acceptodds
Under review as a conference paper at ICLR 2027

Bounded-Abstention Multi-horizon Time-series Forecasting

Abstract

Multi-horizon time-series forecasting involves making predictions for a consecutive sequence of subsequent time steps. This task arises in many application domains, such as healthcare and finance, where mispredictions can have a high cost and reduce trust. The learning with abstention framework tackles these problems by allowing a model to abstain from offering a prediction when it is at an elevated risk of making a misprediction. Unfortunately, existing abstention strategies are ill-suited for the multi-horizon setting: they target problems where a model offers a single prediction for each instance. Hence, they ignore the structured nature of the predictions offered by a multi-horizon forecaster. We fill this gap by formalizing the problem of learning with abstention for the multi-horizon forecasting setting and show that its structured nature admits a richer set of abstention problems. Concretely, we propose three natural notions of how a model could abstain for multi-horizon forecasting. We theoretically analyze each problem to derive the optimal abstention strategy and propose an algorithm that implements it. Extensive evaluation on 24 datasets shows that our proposed algorithms significantly outperform existing baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Bounded-Abstention, Selective Prediction, Learning to Reject, Multi-horizon time-series forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_u8beOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of Bounded-Abstention Multi-Horizon Forecasting, which aims to minimize prediction risk for accepted time steps while satisfying a minimum coverage constraint. Given the sequential temporal structure inherent in multi-horizon forecasting, the authors propose three structured abstention strategies: Full Abstention (FAbFor), Partial Abstention (PAbFor), and Interval Abstention (IntAbFor). The paper establishes a unified optimization framework for these strategies and derives optimal selection functions based on conditional prediction risk. In terms of implementation, the authors employ joint training of a forecaster and a risk estimator, utilizing a calibration set to determine the abstention parameters. Experiments conducted across 24 datasets demonstrate that the proposed methods outperform several baselines in terms of selective prediction risk.

Critical strengths

S1. Existing selective prediction methods primarily focus on single-step or independent predictions, making them ill-suited for multi-horizon forecasting, which requires continuous prediction intervals. To address this, the paper proposes a structured abstention mechanism—an approach with sound practical motivation and significant research value.

S2. The authors unify full, partial, and interval-based abstention into a bounded-abstention optimization framework and analyze optimal selection strategies under various structural constraints. These three strategies follow a clear, progressive relationship, which sheds light on how the flexibility of the selection process influences prediction risk.

Critical weaknesses

W1. The mathematical definition of the randomized calibration mechanism lacks sufficient rigor. In Eq. (9), if the empirical coverage function exhibits discontinuities (jumps), the two defined thresholds may coincide, rendering it impossible to achieve the target coverage via the randomization strategy described in Eq. (10). While a binary search algorithm might circumvent this issue in practice, the mathematical definition in the paper requires further clarification. The relationship between conditional risk and conditional variance requires further elaboration. Section 3 equates the conditional prediction risk (under MSE) to the conditional variance; however, this equality generally holds only if the predictor is equal to the true conditional mean. Although Gaussian β-NLL can be used to estimate the conditional squared prediction error, the authors should explicitly clarify the relationship between their estimation objective and the theoretical conditional risk.

W2. The fairness of comparisons with external baselines warrants further validation. While the authors compare various abstention strategies and risk estimation methods, the external baselines employ different training objectives, which may introduce discrepancies in base predictive performance. Furthermore, risk comparisons based on a fixed target coverage may be confounded by deviations in the actual test-time coverage. Conducting experiments that utilize a fixed predictor while matching the actual achieved coverage would significantly strengthen the credibility of the conclusions.

Questions for authors

Q1. When the empirical coverage function exhibits a jump—causing the two thresholds in Eq. (9) to coincide—how does the mechanism in Eq. (10) ensure that the target coverage is met?

Q2. Could the authors clarify whether the Gaussian β-NLL estimates the conditional variance of the data or the conditional expected squared error relative to the current predictor? How do these relate to the theoretical conditional risk?

Q3. Could the authors clarify whether the external baselines utilize the same point predictions? Additionally, could they provide risk comparisons based on the actual test-time coverage to mitigate the influence of differences in predictors and deviations in coverage on the experimental conclusions?

