usual answer to non-stationary demand, and it spends every origin relearning a periodic level profile that one grouped average over the training split hands back exactly. Put a unit-by-hour-of-week climatology table in that slot, pass the network only the residual, then add the table back. The alternative costs nothing. Fix one causal origin-frozen rolling table before any test number, keep it across four backbones from 27k to 682k parameters, five demand systems in two cities and five horizons, run both sides to convergence, and all 100 cells fall to it against the strongest learned normalizer, by 10.55 to 74.26 percent of MAE, median 36.19, each cell clearing a five-seed screen, surviving Holm correction, ahead on RMSE. Inside it sits the frequency-adaptive normalizer built around periodicity, beaten in 100 of 100 at a median 44.67 percent and the pool's best member in 23. Measured against a threshold fixed before the second city was scored, 39 of its 40 cells reproduce and none fails. Independence is thinner than the count implies, five overlapping panels resting on two lattices and one split design. Grade each head on its own table and almost no credit remains. Under the pinned rolling table the 682k patch transformer returns 1.20 percent over fifteen long-horizon cells, the 27k recurrent head 0.21, two of four below nothing. A cost-free rolling climatology keeps 38 of the 100, the other 62 worth a median 0.56 percent. Work out the climatology floor before any normalization block joins a periodic-data comparison.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, per-instance normalization, non-stationarity, seasonality, evaluation protocol, spatiotemporal forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vb6aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies normalization in online forecasting. It proposes SCAN, a novel method to normalize non-stationary time series. The method is tested on time series of taxi rides at different horizon levels.

Critical strengths

None

Critical weaknesses
  • writing: the paper is hard to read.
    • it's very hard to understand what has been done from the abstract
    • the rest of the paper is similarly ambiguous. Novel and ambiguous terms are used without defining them clearly. Empirical evidence is provided without defining what is being discussed.
  • novelty and empirical results
    • in general, the classic forecasting methods also deal with non-stationarity and seasonality. It would have been interesting to see how they compare against the proposal in this setting
    • the paper lacks an extensive comparison with state-of-the-art methods
    • the benchmarks are limited to mobility data
Questions for authors

See weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer vb6a

We thank the reviewer for the report. The report raises no separate questions, so the answers below address the weaknesses in the reviewer's order, followed by the changes and a short summary.


Answers to weaknesses


W1. The paper is hard to read; it is very hard to understand what was done from the abstract, and the rest of the paper is similarly ambiguous.
The abstract now states the task, the mechanism, and the main evidence in that order. The paper defines the forecasting problem before any results: spatial units u = 1..U, hourly demand y_{u,t}, a 168-hour input window, five horizons, and a normalization slot that transforms the input and reverses the transform at the output. Section 3 gives the three-step interface, Section 4 opens with a roadmap, and per-cell detail is in the Appendix.


W2. Novel and ambiguous terms are used without clear definitions, and empirical evidence is provided without defining what is being discussed.
Table 1 defines the notation used in the method and experiments. Terms are defined at first use. "Climatology" is a per-unit hour-of-week mean estimated from the training split, not a meteorological quantity. "Matched-anchor residual gain", "static anchor" and "origin-frozen rolling anchor" are defined where they first appear; "cell group" in Section 4.3 means a system-by-backbone-by-horizon cell aggregated over seeds. Table 2 reports raw MAE for the recommended arm and the best learned normalizer, so each percentage can be checked against the values it summarizes. Table 3 names the model, the anchor history, and the comparator.


W3. Classic forecasting methods also handle non-stationarity and seasonality. It would have been interesting to see how they compare, and the paper lacks an extensive comparison with state-of-the-art methods.
Seasonal naive and the climatology floors are the classical seasonal baselines for hourly periodic demand. They are first-class arms in the floor contest, and the paper now foregrounds them as such. Per-unit ARIMA and ETS models were not run; the same applies to nearest-neighbor seasonal baselines, STL/MSTL residuals, and Prophet. Appendix M states these limits explicitly. The paper does not claim that the comparison covers every classical method. The learned state-of-the-art set is RevIN, Dish-TS, and FAN, the per-instance normalization methods this paper studies; that scope is now stated rather than implied.


