acceptodds
Under review as a conference paper at ICLR 2027

GateCorrect: Recursive Self-Improvement for Frozen Time-Series Forecasters

Abstract

Even after a time-series forecaster is frozen, its errors can contain patterns predictable from the input history. The unresolved problem is sequential: accepting one correction changes the residual that supplies supervision for the next, so a basis diagnosed for the earlier target may no longer describe the useful correction directions. GateCorrect formulates this process as recursive self-improvement (RSI): each round re-diagnoses the current residual, allocates a new correction unit, and admits it only when validation supports the update. A directional path uses a low-dimensional, predictability-screened output span for uncapped amplitudes; a bounded, level-aware path fits the remaining error while retaining window-level information. Across three forecasters and seven datasets, one unit reduces test MSE by 3.74% over 252 runs. In a separate 216-run matched comparison, RSI raises mean reduction from 2.18% with one unit to 2.81%; later units improve 32 of the 37 runs that accept them. GateCorrect therefore makes residual diagnosis and depth selection part of the learning method, rather than treating successive residual fitting as a fixed stack.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Post-Hoc Correction, Spectral Analysis, Model-Agnostic Adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_u8AQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a post-hoc correction framework motivated by the observation that most existing approaches iteratively use the same weak learners. The authors argue that correction units should dynamically change during recursion steps to adapt to the new residual output space. In essence, they contrast traditional boosting which repeatedly corrects errors using a fixed pattern with an adaptive self-correction mechanism that changes the output family across recursion steps. To achieve this, at each recursion step, the method trains a ridge probe to predict the residual from the historical context, generates a corrected prediction, and performs Singular Value Decomposition (SVD) on it to identify candidate basis functions for the correction. A supplementary statistical test is then applied to these candidates to select a valid subset. Once the basis is chosen, the correction is applied. At test time, all bases are fixed, and corrections are executed using the units saved during training. Overall, the method shows improvements compared to the base forecasters and outperforms δ\delta-adapters, another post-hoc correction baseline.

Critical strengths

(i) The idea of selecting different correction units during recursion is conceptually sound. (ii)The paper provides several thorough ablations on different components of the methodology.

Critical weaknesses

(i)The paper is dense and difficult to read, often introducing technical details without offering sufficient intuition. For instance, it lacks a clear justification for why Benjamini-Hochberg FDR control is chosen over alternative protocols. Furthermore, the text frequently assumes advanced notions without explanation, while simultaneously providing excessive detail for certain experiments, which can confuse the reader. (ii)The approach relies heavily on a validation set, which is critical for performance and for preventing overfitting. It remains unclear how the method behaves if the available validation data has already been used for hyperparameter selection or early stopping. Moreover, in practical applications, practitioners may only have a large training split without a dedicated validation set; the paper does not analyze the impact of taking out validation data for post-hoc correction. (iii)The method lacks comprehensive comparisons with alternative post-hoc correction techniques. Given that the core motivation is to move beyond fixed base learners, a direct comparison against standard boosting using various fixed families of learners should be included. Beyond boosting, several other post-hoc approaches exist in the time-series literature that deserve baseline comparisons.

Questions for authors

1- Why to use Benjamin-Hochberg FDR control instead of other protocols? It would be worthy to state every protocol that could be used for performing the candidate selection and explain this statistical test is the one chosen as well as ablations on alternatives one

2- There are several related works omitted in post hoc correction of time series (including to calibration approaches, context augmentation approaches, finetuning ones) as well as comparison with the proposed approach with these methods. How does the method compare with applying boosting with different correctors family ? See for example Post-Training in Time Series Foundation Models: A Unifying Framework

3- Can you discuss the choice of validation set, its critical impact, how to choose them when giving a big training set and what are the impact of using part of training for the validation which may make the training less efficient. What is the tradeoff. If we take more samples for validation the training is obviously of less quality ... ?

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

The paper proposes a residual correction framework for frozen time-series forecasters (GateCorrect). Specifically, the residual is corrected through ridge regression + SVD to identify the predictable direction of residuals. Amplitudes are fit in a certain subspace for correction. The remaining errors are corrected through a bounded level-aware network. Recursively (RSI), residual is updated repeatedly, subspace is re-estimated, new correction units are allowed by validation.

Critical strengths

(1) Effective interpretable residual correction. 2-path design: unrestricted correction power for predictable output directions. Bounded correction for the rest. Large single unit evaluation. Significant MSE improvement under many forecasters and datasets.

  1. Careful analysis of correction reliability.

Show that direction screening, validation admission are effective for lowering harmful corrections via ablations. Support these via randomized-target controls. The authors explicitly look at evolution of residual subspaces through the recursions.

Critical weaknesses
  1. Weak evidence for core RSI contribution.

From 240 run mech. study: one unit cor. can improve MSE by 3.08% vs 2.79% (fixed-basis recursion) and 2.85% (re-diag recursion). Difference for re-diag over fixed basis is marginal (0.066%). Harmful runs for both recur. increase from 41 to 59.

We can clearly see subspaces are different, but reconfiguration gains little predictive. Best case is in another 216 run expt. w/ different admission protocol (no equivalent fixed basis case). Hard to be sure it's due to re-diagnosis.

Also, need stronger expt. evidence for conceptual differences to usual boosting/stagewise residual learning.