5 comments

We are glad that the Reviewer highlighted the sound practical motivation and significant research value of our paper. We address the reviewer’s comments in the following.

Jumps in empirical coverage. We thank the reviewer for spotting this imprecision. In practice, the thresholds are obtained via binary search (Algorithms 1-2, Appendix D), which keeps ϕ^(g^γ^ℓ)≤cH≤ϕ^(g^γ^r)\hat{\phi}(\hat{g}_{\hat{\gamma}_\ell}) \le cH \le \hat{\phi}(\hat{g}_{\hat{\gamma}_r}), so the randomization always reaches the target coverage. We have revised Eq. (9) to define γ^ℓ,γ^r\hat{\gamma}_\ell, \hat{\gamma}_r as the bracketing thresholds returned by the binary search. The method and results are unchanged.

Conditional variance vs conditional expected squared error. We thank the reviewer for raising this point. The second head estimates the conditional expected squared error of the current predictor, not the conditional variance of the data. This is exactly the conditional risk ρt\rho_t required by Theorems 1-3:

=\mathbb E\bigl[(Y_t-\hat f_t(y_{1:T}))^2\mid Y_{1:T}=y_{1:T}\bigr] =\operatorname{Var}(Y_t\mid Y_{1:T}=y_{1:T}) +\bigl(\mathbb E[Y_t\mid Y_{1:T}=y_{1:T}]-\hat f_t(y_{1:T})\bigr)^2,$$ which reduces to the conditional variance only when $\hat{f}_t$ equals the true conditional mean. For a fixed $\hat{f}$, the expected gradient of Eq. (8) with respect to $\hat{\sigma}^2_t$ vanishes exactly at $\hat{\sigma}^2_t = \rho_t$; the stop-gradient weight only rescales it, and no Gaussian assumption is needed. We have revised Section 3 to better describe the second head as a conditional-risk estimator. **Experiments with the same point-wise predictors.** We thank the reviewer for the question. In section 4.2, the baselines and our methods do not use the same point predictions. Following the reviewer's suggestion, we ran an additional ablation in which all baselines use the same point predictor as FAbFor, and the methods differ only in the scores used for abstention. We report the results in the new Appendix F.11. Even in this setting, FAbFor outperforms the baselines, reducing the selective risk on average by 10% vs AdaptiveCF, by 13% vs MQ-RNN, and more than 20% vs DP-RNN. This confirms the effectiveness of our abstention strategy when all methods use the same forecaster.

Experiments with fixed test-time coverage. We calibrate the thresholds on a held-out calibration set as is standard in selective prediction [1,2]. To fully remove the effect of coverage deviations, we repeat Q1, setting each method’s rejection threshold directly on the test set. Thus, all methods have the same test coverage. While this is not a deployable procedure, it isolates how well each method ranks the series by risk. Across the six target coverages, FAbFor outperforms the competitors in 21/24 datasets and reduces the selective risk by 14% vs AdaptiveCF and MQ-RNN, and 18% vs DP-RNN. The gains are essentially unchanged from the Q1 results (i.e., 14%, 14%, and 19% in Q1, respectively). We report below each dataset’s selective risk for c=0.7c = 0.7.