W4. The paper lacks an extensive comparison with state-of-the-art methods.
The comparison covers RevIN, Dish-TS, and frequency-adaptive normalization on four backbones: GRU-64, DLinear, PatchTST, and iTransformer. It spans 100 system-by-backbone-by-horizon cells in two cities, with five seeds, a five-seed variance screen, week-block bootstrap intervals and Holm-corrected tests within each system-backbone block. The periodicity-aware comparator is the strongest available test of the claim, because it is a learned normalizer built for the periodic structure the anchor captures. The paper states this scope and does not imply a broader benchmark sweep.


W5. The benchmarks are limited to mobility data.
This is an explicit limitation in the Appendix. The evidence covers five hourly urban-demand panels in New York and Chicago; Chicago is the only second-city test. The paper makes no claim that the result transfers to electricity, traffic, weather, or exchange-rate panels. It bounds the family-level reading of the result accordingly.


What we have done to improve

  • Rebuilt the abstract and introduction so the task, mechanism, and evidence are visible before the details; SCAN and FAN are defined at first use.
  • Added a problem setting and notation table (Table 1), plus first-use definitions for the anchor terms and the unit of analysis.
  • The classical seasonal baselines remain first-class arms and are now stated as such; untested ARIMA/ETS, nearest-neighbor, STL/MSTL, and Prophet models are explicit in the Appendix.
  • Added a plain scope statement for the learned comparison: RevIN, Dish-TS, and FAN on four backbones and 100 cells, with five seeds and corrected tests.
  • Added the matched-history control and reorganized the existing FAN, Chicago, and budget evidence; moved per-cell evidence to the appendix and kept raw MAE in the main table.

Summary

ChangeRelated weakness
Abstract, introduction and Section 3 restructured; problem setting before resultsW1
Table 1 notation and first-use definitionsW2
Table 2 raw MAE; Table 3 names model, history and comparatorW2
Classical baselines stated as first-class arms; ARIMA/ETS, nearest-neighbor, STL/MSTL and Prophet limits in Appendix MW3
Explicit comparison scope: RevIN, Dish-TS and FAN on four backbones and 100 cellsW4
Second-city and mobility-only limitation in the AppendixW5

To summarize, a zero-parameter hour-of-week climatology anchor, whose specification was fixed before any test result and never tuned per cell, occupies the normalization slot and beats the best of RevIN, Dish-TS and FAN in all 100 cells of a two-city, five-system, four-backbone, five-horizon demand grid. We have revised the paper for clarity, defined the notation at first use, and stated the untested classical baselines and mobility-only scope. We hope the reviewer will review the revised version and reconsider the score.

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

This paper introduces SCAN, which is a normalization method based on periodic statistics and does not require any trainable parameters. SCAN eliminates periodic components from the input sequence using historical period means and the neural network is responsible for predicting the residuals. The authors conducted experiments on five traffic demand datasets from two cities, covering four different base prediction networks and five different prediction time periods. The results show that in all 100 experimental settings, the performance of SCAN is superior to that of learnable normalization methods such as RevIN, Dish-TS, and FAN.

