acceptodds
Under review as a conference paper at ICLR 2027

PeakTS: Forecasting High-Value Regions via Risk-Conditioned Distribution Correction

Abstract

Time-series forecasting models often achieve strong average performance yet remain unreliable in critical high-value regions, where rare large values are particularly difficult to predict accurately. A key limitation is that existing methods typically lack explicit mechanisms for modeling historical patterns associated with future high-value outcomes. To address this issue, we propose a probability-distribution correction framework, PeakTS, for high-value time-series forecasting. PeakTS predicts an ordered categorical distribution over future value bins and derives risk-sensitive statistics, including the distribution mean and upper quantiles, to adaptively correct deterministic forecasts. We further introduce a Risk-Conditioned Peak Memory module that stores general and peak-precursor prototypes and retrieves relevant patterns from time-, difference-, and frequency-domain views to modulate the predictive distribution. To better learn sparsely represented high-value regions, we employ a tail-aware ordinal objective with frequency-based threshold reweighting and upper-tail weighting. Experiments on multiple forecasting benchmarks demonstrate consistent improvements in high-value forecasting while maintaining competitive overall forecasting performance.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_cghLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PeakTS augments a decomposition-based forecaster with an ordered categorical distribution and general/peak prototype memories queried through time, difference, and frequency representations. A learned gate blends the point forecast with a mixture of the distributional mean and 90th percentile. Training combines L1 forecasting, frequency-weighted cumulative cross-entropy, and consistency losses. Experiments cover nine datasets and four overall forecasting horizons, with high-value evaluation at horizon 96.

Critical strengths
  1. Consistent MAE improvements across diverse datasets. Table 1 reports the best Top-25% MAE on all nine datasets. For example, ETTm1 improves from the best competing 0.395 to 0.378, and CzeLan from 0.310 to 0.294. This is meaningful evidence for usefulness on the evaluated high-value subsets.
  2. Component-level evidence supports the design. Removing correction or memory worsens Top-25% performance on all three ablation datasets (Table 2). Appendix F/Table 5 additionally tests the temporal views; the full combination improves over each reduced-view configuration. These experiments are more informative than aggregate benchmark wins alone.
Critical weaknesses
  1. Comparisons do not sufficiently isolate the contribution. PeakTS explicitly optimizes L1 (Eq. 17), but baseline training losses and matched tuning budgets are not documented in Section 4.1/Appendix C. Thus, the especially consistent MAE gains cannot be cleanly attributed to the proposed mechanisms. The closest ordinal baseline, OCE-TS, has Weather overall MSE 36.91 versus PeakTS's 0.155 (Table 1). The authors acknowledge modified preprocessing to remove leakage (p. 7), but do not specify or validate that modification. This is not evidence of an implementation error, but it leaves a crucial comparison unauditable. No seed variability is reported, including for small gains such as Electricity Top-25% MAE 0.261 to 0.259.
  2. The distributional correction's tail behavior is insufficiently specified. Figure 1 standardizes using historical mean/std, while Section 3.1 initializes bin support near [-1,1], without explaining boundary adaptation or out-of-support targets. For normalized values, if y_true >= y_raw > the upper bin edge, the bounded mean/quantile target pulls the forecast downward whenever the gate is positive. Proposition 3.1 assumes appropriately directed, non-overshooting correction; it does not establish how frequently that condition holds. Boundary-case evidence is therefore important. More generally, Top-25% evaluation at horizon 96 supports a narrower claim than reliable rare-peak forecasting. Its MSE gains are also not universal: Weather is 0.311 versus the best baseline's 0.283, and ZafNoo 0.646 versus 0.614.
Questions for authors
  1. Do the gains persist against L1-trained, equally tuned DLinear and a verified input-only-normalized OCE-TS? Please specify the OCE-TS preprocessing changes and report variability across seeds.
  2. How are bin boundaries updated and out-of-support targets handled? Please give the actual normalization, smoothing/stability parameters, out-of-support frequency, and correction performance on those cases.
Reviewer_yvYbOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses the challenge of predicting high-value regions in time series forecasting and introduces the PEAKTS framework. The proposed method predicts an ordered categorical distribution for future values and utilizes the distribution’s mean, upper quantiles, and confidence levels to adaptively refine point forecasts. A risk-conditioned peak-memory module retrieves prototypical patterns representing both typical and peak behaviors from the time, differential, and frequency domains to capture historical precursors to peak events. The experimental results demonstrate the effectiveness of the proposed method in forecasting high-value regions.