DatasetFAbForAdaptiveCFDP-RNNMQ-RNN
chlorine.0048 ± .0003.0057 ± .0005.0066 ± .0006.0056 ± .0006
ERA5.0045 ± .0008.0065 ± .0009.0135 ± .0016.0066 ± .0013
handoutlines.0045 ± .0023.0054 ± .0022.0070 ± .0028.0124 ± .0114
inlineskate.0031 ± .0006.0036 ± .0020.0035 ± .0004.0049 ± .0019
freezer.0026 ± .0027.0117 ± .0164.0294 ± .0136.0038 ± .0038
wafer.0002 ± .0001.0027 ± .0047.0099 ± .0037.0009 ± .0005
ecg.0591 ± .0046.0833 ± .0126.1307 ± .0165.0794 ± .0102
ITpower.0715 ± .0154.0942 ± .0217.0880 ± .0147.0918 ± .0189
melbourne.0333 ± .0078.0857 ± .0829.0752 ± .0225.0640 ± .0270
mixedshapes.0288 ± .0035.0400 ± .0062.0534 ± .0058.0423 ± .0101
symbols.0259 ± .0135.0363 ± .0109.0467 ± .0148.0452 ± .0119
yoga.0192 ± .0040.0221 ± .0021.0324 ± .0057.0244 ± .0049
covid.2782 ± .1865.2965 ± .0945.2520 ± .0564.3934 ± .1361
eeg.1530 ± .0045.1702 ± .0188.2200 ± .0255.1851 ± .0215
fiftywords.2765 ± .1283.4074 ± .0705.3332 ± .0917.4241 ± .1259
wordrecognition.1426 ± .0239.1552 ± .0238.2017 ± .0179.1656 ± .0273
middlephalanx.1167 ± .0247.1409 ± .0321.1677 ± .0237.1613 ± .0321
swedishleaf.1636 ± .0469.1552 ± .0303.2529 ± .0342.1514 ± .0212
faceall.3987 ± .0436.4782 ± .0669.4985 ± .0248.4732 ± .0521
fetalecg.3933 ± .0906.4743 ± .1128.4918 ± .0941.4945 ± .1245
ford.4843 ± .0162.4895 ± .0234.5757 ± .0159.4738 ± .0335
refrigeration.5492 ± .0609.7506 ± .1131.8311 ± .0901.7299 ± .1092
sonyaiborobot.4136 ± .0351.4568 ± .0352.4860 ± .0621.4648 ± .0343
temperature.6326 ± .1106.7573 ± .2946.9299 ± .4246.8066 ± .3589

[1] Hendrickx, Kilian, et al. "Machine learning with a reject option: A survey." Machine Learning 113.5 (2024): 3073-3110.

[2] Ruggieri, Salvatore, and Andrea Pugnana. "Things machine learning models know that they don’t know." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 27. 2025.

Dear reviewer u8be,

Thanks again for the feedback. As the author-discussion phase comes to an end, please let us know whether we have addressed your concerns or if there are any remaining points we can clarify.

Reviewer_UaowOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies selective multi-horizon forecasting under a minimum coverage constraint. It formalizes three abstention settings: accepting or rejecting the entire horizon, selecting a prefix, and selecting a contiguous interval. The authors derive optimal selection rules assuming known conditional risks and implement them using learned risk estimates and calibration. Experiments on various datasets, with additional backbone and loss-function analyses, report reductions in selective risk.

Critical strengths

++ Requiring accepted predictions to form a contiguous sequence addresses a meaningful limitation of independent, step-wise rejection. The unified formulation and corresponding theoretical characterization provide a clear framework for studying the trade-off between prediction availability and reliability.

Critical weaknesses

-- The algorithms largely combine existing conditional-risk estimation with thresholding, Lagrangian optimization, and calibration. The paper could more clearly identify the nontrivial advances beyond established bounded-abstention principles. Related horizon-design research, such as "The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting. ICML, 2026.", also deserves discussion, while distinguishing its training-target selection from the present inference-time abstention problem.

-- Although 24 datasets are evaluated, 20 originate from the UCR classification archive, including image-derived sequences. Predicting their suffixes provides limited evidence for real-world temporal forecasting. The main-text experiments rely on a small LSTM. Although Transformer results are provided in the appendix, key evidence supporting the framework’s applicability across architectures should be summarized in the main paper.

Questions for authors
  1. Which theoretical or algorithmic components require substantively new insights beyond existing bounded-abstention methods? How does the framework relate to recent work on horizon selection?

  2. Can the authors report aggregate results separately for genuinely temporal datasets and clarify whether any experiments evaluate generalization across successive forecast origins, rather than only across held-out series?

3 comments

We are glad that the Reviewer appreciates that our paper addresses a meaningful limitation, and our framework is clear. We address the reviewer’s comments in the following.

Advances beyond established bounded-abstention principles. Our key advance is three optimal abstention rules, one per setting, that choose which contiguous part of the forecast horizon to predict, whereas existing bounded-abstention methods make a binary accept/reject decision on a single prediction by simply thresholding its conditional risk. This is non-trivial because the steps cannot be thresholded one at a time: whether a step is accepted depends on the other steps in the interval. Finally, we provide algorithms to learn these rules from data and calibrate them to meet the target coverage.

