acceptodds
Under review as a conference paper at ICLR 2027

Learning from Forecasting Errors: Separating Correctable Errors from Predictive Uncertainty in Time Series Forecasting

Abstract

Existing time series forecasting methods typically treat forecasting errors as optimization residuals to be minimized, while rarely exploiting the information contained in forecasting failures themselves. However, forecasting errors differ substantially in correctability: some exhibit stable structures that can be further corrected, some recur under similar historical conditions and can be reused as historical evidence, whereas others lack reliable correction evidence and are better represented as predictive uncertainty. We propose EUD-RM (Error-Uncertainty Decoupling with Residual Memory), an error-centric framework that handles forecasting failures through a reduce-correct-reuse-quantify process. A Channel-Aware Mixer (CAM) first strengthens cross-variable interactions to reduce potential systematic errors. Learnability-Gated Error Correction (LGEC) then identifies prediction positions with reliable correction evidence and applies bounded parametric corrections. Residual-Aware Retrieval Memory (RARM) stores historical forecasting residuals and retrieves consistent past failures under similar input and prediction states as non-parametric correction evidence. Finally, an independent LGEC-U branch models the remaining predictive risk after deterministic corrections and produces calibrated prediction intervals without altering point forecasts. Rather than assuming a strict decomposition into deterministic error and random noise, EUD-RM exploits empirical correctability and historical reproducibility to determine how forecasting failures should be handled. Experiments on long-horizon multivariate forecasting benchmarks demonstrate consistent gains, with EUD-RM achieving leading results on most benchmark settings.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Forecast Error Correction, Residual Memory, Retrieval-Augmented Forecasting, Predictive Uncertainty

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

This paper proposes EUD-RM, an error-centric framework for multivariate time-series forecasting. It progressively reduces systematic errors, applies gated parametric correction, retrieves similar historical residuals for non-parametric correction, and models the remaining predictive uncertainty. Experiments on seven long-horizon benchmarks show strong point-forecasting performance and generally support the proposed component roles.

Critical strengths
  1. The error-centric formulation is well motivated, and the staged design provides a coherent way to treat different types of forecasting failures.

  2. The empirical evaluation is extensive, with strong results on seven benchmarks, detailed ablations, mechanism analyses, computational measurements, and multi-seed stability tests.

Critical weaknesses
  1. The uncertainty evaluation is relatively limited. The main interval comparison is performed in the controlled pre-RARM setting and mainly against a constant-scale baseline, although the complete model uses post-RARM LGEC-U.

  2. The framework introduces substantial architectural and retrieval complexity. RARM can add considerable inference latency and memory cost, while its incremental point-forecasting improvement is relatively small in several ablations.

Questions for authors
  1. Could the authors report prediction-interval results for the complete post-RARM EUD-RM and compare them with stronger probabilistic or conformal forecasting baselines?

  2. Can the authors further quantify the accuracy-efficiency trade-off of RARM, for example by varying memory capacity or retrieval size and reporting forecasting gain versus latency and memory cost?

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

EUD-RM is a framework for multivariate time-series forecasting that combines channel mixing, gated error correction, historical residual retrieval, and uncertainty estimation. It aims to distinguish correctable errors, reusable residuals, and remaining uncertainty. The evaluation covers seven forecasting datasets, module ablations, and uncertainty estimation.

Critical strengths
  1. Combining selective residual correction with historical residual retrieval is a useful idea. Separating point-forecast correction from uncertainty estimation also gives the framework a clear organization.

  2. The evaluation goes beyond benchmark accuracy to include module ablations, uncertainty estimation, and random-seed stability. The competitive forecasting results make the approach worth considering.

Critical weaknesses
  1. The ablations establish that the components help, but do not isolate the benefit of the proposed error-decoupling approach. A simpler residual corrector with comparable capacity could help determine whether the learnability gate is necessary.

  2. The distinction between correctable errors, reusable residuals, and uncertainty is supported mainly by empirical correlations. These observations do not yet establish that the modules identify different error regimes rather than successively reducing the same residuals.

  3. RARM depends on similar historical forecasting states, while correction strength and uncertainty scale rely on validation-based calibration. Their reliability when the residual distribution changes at test time remains unclear.

Questions for authors
  1. How does LGEC compare with a residual correction network of similar capacity without explicit learnability supervision? This comparison would directly affect my assessment of the gate’s contribution.

  2. What evidence shows that LGEC, RARM, and LGEC-U identify different error regimes? For instance, are errors rejected by LGEC consistently corrected by RARM, and does that distinction hold when residual patterns change?