Critical strengths
  1. The paper offers a fresh perspective on time-series forecasting by focusing on rare but practically important high-value regions.

  2. PeakTS proposes a novel framework that combines risk-conditioned distribution correction with peak-precursor memory retrieval to improve high-value forecasting.

  3. The paper is clearly written and supported by comprehensive experiments on nine datasets, while the anonymous code repository supports reproducibility.

Critical weaknesses
  1. Although the manuscript evaluates the method on the top 75%, 50%, and 25% high-value regions, these evaluations are insufficient to support its applicability to rare peaks or extreme events. The experiments lack results for more extreme high-value regions, such as the top 10% or top 5%.

  2. In Equation (11), the correction strength parameter alpha is set to 0.8. However, the authors did not provide justification for this choice.

Questions for authors
  1. In Equation (10), why does the correction gate take elements such as X~\tilde{X}, μ\mu and qq as gate inputs? Are they all necessary for determining the correlation strength?

  2. How do the authors ensure that PEAKTS improves forecasting accuracy in high-value regions while preserving overall forecasting performance?

  3. How much additional memory overhead does the Risk-Conditioned Peak Memory (RCPM) module introduce?

1 comment

I have read the authors’ rebuttal. My questions have been addressed. The additional Top-5% experiments address my concern about forecasting performance in more extreme high-value regions. I would like to raise my rating.

Reviewer_yGK7OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose PeakTS, a time-series forecasting framework that predicts an ordered categorical distribution over future values and employs a risk-conditioned memory module to retrieve historical patterns associated with high-value events.

Critical strengths
  1. The proposed method is well-motivated and addresses an important and practically relevant forecasting task.

  2. The experiments and ablation studies are well-designed, and the results demonstrate the effectiveness and benefits of the proposed method.

Critical weaknesses

Some methodological choices require further justification and validation. Additionally, the experimental evaluation could be strengthened with further analyses to better support the effectiveness of the proposed method.

Questions for authors
  1. Could the authors clarify the motivation for using Gaussian smoothing to construct the soft target distributions over KK ordered bins in Section 3.1?

  2. The related work section could be strengthened by providing a more comprehensive discussion of existing studies related to the proposed components, For example, memory-based approach, although RAFT [1] is cited, its memory-based approach in relation to the proposed method. Other relevant studies, based on my quick literature search, such as SMLDR [2], could also be considered. Ideally, including those introduced methods (per each component) as experimental baselines would further demonstrate the advantages of the proposed approach.

[1] Retrieval augmented time series forecasting.

[2] SMLDR: Spectral Memory Learner with Dual-Retrieval for Time Series Forecasting

  1. I believe the proposed method shows promising results, while the authors may interested in see the same results in longer prediction horizon, since the main contribution that authors stated as high-value region forecast is evaluated in the shorted prediction regions in {96,192,336,720}\{96, 192, 336, 720 \}

  2. Proposition 3.1 appears to follow directly from Equation (11). However, this result holds for any correction target satisfying the same assumption and does not establish a theoretical advantage specific to the proposed distribution-guided correction mechanism based on my understanding. Could the authors clarify the significance of this proposition and provide additional justification?

  3. In Section 3.4, the authors use the 90th percentile of the predicted distribution and set the correction strength to α=0.8\alpha=0.8. Additionally, the risk-aware target is constructed using an equally weighted combination of the predictive mean and upper quantile. However, the manuscript does not provide sufficient justification for these choices. Could the authors clarify how these values were selected and ideally provide a sensitivity analysis for the upper quantile level, correction strength, and mean–quantile combination weights? Since these parameters directly influence the proposed distribution-guided correction mechanism, such an analysis would help establish the robustness of the method.

  4. Does Figure 3.b shows that retrieval is not actually provide significantly useful information since the results look insensitive to the choices of top-N?

1 comment

Thank you for the rebuttal. Most of my concerns have been addressed. For Q3, it will be useful if the results with more prediction horizon are provided, but given the short rebuttal period, i will not consider this as significant limitation.

I will keep my score positive as weak accept, and increase confidence and Informativeness in the detailed rubric score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.