Relation to horizon-selection paper. We thank the reviewer for the pointer. We have added a discussion of horizon-design research, including The Label Horizon Paradox (ICML 2026), to the Related Work (Section 5). We consider this approach complementary to ours. Indeed, our optimal selection functions are derived for any given forecaster - including one trained with an optimized label horizon - provided that a corresponding conditional risk estimator is fitted (e.g., post hoc using the forecaster’s squared residuals).

Key experimental evidence in the main paper. In the revised version, we have discussed the Transformer results in Section 4.2. Moreover, Figure 3 now shows the CD plots for both backbones. With the Transformer, FAbFor ranks first at every target coverage and is significantly better than DP-RNN and AdaptiveCF across all coverages. At c=0.70c=0.70, FAbFor is also significantly better than MQ-RNN.

Other real-world temporal forecasting datasets. Following Reviewer iFK2’s suggestion, we added M4, Electricity, and Traffic to our experiments. We report the results in Appendix F.10. Across these datasets, FAbFor reduces the selective risk by 10% on average against the best competitor, AdaptiveCF, and always achieves the lowest risk. PAbFor and IntAbFor reduce the risk further on 6 of the eight datasets. In the next version of the paper, we will integrate these datasets into the main evaluation (Q1–Q3) instead of reporting them in the appendix, and we will run on them all the ablations of Appendix F (e.g., Transformer backbone, MAE as risk measure).

Temporal datasets. The tables below report the results on the 14 genuinely temporal datasets. The included datasets cover health and physiological signals (covid, eeg, ecg, fetalecg), weather (ERA5, temperature), energy and count data (ITpower, freezer, refrigeration, melbourne), industrial sensors (ford, wafer) and motion recordings (sonyaiborobot, inlineskate). The results closely mirror those on all datasets. In full abstention, FAbFor has the best average rank at every coverage (1.17 on average, vs. 1.21 on all datasets) and its relative improvements are of similar magnitude it reduces the selective risk by 18% vs. DP-RNN, 14% vs. MQ-RNN and 13% vs. AdaptiveCF, compared to 19%, 14%, and 14% on all datasets. PAbFor and IntAbFor outrank FAbFor at every coverage, and Accept-cH ranks last.

Full abstention: average relative reduction in selective risk of FAbFor over the baselines across the six target coverages on the temporal datasets.

Datasetvs AdaptiveCFvs DP-RNNvs MQ-RNN
covid3.0%−3.2%25.7%
eeg13.9%27.3%14.3%
ERA516.2%45.2%19.2%
temperature11.3%22.1%14.3%
ecg23.2%41.9%19.7%
fetalecg18.8%19.1%19.4%
ford2.1%10.2%−0.4%
freezer45.0%74.3%26.4%
inlineskate9.8%7.5%29.8%
ITpower18.0%9.7%15.0%
melbourne36.8%17.7%18.6%
refrigeration20.7%24.7%19.3%
sonyaiborobot6.2%6.0%5.2%
wafer74.0%92.8%41.1%
Average21.4%28.2%19.1%

Full abstention: average rank (± std) across the temporal datasets (lower is better).

cFAbForAdaptiveCFDP-RNNMQ-RNN
.701.07 ± 0.262.68 ± 0.593.46 ± 0.772.79 ± 0.94
.751.14 ± 0.352.86 ± 0.643.36 ± 0.972.64 ± 0.89
.801.14 ± 0.352.86 ± 0.643.21 ± 1.012.79 ± 0.94
.851.18 ± 0.363.07 ± 0.703.04 ± 1.142.71 ± 0.80
.901.21 ± 0.563.14 ± 0.643.14 ± 1.122.50 ± 0.73
.951.29 ± 0.703.07 ± 0.703.14 ± 1.122.50 ± 0.73

Abstention strategies: average rank (± std) across the temporal datasets (lower is better).

cFAbForPAbForIntAbForAccept-cH
.702.79 ± 0.671.86 ± 0.521.43 ± 0.623.93 ± 0.26
.752.71 ± 0.701.86 ± 0.611.50 ± 0.713.93 ± 0.26
.803.00 ± 0.381.61 ± 0.471.46 ± 0.613.93 ± 0.26
.853.00 ± 0.381.71 ± 0.561.36 ± 0.443.93 ± 0.26
.902.82 ± 0.701.96 ± 0.551.32 ± 0.593.89 ± 0.28
.952.57 ± 0.731.79 ± 0.591.64 ± 0.744.00 ± 0.00

Held-out series vs. successive forecast origins. We evaluate the performance of the methods using held-out series, consistent with our stated setting (see lines 85-86).

