acceptodds
Under review as a conference paper at ICLR 2027

How Good Can Linear Models Be for Time-Series Forecasting?

Abstract

Time-series forecasting research has been moving steadily toward larger architectures, from specialized transformers to general-purpose foundation models, on the assumption that capacity is what unlocks accuracy. We take the opposite position: most of the gap can be closed at far lower cost by tuning preprocessing rather than scaling models. We use Ridge regression as the testbed, since it has a closed-form solution and interpretable weights, which let the optimal hyperparameters be read off the search directly. We search over context length, local normalization, regularization, and augmentation on eight standard benchmarks and find three patterns. (1) Optimal lookback is strongly series-specific and often non-monotonic in forecast horizon, with fitted power-law exponents ranging from on ETTm2 to on Exchange and Traffic, challenging the convention that longer horizons need longer history. (2) Normalizing over a learned trailing fraction of the context, rather than its entirety, is almost universally preferred. (3) Series within the same dataset often disagree on hyperparameters; the optimal degree of cross-series sharing varies from fully shared to fully per-series. The resulting models beat prior linear forecasters on 7 of 8 benchmarks (a near-tie on the 8th), and attain lower average MSE than Transformer and MLP baselines under matched settings on seven of eight. The optimized hyperparameters also serve as a diagnostic on the data itself, revealing structures that larger models absorb silently into their learned parameters.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series Forecasting, Linear Models

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_xqY4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper compares simple linear models (e.g., OLS, FITS, DLinear) and nonlinear deep learning models (PatchTST, TimeMixer, PatchMLP) against closed-form Ridge regression paired with an extensive preprocessing and hyperparameter tuning pipeline called SearchCast. By searching over lookback length, trailing-window local normalization, regularization, and noise augmentation across grouped horizons and series, SearchCast outperforms the compared linear and nonlinear baselines on 7 out of 8 standard benchmarks

Critical strengths
  1. This paper builds directly on prior work showing that linear models can achieve strong forecasting performance and that many recent linear variants are functionally equivalent to linear regression over augmented feature sets. I really appreciate the approach that rather than proposing yet another linear architecture, the authors usefully focus on the preprocessing side.
  2. Although I have some concerns regarding the evaluation setup, the empirical results generally support the claim that SearchCast improves over standard linear and nonlinear baselines on the considered benchmarks at a low computational cost.
Critical weaknesses
  1. There is a noticeable discrepancy between the framing in the Abstract/Introduction and the actual evaluation. The opening claims suggest that tuning preprocessing on a linear model can close the accuracy gap with very high-capacity architectures, explicitly citing general-purpose foundation models (e.g., first five lines of the Abstract and Section 2). However, the evaluation (a) does not include any foundation models or truly high-capacity models, and (b) relies on small/standard LTSF benchmarks where 4 of the 8 datasets (ETTh1, ETTh2, ETTm1, ETTm2) come fromvery similar sources and on the largest, highest-dimensional benchmark (Traffic, 862 series), SearchCast falls behind PatchTST.
  2. None of the baseline methods are given access to the key preprocessing improvements introduced for SearchCast. While running a full per-group search for deep models is computationally expensive at least the trailing-window local normalization (r<1) should have been evaluated on the baselines, especially given the paper's central finding that "full-window normalization is essentially never selected."
  3. The text implies that time- and frequency-domain noise augmentation is a meaningful contributor to performance, yet Figure 8(a) shows that Freq, Time, and None are each selected roughly one-third of the time across datasets, which resembles uniform random sampling. Furthermore, also from a theoretical standpoint, adding zero-mean noise to inputs in closed-form Ridge regression acts similarly to l2l_2 regularization, which should be largely redundant given the grid search over α\alpha
Questions for authors

Beyond the weaknesses above, I would like to ask the authors:

  1. Table 6 gives standard deviations only for dataset averages, and only for SearchCast and the nonlinear baselines. Could you also provide standard deviations in Table 1?
  2. Can you provide an ablation table isolating the marginal MSE gain of each preprocessing step?
4 comments
AuthorsOpenReview ↗

Response to Reviewer xqY4 [1/2]

Thank you for the close reading and constructive suggestions.

W1. Does the evaluation match the framing?