Critical strengths
  1. The research motivation is valuable because it challenges the default assumption that complex models are superior to simpler methods.
  2. The approach is simple and highly practical for engineering applications.
Critical weaknesses
  1. SCAN builds an hour-of-week periodic statistics table from eight weeks of history, whereas RevIN and Dish-TS use only a one-week input window. Thus, SCAN’s gains may come from longer history and stronger periodic priors rather than normalization itself. The lack of fair comparisons is a major concern.
  2. SCAN is essentially seasonal decomposition plus residual forecasting, with limited methodological novelty. It also does not test integration with learning-based normalization methods such as RevIN or Dish-TS. The experiments therefore do not show whether the two approaches are mutually exclusive or complementary.
  3. The experimental scenarios are too narrow, and generalization remains unverified.
Questions for authors
  1. SCAN uses eight weeks of history and explicit periodicity, while other learning-based normalization methods use only a 168-hour input window. How do the authors ensure fair comparison and rule out gains from the extra historical information?
  2. How do the authors show that the gains come from the normalization mechanism itself, not from the explicit periodic priors?
2 comments
AuthorsOpenReview ↗

Response to Reviewer AbfB

We thank the reviewer for the comments. We answer the questions in order, then summarize the improvements made in response to the weaknesses.


Answers to questions


Q1. How do the authors ensure a fair comparison and rule out gains from the extra historical information?

We hold each anchored checkpoint fixed and re-anchor the same trained model at evaluation with W = 1, 2, 4, or 8 weeks. Only W changes. W = 1 uses 168 hours, exactly the history RevIN and Dish-TS receive; W = 8 recovers the eight-week result. The checkpoints are the GRU and DLinear anchored models from the shorter-budget run.

Table 3 gives the h = 24 comparison:

SystemBackboneBest learnedSCAN W=1SCAN W=8Margin at W=1
TaxiGRU-64RevIN 22.43813.98911.838+37.65%
TaxiDLinearRevIN 17.20414.17611.733+17.60%
SubwayGRU-64RevIN 37.33029.25125.076+21.64%
SubwayDLinearDish-TS 50.56437.28933.221+26.25%

Across all 20 system-by-backbone-by-horizon cell groups, W = 1 beats the best of the three learned normalizers in 20 of 20, with a median margin of +29.20% and a minimum of +1.29%. W = 8 also wins 20 of 20 at a median of +38.66%. W = 1 MAE is a median 17.26% higher than W = 8, so longer history helps, but the advantage does not depend on it. A model-free sweep gives the lower bound: at W = 1 the rolling climatology floor alone wins 10 of 15 system-horizon cells, median +10.63%.

The control uses two systems, two backbones, and three seeds. It is reported separately from the five-seed headline grid, and no number is merged across the two protocols.


Q2. How do the authors show that the gains come from the normalization mechanism itself, not from explicit periodic priors?

We agree that the mechanism and the periodic prior cannot be separated in this design. The matched-history control isolates history length, not the origin of the prior. It shows that the gain does not require more history than the baselines receive; the remaining difference is where the periodic statistic is computed and applied. The periodicity-aware FAN arm loses all 100 cells at a median 44.67%, and the RevIN layer taken verbatim from its authors loses 100 of 100 at a median 38.45%. These comparisons are consistent with the slot being the operational difference, but they do not prove that the prior and the slot act independently. We state that limit rather than claim a separation the design cannot establish.


What we have done to improve


  • Added the matched-history control in Section 4.3 and Table 3. The trained checkpoint is fixed and only W changes; W = 1 matches the learned normalizers' 168-hour history. A model-free sweep supplies the conservative bound.
  • Made the periodicity-aware FAN comparison the explicit test of the periodic alternative: it loses all 100 cells at a median 44.67%, and the verbatim RevIN layer also loses 100 of 100 at a median 38.45%.
  • Narrowed the claim: the paper no longer implies that the mechanism and prior are separated, and it records that a learned normalizer applied to the residual is untested. Because that hybrid is untested, the paper does not claim that the two approaches are mutually exclusive; it treats them as alternatives in the same slot and names the hybrid as the first follow-up.
  • Stated the scope plainly: two cities, five demand systems, hourly resolution, and 100 cells; Chicago is a second-city test, not an external replication. The Limitations appendix bounds family-level and benchmark-level claims.