Dear reviewer Uaow,

Thanks again for the feedback. As the author-discussion phase comes to an end, please let us know whether we have addressed your concerns or if there are any remaining points we can clarify.

Reviewer_iFK2OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper extends the abstention framework of Franc et. al (2023) to the case of multi-horizon forecasting. It covers three cases: full-abstention over the whole horizon, partial-abstention, where the model predicts until some step ee , and interval abstention, where it may abstain outside a subinterval (s,e)(s,e). The authors analyze the optimal selection functions in these three settings, which depend on the conditional risks. Then, they present learning algorithms for each type of selective forecaster, and empirically compare them based on selective risk and constraint satisfaction.

Critical strengths
  1. Extending the abstention framework to multistep forecasting is a valuable research contribution. This is a technically solid paper with clear and well-written theory. I did not check the proofs, but the Theorems make intuitive sense: the model should abstain where the cumulative risk over the horizon exceeds quantile of the cumulative risk. In the partial abstention case, the intervals are extended only where the marginal risk is lower than γ∗\gamma^*.
  2. The experimental results are solid and align with the theory. FAbFor exhibits superior selective risk over baselines in a variety of datasets. As expected,IntAbFor exhibits high constraint satisfaction, and its selective risk is as good or better than PAbFor.
Critical weaknesses
  1. The results focus on LSTMs and old RNN baselines. Furthermore, many of the common multi-horizon time series forecasting datasets are missing (e.g. see TSLib). How effective are these abstention methods on modern long-horizon models?
  2. I would have liked to see a more detailed discussion on the statements of Theorems 1,2,3.

Particularly if weakness 1 were addressed I would be open to increasing my score.

Questions for authors
  1. It seems the partial abstention methods are as good as or better than full abstention. Why not plot all methods (or the best of the three) in Figure 2?
  2. Do any of the baselines utilize stepwise abstention? If not, I am curious about how this compares to the multi-horizon abstention methods.

Clarification: 3. Why not present the transformer results in the main body?

5 comments

We are glad that the Reviewer found our theory clear and well-written, and that the results are solid. We address the reviewer’s comments in the following.

TSLib datasets. We thank the reviewer for pointing us to TSLib. We added M4, Electricity, and Traffic, the TSLib datasets with enough distinct series for our setting. For M4, we treat each frequency as a separate dataset with its standard horizon H and use the last T = 2H steps before that as input [1]. For electricity and traffic, we adapt them to our setting by treating each channel as a distinct series and taking its last 288 observations, with H = 96 and T = 2H.

We report the results in Appendix F.10. Across the eight datasets, FAbFor reduces the selective risk by 10% on average against the best competitor, AdaptiveCF, and it always achieves the lowest risk. PAbFor and IntAbFor reduce the risk further on 6 of the eight datasets. In the next version of the paper, we will integrate these datasets into the main evaluation (Q1–Q3) instead of reporting them in the appendix, and we will run on them all the ablations of Appendix F (e.g., Transformer backbone, MAE as risk measure).