Reviewer_wRFdOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes EUD-RM, a forecasting pipeline that handles errors in stages. A channel-mixing block is added to a multi-scale mixer backbone. A gated head (LGEC) adds a bounded correction where a training-time “learnability” target says the residual is smooth and large. A memory module (RARM) retrieves training residuals from similar states and adds a confidence-weighted average. A separate head (LGEC-U) predicts a scale for calibrated intervals. On seven long-horizon benchmarks it reports the best average MSE on most datasets, with ablations on ETT and several mechanism analyses

Critical strengths
  • The idea of reusing model's own residuals as retrieval evidence is good, even though it is not new
Critical weaknesses

Before getting to the weaknesses, I think the writing needs a lot of work. There are five acronyms, and the same pipeline is described again and again from the abstract to the conclusion. Many sentences say what a module is not rather than what it does. Meanwhile, the things I actually needed, like the learnability target and the LGEC losses, are only in the appendix. Several presentation errors are easily seen in the paper. For example, in figure 1 R and D are used without being defined. I list two major concerns i have for the paper below

  • The separation promised in the title is not really shown. LGEC is trained jointly with the backbone through the forecasting loss, so Y^(0)\hat Y^{(0)} is an intermediate output rather than a forecast whose errors are later corrected. The correction gain in Table 3 is measured against that intermediate. The gate also seems to track error size more than correctability. High-G positions have the largest base MAE (Table 3) and also the largest LGEC-U scale (Table 6), so the positions called most correctable are the ones called most uncertain. RARM stores in-sample training residuals, which are typically smaller than test errors, and the validation α absorbs the gap. What remains is an extra output head, a kNN residual adjustment and a locally adaptive interval, all familiar, with no evidence that they capture different kinds of error.

  • I also have doubts about the reported numbers. The five-seed means in Table 12 are worse than Table 9 on every ETT average, and on ETTm2-720 the gap is 0.345 versus 0.375. With the seed means, EUD-RM is no longer best on ETTm1 (0.338 vs 0.337 for Times2D) or ETTm2 (0.255 vs 0.245). This gap is comparable to the CAM and LGEC effects in Table 2 and larger than RARM’s gain of about 0.002 MSE, which sits within seed noise. The baselines also seem copied from the original papers (iTransformer matches its L = 96 results), while EUD-RM uses L = 512 or 720. That contradicts the same-input-length claim in Appendix F.1.

Questions for authors
  • With smax⁡=0.05s_{\max}=0.05, LGEC can move a prediction by at most 0.05, so its MAE gain should not exceed 0.05. Yet Traffic High-G in Table 3 drops from 0.499 to 0.295. In which space is the bound applied?

  • Which run is reported in Table 9, and why does it differ from the seed means in Table 12?

  • Were the baselines rerun with the same input length?

  • Does the gate still stratify correction gain within bins of base error? How does LGEC-U compare with a per-horizon conformal interval?

1 comment
Reviewer_wRFdOpenReview ↗

reply to authors' rebuttal

I thank the authors for the detailed rebuttal. As i have mentioned in my initial review, i find the paper very complicated with different components described repeatedly. I acknowledge the changes made in the revised revision, but to me, it is unclear. As i have experience working on time-series, it took me a lot of time to understand the descriptions in the paper, and i reckon that is not good writing. Good researcch should allow audience even from other fields to understand the main idea. I do not think the paper achieves it within this current state.

Other than the presentation problem, i find some contradicting facts in the rebuttal of the paper, which eventually leads to me taking a stand with my rejection for the paper. I list some of the main errors here. I do care about the details, and rest assured the reason I list the reasons below is i want the paper quality to get better. My recommendation for the paper at this stage is reject, but i think after other rounds of revision the paper could improve.

  • Other than the original W2, I still find some inconsistencies in terms of the reported results. For ETTh1 at H=96 on the full test set, Tables 10 and 14 report MSE 0.339 (seed std 0.00016). Table 19 evaluates the frozen complete pipeline on all 2,785 test windows (the H=96 test set) and reports 0.3493

  • There could be problems with your independent runs. The five-seed ablation you posted on 17 Nov means equal the single-seed Table 11 values with a fourth digit appended (0.344 and 0.3442, 0.381 and 0.3817). Independent runs do not do this.

  • In your own fixed-head test (Table 13), a constant gate keeps 58% of the LGEC gain and the whole correction changes MSE by 0.34%. The gate is therefore not shown to identify correctable errors.

With all the reasons above, I stand by my decision to reject the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.