Yes, with the two additions you ask for. The framing is point forecasting on standard long-horizon benchmarks, and the abstract and introduction state it. The evaluation now includes a zero-shot foundation model (TimesFM 3.0), against which SearchCast is ahead on six of the seven datasets outside its pretraining corpus, and PEMS08, a dataset with 170 sensors and strong cross-channel dependence, where SearchCast is ahead of all five deep forecasters (General Q2). The revision states that the four ETT datasets share one source. On Traffic, SearchCast is within 0.010 of the best deep forecaster (0.407 vs. 0.397) at a small fraction of the training cost.

W2. Does trailing-window normalization help the deep baselines?

No. We changed PatchTST to compute its normalization mean over only the last 30% or 10% of the input window, or over the window SearchCast chose per dataset, with everything else fixed:

DatasetPatchTSTLast 30%Last 10%SearchCast's windowPatchTST, range over 3 seeds
ETTh10.4190.4170.4190.4260.419–0.442
ETTh20.3510.3520.3490.3510.351–0.355
ETTm10.3490.3470.3490.3520.349–0.349
ETTm20.2560.2560.2570.2570.256–0.260
Exchange0.3530.3530.3530.3550.353–0.378
Weather0.2240.2240.2250.2270.223–0.225
Electricity0.1710.1710.1710.1750.170–0.171
Traffic0.3970.3980.3980.4040.397–0.398

The best of the three windows barely changes PatchTST's MSE. With α large in most cells, Ridge shrinks the forecast toward the subtracted level, so that level acts as the forecast's anchor and choosing it matters. PatchTST's output is not shrunk toward its normalization level in this way, which may be because of its non-linear nature.

W3. Augmentations

Is augmentation actually selected, or does Figure 8a just mirror uniform sampling? Where the search selects augmentation, the noise is small: the median selected σ lies between 0.003 and 0.015 per dataset. Figure 8a counted all trials; the revised figure counts the selected configurations. At these noise levels, augmentation changes average MSE by at most 0.001 on the eight benchmarks and lowers the three-seed mean on PEMS08 by 0.003 (0.187 → 0.184). Augmentation stays in the search space as an optional axis, and the revision reports the selected levels and their effect.
Isn't zero-mean input noise in closed-form Ridge redundant with L2 regularization? For additive time-domain noise, yes: in expectation it adds Nσ² to α (Bishop, 1995). Frequency-domain noise is different: it scales the forecast by roughly 1/(1+σ²), which α alone cannot express. The search keeps both as options and selects small noise levels.

AuthorsOpenReview ↗

Response to Reviewer xqY4 [2/2]

Q1. Can you report standard deviations in Table 1?

The revision moves these into Table 1, which now shows mean ± standard deviation over three runs at every horizon, for SearchCast and all rerun baselines. OLS is deterministic, and FITS and DLinear are quoted from Toner & Darlow (2024), so they, alongside TimesFM 3.0, remain single numbers.

Q2. Ablation that isolates each preprocessing step

Yes. Starting from OLS (lookback 720, full-window normalization), each column adds one step; all channels share one group until the last column (3 seeds):

DatasetOLS (L = 720)+ α search+ lookback search+ trailing-window norm+ augmentation+ series grouping
ETTh10.4250.4140.4060.4050.4040.404
ETTh20.3340.3310.3320.3330.3340.334
ETTm10.3560.3550.3490.3490.3470.347
ETTm20.2490.2480.2370.2380.2390.239
Exchange0.3940.3930.3470.3420.3420.347
Weather0.2390.2390.2370.2370.2370.215
Electricity0.1620.1620.1600.1600.1600.160
Traffic0.4100.4100.4060.4060.4060.407

Removing one step at a time from the full model:

DatasetFull− lookback search− trailing-window norm− augmentation
ETTh10.4040.4100.4060.405
ETTh20.3340.3360.3320.333
ETTm10.3470.3540.3470.347
ETTm20.2390.2490.2370.238
Exchange0.3470.4280.3580.346
Weather0.2150.2150.2150.215
Electricity0.1600.1620.1600.160
Traffic0.4070.4090.4070.407
  • Lookback search carries most of the gain. Adding it lowers MSE on 7 of 8 datasets, by up to 0.046 (Exchange), and removing it from the full model costs 0.002–0.081 on 7 of 8.
  • Per-channel models carry Weather (0.237 → 0.215), whose 21 channels measure different physical quantities.
  • α search matters on ETTh1 (−0.011), and trailing-window normalization on Exchange (−0.005 when added, +0.011 when removed).
  • Augmentation stays within 0.0014 on the eight benchmarks and lowers MSE by 0.003 on PEMS08 (W3).