Summary

ChangeRelated weakness or question
Section 4.3 and Table 3: fixed checkpoints, W swept over 1, 2, 4, and 8 weeksW1, Q1
Model-free anchor-history sweep reported as the lower boundW1, Q2
Claim narrowed: mechanism and prior are not separated by this designW2, Q2
Explicit statement that a learned-normalizer-on-residual hybrid is untestedW2
Periodicity-aware FAN comparison and verbatim RevIN control made the explicit evidenceW2, Q2
Scope of five systems, two cities and 100 cells; Chicago as a second-city testW3
Bounded claims and calibrated omissions in Appendix MW3
Reference control labeled separately from the five-seed headlineW1, Q1

The zero-parameter hour-of-week anchor beats the best learned normalizer in all 100 cells, and still wins 20 of 20 when its anchor history is cut to 168 hours. We have added matched-history control, stated the mechanism-versus-prior limit, and clarified the five-system, two-city scope. We hope the reviewer will review the revised version and reconsider the score.

Reviewer_SY6eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a novel, parameter-free preprocessing/normalization method. It is implemented through R(h), matched-anchor residual accounting, which roughly works by placing anchor points on the time series and then performing different normalizations. Experiments demonstrate that this method is more effective than other parameterized normalization methods.

Critical strengths
  1. The authors propose a zero-parameter normalization method and demonstrate its effectiveness in the meteorological domain.
  2. The experimental setup is very detailed and rigorous.
Critical weaknesses
  1. The writing reads more like a technical report than an academic paper. The authors use extensive data in the abstract and introduction to demonstrate the method's effectiveness, but lack an explanation of the method and its underlying principles.
  2. The paper has no problem definition, and I find it difficult to understand what task the authors are targeting. CLIMATOLOGY is a very broad concept; judging from the actual experiments, this appears to be forecasting into the future, but I believe the authors should state this explicitly, otherwise it is hard to follow.
  3. The authors do not adequately explain the motivation behind the method design. Placing anchor points is an intuitive and simple approach — I am very fond of simple work — but only on the premise that the authors can effectively justify the approach.
  4. I genuinely found the experimental section hard to read; it is filled with a large number of conclusions but lacks summarization.
  5. The use of generative AI is not in itself a flaw of a paper, provided that a human has thoroughly checked it to ensure readability and quality.
Questions for authors
  1. Line 140: is TtrainT_{train} the training set? Please clarify.
  2. Line 145: the first letter should be capitalized.
  3. In Figure 1, what is the difference between the two yellow dashed lines? This further makes the figure difficult to understand.
  4. Lines 153–155 appear to be defensive LLM-style writing; there is no need to emphasize that hyperparameters are not model parameters. Suggest revising.
  5. What does hh mean in Equation 4? It also appears in Equation 3, but the authors do not seem to define this symbol. Formally it looks more like the raw data, yet line 205 states h∈{24,72,168}h \in \{24, 72, 168\}, which is confusing.
  6. Is Section 3.3 the experimental setup or the method? The content looks more like experimental setup.
  7. Why is the experimental configuration repeatedly mentioned both in lines 285–289 and in the REPRODUCIBILITY STATEMENT?
  8. The presentation of Figure 2 and Table 1 — comparing against the best result rather than the raw values — is really unintuitive. I suggest comparing against the raw data instead.
  9. I do not understand why the use of SHA-256 is emphasized. This is clearly not a blockchain-related paper. I would like the authors to explain the necessity of this writing choice; at the very least, I believe open-sourcing the code and raw data would demonstrate the correctness of the results far better than providing hash values, and the authors have not done the latter. In my view, a paper generated directly by an LLM without human review should be rejected.
2 comments
AuthorsOpenReview ↗

Response to Reviewer SY6e

We thank the reviewer for the detailed report. We answer the questions in order, then summarize the improvements made in response to the weaknesses.


Answers to questions


Q1. Line 140: is the quantity the training set?