Abstention methods on modern long-horizon models. Our framework is agnostic to the choice of forecaster and conditional risk estimator: the selection rules only require an estimate of each step's conditional risk. This estimate can come from (i)(i) a separate regression model fit on the forecaster's residuals, or (ii)(ii), as in the paper, an additional output head trained end-to-end. The quality of the abstention rules depends on the quality of this estimate (Theorems 1-3), not on the backbone itself. To test a modern forecaster directly, we tested option (i)(i), adding an experiment with a frozen pretrained foundation model, Chronos-2 [2]. We use Chronos-2 zero-shot: its median forecast is the point forecast on which all methods abstain. We fit every risk estimator on the training split and calibrate the abstention thresholds on the calibration split. Everything else follows the same setup as in Section 4. FAbFor, PAbFor, and IntAbFor use the same network as in the paper with the mean fixed to the Chronos-2 forecast, so only the risk head is trained. AdaptiveCF and MQ-RNN are fit on the Chronos-2 residuals, while DP-Chronos is obtained by applying MC dropout through Chronos-2 itself and uses the standard deviation of the median forecast over 100 forward passes as its score. Finally, we include Chronos-PI, which abstains based on the width of Chronos-2’s own 90% predictive interval. We report the results in Appendix F.12. Even in this setting, FAbFor outperforms the baselines: its selective risk is lower than that of AdaptiveCF in 84% of the cases, MQ-RNN in 84%, and DP-Chronos in 90%. Moreover, FAbFor reduces the selective risk by 17% on average compared to Chronos-2's own predictive intervals. Moreover, PAbFor and IntAbFor further reduce the selective risk in all datasets compared to FAbFor. This confirms the effectiveness of our abstention strategy when all methods use the same frozen forecaster.

Clarity and presentation. We thank the reviewer for the suggestions. In the revised version, we have expanded the discussion following each theorem to explain the intuition behind the result. Moreover, we have added the discussion for the Transformer results to the main paper. Figure 3 now reports the CD plots for both backbones, and Section 4.2 discusses the Transformer results for Q1 and Q2.

Comparison with step-wise abstention. None of the baselines uses step-wise abstention. We did not include such a strategy because abstaining on individual steps produces fragmented forecasts, which do not fit the purpose of our work. Moreover, note that if we allow abstaining on non-contiguous points, such an abstention strategy should generally perform better than other methods.

Still, we tested a step-wise abstention strategy that treats each step’s estimated risk separately and abstains on all steps whose risk exceeds a single threshold, calibrated as the c-th quantile of all step-level risks in the calibration set. Since step-wise abstention has the highest flexibility, it achieves the lowest selective risk in 13 out of 24 datasets. However, the differences with IntAbFor are small and mostly within one standard deviation. This suggests that requiring contiguous forecasts comes at a limited cost in terms of selective risk. We will include these results in the paper if the reviewer finds them a valuable addition. We report below the results for c=0.7c=0.7.

DatasetFAbForPAbForIntAbForStep-wise
chlorine.0048 ± .0003.0038 ± .0002.0034 ± .0003.0023 ± .0002
ERA5.0044 ± .0006.0043 ± .0005.0043 ± .0006.0044 ± .0007
handoutlines.0045 ± .0021.0041 ± .0020.0044 ± .0023.0044 ± .0022
inlineskate.0031 ± .0005.0028 ± .0004.0028 ± .0004.0027 ± .0004
freezer.0026 ± .0026.0023 ± .0024.0057 ± .0053.0067 ± .0047
wafer.0002 ± .0001.0003 ± .0003.0003 ± .0005.0001 ± .0000
ecg.0593 ± .0043.0423 ± .0043.0419 ± .0044.0420 ± .0041
ITpower.0669 ± .0104.0689 ± .0081.0639 ± .0065.0617 ± .0044
melbourne.0336 ± .0074.0307 ± .0061.0305 ± .0058.0299 ± .0090
mixedshapes.0291 ± .0033.0198 ± .0032.0198 ± .0026.0198 ± .0032
symbols.0263 ± .0133.0233 ± .0105.0233 ± .0105.0229 ± .0101
yoga.0194 ± .0039.0144 ± .0034.0145 ± .0034.0144 ± .0033
covid.2950 ± .2312.2927 ± .2304.1974 ± .1894.1559 ± .1940
eeg.1445 ± .0044.1444 ± .0035.1444 ± .0039.1442 ± .0039
fiftywords.2781 ± .1311.1865 ± .0623.2039 ± .0720.1866 ± .0618
wordrecognition.1366 ± .0266.1354 ± .0248.1448 ± .0233.1464 ± .0239
middlephalanx.1180 ± .0320.1112 ± .0232.1107 ± .0211.1028 ± .0201
swedishleaf.1735 ± .0432.1694 ± .0393.1465 ± .0232.1480 ± .0238
faceall.3953 ± .0438.3868 ± .0438.3855 ± .0416.3911 ± .0399
fetalecg.3919 ± .0863.3261 ± .0576.3262 ± .0575.3205 ± .0576
ford.4816 ± .0162.4199 ± .0162.4180 ± .0114.4202 ± .0151
refrigeration.5739 ± .0763.5203 ± .0779.5138 ± .0785.5158 ± .0759
sonyaiborobot.4444 ± .0358.3373 ± .0404.2794 ± .0381.2867 ± .0335
temperature.6102 ± .1080.4934 ± .0958.4919 ± .0950.4866 ± .0916