Thank you for your response. Overall I mantain my assessment: the idea is interesting, but several claims remain overstated or insufficiently supported by the evaluation. In particular:

  • Contribution (2) in the abstract and introduction claims that trailing-window normalization is almost universally preferred and "consistently improves accuracy". Which is why I asked to test it on the baselines. However, the ablation in Q2 shows that it makes essentially no difference in most of the datasets (sometimes slightly hurting, only helping in 1), while most gains come from lookback search. Moreover, W2 shows it does not help PatchTST either, which conflicts with the unchanged claim in the Conclusion that these preprocessing choices transfer to any model class.

  • Adding a single foundation model that you could not test on all the datasets provides limited support for the broad opening framing in the Abstract and Introduction (which remains unchanged in the PDF). Once Traffic and PEMS08 are excluded for pretraining overlap, 4 of the 7 evaluated datasets are ETT variants from the same source, and TimesFM 3.0 matches/slightly edges out SearchCast on Electricity

  • The response confirms W3, yet augmentation is still presented throughout the paper as a core preprocessing axis

These are prominent examples of flaws in the evaluation for which I cannot recommend acceptance.

Dear Reviewer,

The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.

Thank you for your time and thoughtful consideration.

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

This paper investigates how far carefully tuned linear models can go in long-term time-series forecasting. Instead of increasing model capacity, the authors optimize preprocessing choices for Ridge regression, including context length, local normalization, regularization, and data augmentation, with per-horizon and per-series hyperparameter search. Experiments on eight standard benchmarks show that the resulting SearchCast pipeline outperforms prior linear forecasters on seven datasets and is competitive with or better than Transformer and MLP baselines on most benchmarks. The analysis further shows that optimal lookback, normalization scope, and cross-series parameter sharing are highly dataset- and series-dependent, suggesting that preprocessing choices themselves provide useful diagnostics of time-series structure.

Critical strengths

Simple yet effective methodology. The paper shows that substantial forecasting gains can come from preprocessing and hyperparameter tuning without increasing model complexity.

Strong empirical performance. SearchCast outperforms prior linear models on seven of eight benchmarks and remains competitive with stronger nonlinear baselines.

Comprehensive hyperparameter analysis. The study systematically examines context length, normalization, regularization, augmentation, and parameter sharing across horizons and series.

Good interpretability and diagnostic value. The optimized hyperparameters and learned Ridge weights provide useful insights into nonstationarity, seasonality, and cross-series heterogeneity.

Critical weaknesses

The paper demonstrates that preprocessing can close much of the gap, but does not establish that model capacity is generally unnecessary. The results are specific to standard long-term forecasting benchmarks, and more complex settings involving probabilistic forecasting, missing values, covariates, or cross-domain transfer may still benefit substantially from richer representations.

The method remains channel-independent at the prediction stage. Cross-series interactions are not directly modeled by Ridge, which may explain the weaker performance on Traffic and could limit performance on datasets with strong inter-variable dependencies.

Baseline tuning is not fully budget-matched. SearchCast includes the full search pipeline, whereas several baselines use previously selected or fixed configurations. This makes it difficult to isolate whether the gains come from better preprocessing or simply more aggressive tuning.

Questions for authors

How sensitive are the reported gains to the hyperparameter-search budget, and would the advantage remain under a strictly matched tuning budget for all baselines?

Can you evaluate SearchCast on more diverse or distribution-shifted datasets to verify whether the discovered preprocessing rules generalize beyond standard long-term forecasting benchmarks?

3 comments
AuthorsOpenReview ↗

Response to Reviewer stDW

Thank you for the clear summary and the precise questions.

W1. What does the paper claim about model capacity?

We do not claim that model capacity is generally unnecessary. Our claim is the one in the abstract: “most of the gap can be closed at far lower cost by tuning preprocessing rather than scaling models.” The reviewer’s first clause actually restates this claim, and the results support it. With preprocessing search, a closed-form linear model attains lower error than five retrained deep forecasters on 8 of 9 datasets, at 0.2 to 9 minutes per dataset. We evaluate on the standard long-horizon benchmarks because the deep forecasters were proposed and compared on them. Probabilistic forecasting, missing values, covariates, and cross-domain transfer are out of the scope of this paper.