Yes. The quantity is the training period, not the validation or test block. It is named at first use and defined in Table 1 as the period used to fit the static anchor.


Q2. Line 145: the first letter should be capitalized.

Fixed. The sentence now reads "The result is a U × 168 table computed once...".


Q3. In Figure 1, what is the difference between the two yellow dashed lines?

One trace is the frozen hour-of-week anchor. The other is the per-window mean that RevIN hands the same backbone. Their vertical separation is the level the anchor removes. The Figure 1 caption now identifies both traces.


Q4. Lines 153-155 appear to be defensive LLM-style writing.

The passage asserting that hyperparameters are not model parameters has been removed.


Q5. What does the symbol in Equation 4 mean? It also appears in Equation 3 and is not defined.

Table 1 now defines the symbols used across Equations 3 and 4: the anchor a, the residual r after subtracting the anchor, and the backbone f_theta. The interface text uses them consistently, and the sentence that made one of them look like raw data has been rewritten.


Q6. Is Section 3.3 the experimental setup or the method?

It is the evaluation design. It is now Section 3.4, "Evaluation design"; the setup details are in Section 4.1.


Q7. Why is the experimental configuration repeated in the reproducibility statement?

The configuration is consolidated in the canonical hyperparameter table (Table A2). The method text points to that table, and the Reproducibility Statement records the environment rather than repeating the configuration.


Q8. Why do Figure 2 and Table 1 compare against the best result instead of raw values?

The headline table (Table 2) now gives raw MAE first for the recommended arm and the best learned normalizer, with the margin as a secondary column. The full 100-cell grid is a raw-MAE table (Table C1) in the appendix. The relative-margin heatmap is no longer the main presentation of the evidence.


Q9. Why is SHA-256 emphasized instead of releasing code and data?

The SHA-256 narrative and the hash column have been removed. The supplementary package contained the evaluation code, its run script and manifest, the three New York demand panels, and the matched-history results.


What we have done to improve


  • Rebuilt the abstract and introduction so the task, mechanism, and evidence appear in that order. SCAN and FAN are defined at first use.
  • Added a problem setting and notation table (Table 1) before any results; defined climatology as a per-unit hour-of-week mean from the training split.
  • Justified the method design: the grouped average is the parameter-free estimator of the periodic profile, and the rolling variant is causal because it is frozen at the forecast origin.
  • Reorganized the experiments around four questions, opened with a roadmap, and moved per-cell detail to the appendices. Table 2 reports raw MAE; Table C1 carries the full grid.
  • Added the matched-history control (Section 4.3, Table 3) and folded the existing FAN, Chicago, and budget evidence into the new roadmap.
  • Completed a readability pass: defensive metadiscourse and over-signposting were removed. The LLM-use statement names the tools and the scope of their use.
  • Removed the SHA-256 manifest narrative.

Summary

ChangeRelated weakness or question
Abstract, introduction, and Section 3 restructured around task, mechanism, and evidenceW1
Problem setting, first-use definitions and Table 1 notationW2, Q1, Q5
Method motivation for the grouped average and the causal rolling variantW3
Experiment roadmap, raw-MAE Table 2 and full grid in Table C1W4, Q8
Readability passW5
Capitalization correctedQ2
Figure 1 caption identifies both dashed tracesQ3
Passage removedQ4
Evaluation design moved to Section 3.4; setup in Section 4.1Q6
Configuration consolidated in Table A2Q7
SHA-256 removed; supplementary code, panels and matched-history results keptQ9

A fixed-by-design, zero-parameter hour-of-week anchor takes the normalization slot and beats the best of RevIN, Dish-TS and FAN in all 100 demand-forecasting cells. We appreciate the reviewer’s careful attention to AI use and confirm that the study and manuscript were authored and reviewed by authors, with LLMs used as disclosed, and we have revised the manuscript substantially for clarity, added details of the problem setting and notation, and kept the code, New York panels, and matched-history results in the supplementary package.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.