  1. Baseline similarity and efficiency poor.

For primary δ-Adapter comparison, take out input-correction part. (Smaller 13 cell comparison, full δ-Adapter wins standalone GateCorrect for 11 cells.)

Lack a baseline for stagewise residual learning that has similar compute. And no/weak assessment of extra training/parameter/inference cost for recursive correction. Plus Electricity/Traffic gain dominates with recursion.

Questions for authors

Q1) Can you also compare one-unit fixed-basis re-diag along with your primary (same primary: 216 runs chronological A/B, pair wise CI) for one-unit, fixed-basis and re-diag to establish the independent value of re-diagnosis? This is crucial.

Q2) How does GateCorrect differ from traditional stagewise residual learning, under the same network capacity/recursion/validation?

Q3) Sensitivity of temporal dependence on your recursive admit decision, what extra training time/params/inference latency for each unit admit?

Reviewer_AFw2OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper proposes GateCorrect, a recursive residual correction framework for frozen time-series forecasters. It uses Ridge regression and SVD to identify predictable residual directions, followed by an unconstrained directional correction and a bounded level-aware correction for the remaining errors. The proposed recursive self-improvement (RSI) mechanism re-diagnoses the residual after each accepted correction, while a validation gate controls the recursion depth. Experiments on three forecasting models and seven datasets report an average MSE reduction of 3.74% with a single correction unit. In a separate matched evaluation, RSI improves the average reduction from 2.18% to 2.81%.

Critical strengths
  1. Well-motivated correction design. Separating predictable residual directions from the remaining errors provides a reasonable way to allocate correction capacity. The combination of unconstrained low-dimensional correction and bounded full-dimensional correction is intuitive and supported by the single-unit ablations.

  2. Useful experimental analysis. The evaluation covers multiple forecasting models and datasets. The ablations on direction screening, validation gating, and subspace re-diagnosis provide evidence beyond aggregate performance.

Critical weaknesses
  1. The benefit of recursive re-diagnosis is not convincing. Table 13 shows only a 0.066 percentage-point improvement over fixed-basis recursion. More importantly, both recursive variants perform worse than the single-unit method (2.85% and 2.79% vs. 3.08%). Although the selected subspaces change across iterations, the results do not establish that re-diagnosis consistently improves forecasting accuracy.

  2. The contribution of re-diagnosis is not sufficiently isolated. Table 2 reports an additional 0.63 percentage-point gain from RSI, but this may result from adding more correction units rather than rebuilding the output subspace. The fixed-basis comparison uses a different admission protocol, making direct attribution difficult.

  3. The recursive gains are concentrated in a few datasets. Among the 216 matched runs, only 37 accept additional correction units, and 30 of the 32 improved runs come from Electricity and Traffic. Recursion provides little additional benefit elsewhere, including only a 0.07 percentage-point average gain for iTransformer. The paper does not sufficiently explain when recursive correction is useful or why its benefits vary so substantially.

  4. Important baselines and practical comparisons are limited. The Ridge probe already predicts residuals, but a simple Ridge-only correction baseline is missing. Moreover, the main δ-Adapter comparison excludes input adaptation, while the full-method comparison covers only 13 cells.

Questions for authors
  1. Recursive re-diagnosis: Is the 0.066 percentage-point improvement in Table 13 statistically significant?

  2. Recursive performance: Why do both recursive variants underperform the single-unit method in Table 13, while RSI improves performance in Table 2? How much of this difference can be attributed to the admission protocol?

  3. Simple baselines: How would a direct Ridge-based residual correction perform under the same validation gate? What additional benefit do the correction networks provide?

  4. Applicability: Why are most recursive improvements concentrated in Electricity and Traffic? Can the proposed predictability diagnostics reliably identify when further recursion will be useful?

Comments on the paper

1 comment

I appreciate the author's effort. Therefore, I am raising my score to 2. However, the following issues remain:

  1. Recursive gains: In the vast majority of cases, these do not yield additional benefits. Actual gains remain concentrated within specific datasets and backbones—most notably, Electricity and Traffic. The effectiveness of RSI remains highly conditional.

  2. Statistical evidence: The primary recursive experiments cover only 216 out of 252 runs, failing to provide complete coverage. Furthermore, the authors explicitly state that their conclusions apply only to a fixed benchmark panel, which is insufficient to support broader generalizations.

  3. Baseline comparison: The comparison with Direct Ridge is limited to the ETTh1, Weather, and Electricity datasets, and is conducted solely under the H96 setting. Even on these datasets, it is not demonstrated that GateCorrect consistently outperforms Direct Ridge. Additionally, the comparison with Full δ-Adapter covers only 13 cells.

  4. Computational cost: For DLinear/Electricity, the original backbone contains only 0.0186M parameters, whereas a single correction unit adds 79.45M parameters. This leads to a substantial increase in both the number of parameters and deployment costs—a concern that is particularly significant given that recursive gains are not universally observed.

  5. Presentation and readability: I concur with reviewer u8AQ’s assessment regarding the paper’s readability: "Furthermore, the text frequently assumes advanced notions without explanation, while simultaneously providing excessive detail for certain experiments, which can confuse the reader." Some advanced concepts in the paper lack necessary intuitive explanations, while certain experimental details are overly verbose. Moreover, the organization of information in Figure 1 and Figure 3 lacks clarity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.