W2. Could channel independence explain the Traffic result and limit SearchCast on correlated data?

Channel independence does not explain Traffic, and it does not put SearchCast behind on PEMS08. As you note, each SearchCast forecast takes only its own channel's history as input. However, across channels, series grouping shares one model within each group (kTJY W3&Q3). We further validate this question using PEMS08, which has the strongest inter-variate dependence of our nine datasets: other sensors' history lowers a ridge forecaster's error by 13.2% for the median sensor. On it, SearchCast (0.184) is ahead of all five deep forecasters, including the channel-dependent iTransformer (0.297), which reads all sensors' history. We note that series grouping contributes much of this gain: one model per sensor reaches only 0.222 MSE.

W3. Is the tuning budget matched between SearchCast and the baselines?

Yes, in two senses.

  • Wall-clock: SearchCast's whole search (see Table below) costs less than one training run of any deep baseline, and the baselines' TFB configurations were themselves already tuned per dataset and horizon.
  • Same trials per search unit: Appendix G (Table 11) gives PatchTST and TimeMixer 20 configurations per horizon, SearchCast's number of trials per cell, on ETTh1 and ETTh2. We extend the PatchTST search (5 lookbacks × 4 widths, selected by validation loss, seed 2021) to four more datasets. PatchTST is the strongest deep baseline on five of the six.
DatasetSearchCastPatchTST (TFB config)PatchTST + searchPatchTST + search (min)SearchCast (min)
ETTh10.4040.4190.42438.60.3
ETTh20.3340.3510.36637.70.2
ETTm10.3470.3490.350147.80.7
ETTm20.2390.2560.254136.00.8
Weather0.2150.2240.221210.16.8
Exchange0.3470.3530.39822.70.9

The search changes PatchTST's MSE by −0.003 to +0.045 and never passes SearchCast. Each PatchTST search costs 25–210× SearchCast's whole pipeline. TimeMixer, searched on ETTh1/ETTh2 (Appendix G), did not improve.

Q1. How sensitive are the gains to the search budget?

Appendix F covers larger budgets: with 10 to 50 trials per cell (Table 10), average MSE stays within 0.006 of the default 20 trials on every dataset and saturates at about 15 trials. We extend the sweep down to a single trial, with only the budget changes (3 seeds):

DatasetNo search (OLS, L = 720)1 trial2 trials5 trials10 trials (App. F)20 trials (default)PatchTST
ETTh10.4250.4190.4170.4130.4040.4040.427
ETTh20.3340.3510.3480.3420.3340.3340.353
ETTm10.3560.3590.3600.3550.3530.3470.349
ETTm20.2490.2640.2630.2460.2440.2390.258
Exchange0.3940.3620.3550.3460.3450.3470.368
Weather0.2390.2270.2270.2170.2170.2150.224
Electricity0.1620.1750.1730.1630.1610.1600.170
Traffic0.4100.4600.4540.4120.4100.4070.398

With 5 or 10 trials, SearchCast is ahead of PatchTST on 6 of 8 datasets; with the default 20, on 7 of 8. A single random trial is worse than the fixed α-only setting on 6 of 8 datasets, so the gain comes from what the search selects, not from the sampling range.

Q2. Can you test on more diverse or distribution-shifted datasets?

Yes. PEMS08 adds a new sampling rate (5 minutes) and strong cross-channel structure, and SearchCast is ahead of all five deep forecasters there (0.184 vs. 0.297–0.355). Please refer to Table 1 for detailed results. The search adapts to it: it selects lookbacks of 1,399–2,044 steps (about 5–7 days) in every cell and shares one model across two groups of 85 sensors.
Exchange is the shift case: the search selects short contexts (median 35 steps) and last-value centering, and SearchCast is ahead of all five deep forecasters there (0.347 vs. 0.360–0.402).

Dear Reviewer,

The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.

Thank you for your time and thoughtful consideration.

We thank the authors for their detailed replies and supplementary experiments.

However, I do not think these explanations adequately dispel my doubts about the generalization of the method and the fairness of the experiment. On the basis of all the evidence, I decided to maintain my rating.