Figure 2’s methods. We chose to report the results in two separate figures to make the comparisons easier to follow. Figure 2 compares FAbFor with the baselines, which rely on a different criterion for deciding when to abstain, while Figure 4 (in the revised version of the paper) shows how much the additional flexibility of partial and interval abstention reduces the selective risk. We would be happy to include all methods in Figure 2 in the next version of the paper if the reviewer thinks this helps the paper’s readability.

[1] Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. In International Conference on Learning Representations, 2023.

[2] Ansari, Abdul Fatir, et al. "Chronos-2: From univariate to universal forecasting." arXiv preprint arXiv:2510.15821 (2025).

Dear reviewer iFK2,

Thanks again for the feedback. As the author-discussion phase comes to an end, please let us know whether we have addressed your concerns or if there are any remaining points we can clarify.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author-Reviewer Discussion summary

We thank the Area Chair and the reviewers for their time and constructive feedback. We are glad the reviewers find our contribution valuable, our theory technically solid, and our experimental results aligned with the theory. Additionally, we thank reviewers iFK2 and u8be for finding our responses detailed and comprehensive. We synthesize below the reviewers’ main points.

Novelty. We are glad that the reviewers recognize the novelty of our contribution. Reviewer iFK2 states that extending the abstention framework to multi-step forecasting is “a valuable research contribution”. Reviewer u8be finds that our structured abstention mechanism has “sound practical motivation and significant research value”. Reviewer Uaow notes that requiring contiguous accepted predictions “addresses a meaningful limitation of independent, step-wise rejection”. In our response to Reviewer Uaow, we further highlighted the significance of our framework, which accounts for the structured nature of multi-horizon forecasting and leads to optimal decision rules that go beyond standard threshold-based abstention.

Soundness. Reviewer iFK2 describes our work as “a technically solid paper”. We thank Reviewer u8be for pointing out two imprecisions, which we have fixed in the revised version of the paper without changing the method or the results. Specifically, we clarified the definition of the calibration thresholds in Eq. 9, and we corrected the description of the risk estimator, which targets the forecaster’s conditional expected squared error rather than the conditional variance of the data.

Setting and clarifications. We have clarified our setting and the scope of our framework. First, the framework does not depend on the choice of forecaster: Theorems 1-3 give the optimal selection function for any model that outputs H-step point forecasts, so they apply directly to modern long-horizon architectures (Reviewer iFK2). Second, we have added a discussion of horizon-design research, including The Label Horizon Paradox (ICML 2026), to the Related Work. We see this line of work as complementary to ours, since our selection functions apply to any forecaster, including one trained with an optimized label horizon (Reviewer Uaow). Finally, we clarified that our evaluation measures generalization across held-out series, consistent with our stated setting at lines 85-86 (Reviewer Uaow).

Additional experiments. We are glad that Reviewer iFK2 finds our experimental results “solid and aligned with the theory”. We ran the additional experiments the reviewers requested and reported them in the individual responses. For Reviewer iFK2, we added results on the TSLib datasets, results with a frozen pretrained foundation model, and a step-wise abstention baseline. For Reviewer Uaow, we added a separate analysis of the temporal datasets. For Reviewer u8be, we added experiments where all methods share the same forecaster, and comparisons at matched test-time coverage. All new results are consistent with our original findings.

Presentation and clarity. Reviewer iFK2 praises our “clear and well-written theory”, Reviewer Uaow notes that “the unified formulation and corresponding theoretical characterization provide a clear framework”, and Reviewer u8be remarks that the three strategies “follow a clear, progressive relationship”. We have further improved the presentation in two ways. First, as requested by Reviewer iFK2, we have added a discussion after each theorem explaining the intuition behind the optimal selection rule. Second, following Reviewers iFK2 and Uaow, we have moved the discussion of the Transformer results into the main paper, showing that our findings hold across architectures.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.