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

The paper presents the conceptual idea and an actual time-series forecaster, showing that most of the performance gains can still be achieved by moving compute from the model architecture to the pre-processing pipeline (HP tuning, etc.). The resulting Ridge regression model + searched pre-processing (SearchCast) is shown to outperform several linear and 3 Transformer/MLP baselines on standard benchmarks. The analysis in Section 5 also dissects the search results into several insights.

Critical strengths
  • SearchCast a very compute-friendly approach that rivals some more complicated models at a fraction of the cost. In an era in which increasing data collection and computation are a must, I am in favor of simple, compute-efficient approaches.

  • The insights coming from the analysis in Section 5 are interesting, in particular (1) "normalizing over a learned trailing fraction of the context" and (2) "series within the same dataset often disagree on hyperparameters". From these findings, it is likely that future time-series forecasters can build on existing heuristics and devise new ones that dynamically normalize context and perform per-series tuning.

  • The paper is well-organized and written. It also has several plots and interesting analyses.

Critical weaknesses
  • The idea of keeping a very simple baseline and tuning its components to raise performance is not new in the time series field (AutoARIMA [1], AutoAR [2]). The paper is missing references and experimental comparisons.

  • The paper dedicates a related work section to FMs but does not evaluate any of them.

  • The paper follows the standard benchmark suite of ETT/Electricity/Traffic, etc., but does not evaluate SearchCast and baselines on datasets that might have more correlated channels. SearchCast treats channels completely independently, and I suspect it could be less effective in scenarios where inter-channel dependencies are stronger. Current benchmarks mostly have intra-variate dependencies [4].

  • The paper overlooks more recent research in time series forecasting, as the majority of references date back to 2024, despite the boom in publications in the field in the past 2 years. This leads to some claims that are somewhat misleading: (1) "fixing context length" has been criticized in recent work (see [3]), (2) linear models are not considered that weak, as they have been shown to rival more complicated models on several occasions [DLinear,2,3].

  • Figures 1 and 6 need better presentation or reframing

  • The paper doesn't explicitly reference the appendix and leaves some claims not backed up: e.g., in the conclusions it mentions CNN baselines but the main paper does not contain any; it talks about a fraction of the cost without referring to the computational analysis in the appendix; finally, the "preprocessing choices transfer to any model class" without actually backing up this claim (unless I missed an analysis in the appendix).

[1] Nhits: Neural hierarchical interpolation for time series forecasting, AAAI 2023

[2] Specialized Foundation Models Struggle to Beat Supervised Baselines, ICLR 2025

[3] There are no Champions in Supervised Long-Term Time Series Forecasting, TMLR 2026

[4] A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle, ICML 2025

Questions for authors

I like the main idea of the paper and the insights coming from the analyses. However, I would require answers to the following questions before it is ready for acceptance:

  • Can the authors compare SearchCast against AutoARIMA and AutoAR?

  • Can the authors compare SearchCast against FMs?

  • Can the authors evaluate SearchCast on datasets with more inter-variate dependencies?

  • Can the authors update the related work section with newer works?

  • Can the authors clarify what exactly are presenting in Figures 1 and 6?

4 comments
AuthorsOpenReview ↗

Response to Reviewer kTJY [1/2]

Thank you for the detailed review and the pointers to recent work.

W1&Q1. How does SearchCast relate to tuning a simple model (AutoARIMA, AutoAR), and how does it compare?

SearchCast shares the tuning principle with AutoARIMA [1] and Auto-AR [2], which the revision cites and compares against. It differs in what it selects, and it is ahead of both on all nine datasets (table below).
AutoARIMA and Auto-AR select one model order per series or dataset with an in-sample criterion; Auto-AR shares one AR model (order up to 192) across all channels and forecasts recursively. SearchCast selects the lookback, normalization window, regularization and augmentation for each group of forecast steps and each group of series, by cross-validation, and forecasts each step directly from lookbacks up to 2,048. The closed-form fit makes this affordable: 300 trials on ETTh1 take 0.3 minutes. The two findings you highlight, the trailing normalization window and the disagreement between series, come from this selection; one order per dataset cannot produce them.

DatasetSearchCastAuto-AR (official recipe)AutoARIMA (MSTL)
ETTh10.4040.4380.457
ETTh20.3340.3670.455
ETTm10.3470.3840.549
ETTm20.2390.2760.330
Exchange0.3470.8330.360†
Weather0.2150.9950.289
Electricity0.1600.1880.178
Traffic0.4070.9100.485
PEMS080.1840.8440.213

† Exchange has no seasonal period, so AutoARIMA runs on the raw series.

W2&Q2. Can you compare against foundation models?

Yes. We evaluate TimesFM 3.0 (released August 2026) zero-shot on every test window with a 2,048-step context, forecasting each channel separately. SearchCast is ahead on six of the seven datasets outside TimesFM's pretraining corpus and ties on Electricity (0.1596 vs. 0.1595), using each dataset's training split alone (General Q2).
Note that the training corpus of TimesFM includes GiftEvalPretrain, which contains PEMS08 in full (including the test split), and Monash traffic_hourly, which matches our Traffic file value for about 95% of the test windows. Comparisons on these two datasets are not fair and thus omitted.

W3&Q3. Can you evaluate datasets with stronger inter-variate dependence?

Yes. We added PEMS08 and measured inter-variate dependence on all nine datasets. PEMS08 is the one dataset with strong dependence, and SearchCast is ahead of all five deep forecasters on it.
PEMS08 has strong inter-variate dependence; the eight original benchmarks do not. We measure the data-level counterpart of the inter-variate dependence in [4]: how much other channels' history lowers the test error of a ridge forecaster that already uses the channel's own history, with the penalty selected on validation. On PEMS08, other sensors' history lowers the error by 13.2% for the median sensor. On the eight original datasets, the median effect ranges from a 0.7% increase to a 1.5% reduction, which supports your observation from [4] that they are mostly intra-variate.
SearchCast is ahead on PEMS08. On TFB's 6:2:2 split it reaches 0.184 average MSE, ahead of all five retrained deep forecasters (iTransformer 0.297, PatchMLP 0.329, PatchTST 0.330, TimeMixer 0.342, TimesNet 0.355) and AutoARIMA (0.213).
SearchCast does not fit channels in isolation. Series grouping splits the channels into groups, and each group shares one model fit on the pooled training windows of its channels. On PEMS08, two groups of 85 sensors reach 0.184, against 0.222 with one model per sensor. Each sensor's forecast takes only that sensor's history as input. The channel-dependent iTransformer reads all sensors' history, and SearchCast is ahead of it by 0.113.

AuthorsOpenReview ↗

Response to Reviewer kTJY [2/2]

W4&Q4. Can you update the related work with 2025–26 papers?

Yes. The revision will add the works you cite ([1]–[4]) and recent forecasters and foundation models (e.g., Chronos-2, TimesFM 3.0). It cites [3] and TFB for lookback tuning, and DLinear, [2] and [3] for competitive linear models, and positions our question relative to them: how far a linear model can go with fine-grained preprocessing search, and what the searched settings reveal about the data.

W5&Q5. Explaining Figures 1 and 6.

What exactly does Figure 1 show? Each panel shows one channel around a test time t = 0. Black is the input history, and the colored line is the future, colored by forecast step h. Each horizontal bar has the color of one step h and extends back as far as the lookback the search chose for that step. For example, Exchange uses only about 30 past steps for most horizons, while Weather and Electricity use up to about 2,000.

What exactly does Figure 6 show? It shows the size of each learned Ridge weight for one channel per dataset. The x-axis is the forecast step h, the y-axis is the lag (0 = most recent input), and white areas lie beyond the chosen lookback. The diagonal stripes are seasonal anchors: the model relies on past values exactly one or more periods before the target, which appear as diagonals in these coordinates.

W6. Where does the paper support the conclusion's statements on CNN baselines, cost and transfer?

The revision will support each in the main text and add pointers to every appendix.

  • CNN baselines: As in the table in General Response Q2, Table 1 will include TimesNet (a CNN), and SearchCast has lower average MSE on all 9 datasets, PEMS08 included.
  • Cost: the claim cites Appendix D. The whole SearchCast pipeline takes 0.2–9 minutes per dataset, versus 1–283 minutes for one PatchTST training run.
  • Transfer to other models: the revision scopes this statement to regularized linear models, where the mechanism is explicit: Ridge shrinks the forecast toward the subtracted level, so the choice of that level matters (xqY4 W2).

Dear Reviewer,

The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.

Thank you for your time and thoughtful consideration.

I would like to thank the authors for their rebuttal and the effort they put into it. Unfortunately, parts of my concerns remain unaddressed. To this end, I am keeping my score. Please see the explanations in the following:

  • The comparison with the requested baselines (AutoAR and AutoARIMA) is not included in the revised PDF. The AutoAR results are also lower than the values reported in the original paper (Table 14). If I understood the authors' rebuttal correctly, their AutoAR implementation uses lags up to 192, whereas the original paper reports lags up to 512. SearchCast goes back up to 2,048. This comparison doesn't seem fully fair to me, and I suspect it could be the cause of AutoAR's poor performance.

  • The related work suggested to be incorporated is not part of the revised manuscript (including the two new baselines, AutoAR and AutoARIMA).

  • I understood the meaning of Figures 1 and 6. I would like to thank the authors for the explanations. However, the caption remains too brief to make this immediately understandable, and I consider this a remaining presentation issue.

  • The revised conclusion still states: "These preprocessing choices transfer to any model class".

Comments on the paper

1 comment
AuthorsOpenReview ↗

General responses

We thank all three reviewers for careful and constructive reviews. The common requests were foundation models and auto-tuned classical baselines, a dataset with stronger cross-channel dependence, a component ablation, and a clearer statement of scope. The new evidence, under the Table 1 splits, normalization and metrics:

  • Against five deep forecasters retrained with TFB's tuned configurations (three seeds), SearchCast has the lowest average MSE on 8 of 9 datasets, with a whole search of 0.2–9 minutes per dataset (Q2).
  • On PEMS08, the dataset with strong cross-channel dependency, SearchCast (0.184) is ahead of all five deep forecasters, including the channel-dependent iTransformer (0.297) (kTJY W3&Q3).
  • SearchCast is ahead of Auto-AR and AutoARIMA on all nine datasets (kTJY W1&Q1). Against zero-shot TimesFM 3.0, it is ahead on six of the seven datasets outside TimesFM's pretraining corpus and level on the seventh (Q2).
  • The ablation locates the gain: per-horizon lookback search carries most of it, and series grouping carries Weather, where it selects one model per channel (xqY4 Q2).

Q1. What is new in the revision?

All new methods use the Table 1 splits, normalization and metrics.

  • New baselines: a foundation model (TimesFM 3.0, zero-shot), Auto-AR, AutoARIMA, iTransformer and TimesNet.
  • New dataset: PEMS08, 170 traffic sensors with strong cross-channel dependence.
  • New analyses: a component ablation, PatchTST with our trailing-window normalization, matched-budget searches for PatchTST (extending Appendix G), a trial-budget sweep down to one trial (extending Appendix F).

Q2. How does SearchCast compare with the added baselines?

It has the lowest MSE of the six trained forecasters on 8 of 9 datasets. SearchCast fits closed-form linear models, and its whole search takes 0.2–9 minutes per dataset on one GPU. Average MSE over the four horizons, mean of three seeds, except for TimesFM which is a single deterministic zero-shot run:

DatasetSearchCastPatchTSTiTransformerTimeMixerTimesNetPatchMLPTimesFM 3.0
ETTh10.4040.4270.4410.4370.4690.4510.428
ETTh20.3340.3530.3760.3590.4030.3880.347
ETTm10.3470.3490.3630.3560.4080.3760.361
ETTm20.2390.2580.2680.2590.2970.2780.259
Exchange0.3470.3680.3600.3930.4020.3870.388
Weather0.2150.2240.2330.2270.2570.2300.230
Electricity0.1600.1700.1640.1860.1950.1730.159
Traffic0.4070.3980.3970.4120.6240.445N/A*
PEMS080.1840.3300.2970.3420.3550.329N/A*

* The dataset is in TimesFM 3.0's pretraining corpus. For reference, TimesFM 3.0 reaches 0.348 on Traffic and 0.167 on PEMS08.

  • Compared with trained deep forecasters: SearchCast has the lowest MSE of the six methods on 8 of the 9 datasets. On PEMS08 it is ahead of the best deep forecaster by 0.113 (0.184 vs. 0.297). On Traffic it is within 0.010 of the best (0.407 vs. 0.397).
  • Compared with TimesFM 3.0 (zero-shot): SearchCast is ahead on six of the seven datasets outside TimesFM's pretraining corpus and level on Electricity (0.1596 vs. 0.1595). Traffic and PEMS08 are inside that corpus. SearchCast reaches this from each dataset's training split alone, with no pretraining data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.