acceptodds
Under review as a conference paper at ICLR 2027

TimeSelect: Variable-Wise Historical Context Selection for Time Series Forecasting

Abstract

Most time series forecasting methods use a fixed-length historical window and implicitly assume that all observations are equally informative. In practice, however, the effective context varies across forecasting instances and variables, while distant observations may introduce outdated or irrelevant patterns. To address this issue, we propose TimeSelect, a forecasting framework that adaptively selects informative history and organizes it into a stable temporal representation. Before training, TimeSelect derives temporal bases from lagged correlations in the training data to capture stable dependencies across temporal positions. During forecasting, a lightweight selector estimates a variable-specific context length and constructs a differentiable mask to suppress less informative observations. The selected history is then projected onto the correlation-derived bases and processed by a learnable backbone to model inter-variable dependencies and predict the future trajectory. To preserve recent periodic variations that may be smoothed by global modeling, a period-aware refinement module further adjusts the initial forecast using gated residuals derived from the latest observed cycles. Extensive experiments on diverse real-world benchmarks demonstrate the consistent effectiveness of TimeSelect.

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

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 2 reviews, above 99% of the others.
Reviewer_M48zOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TimeSelect, a time-series forecasting framework that adaptively selects informative historical contexts for individual variables and forecasting instances. Unlike conventional methods that rely on fixed-length input windows, TimeSelect employs a differentiable context selector to suppress potentially irrelevant historical observations. The framework further incorporates correlation-derived temporal bases to capture stable temporal dependencies and a cycle-aware refinement module to preserve recent periodic patterns. Extensive experiments on 13 real-world datasets demonstrate competitive forecasting accuracy against SOTA baselines.

Critical strengths
  1. The paper addresses a meaningful and practical problem in time-series forecasting by recognizing that informative historical contexts vary across variables and forecasting instances.

  2. The proposed framework effectively integrates variable-wise context selection, correlation-derived temporal bases, and cycle-aware refinement. These components are well motivated and complement each other in capturing relevant historical information and temporal patterns.

  3. The experimental evaluation is comprehensive, covering 13 real-world datasets and comparisons with 13 competitive baselines.

Critical weaknesses

The paper could further explore sensitivity to the context-mask temperature, as well as the robustness of cycle-aware refinement under conditions of distribution shift or weak periodicity. Additionally, certain implementation details require clarification; please refer to the question box below.

Questions for authors
  1. How sensitive is TimeSelect to the context-mask temperature? Could the authors clarify whether these hyperparameters require dataset-specific tuning?

  2. The proposed temporal bases are constructed from training-set correlations and remain fixed during inference. How would the model respond to substantial distribution shifts or changes in temporal periodicity?

  3. Since cycle-aware refinement explicitly uses recent periodic observations, how does the framework behave on datasets with weak or irregular periodicity?

  4. Since RevIN is applied before historical context selection, are the normalization statistics computed from the entire input window or only the selected historical context?

Reviewer_x4s3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how to identify informative historical contexts for time-series forecasting. The proposed "TimeSelect" framework adaptively selects variable-specific look-back windows and integrates correlation-derived temporal bases with periodic refinement to improve forecasting accuracy. Extensive experiments on 13 datasets demonstrate superior performance over competitive baselines, along with favorable computational efficiency.

Critical strengths
  • The paper addresses the important problem of identifying informative historical contexts for time-series forecasting, offering a novel perspective beyond fixed look-back windows.

  • The proposed framework is well motivated, with an adaptive context selection mechanism that effectively captures relevant historical information. And extensive experiments systematically demonstrate its effectiveness. The anonymous code repository further supports reproducibility.

Critical weaknesses

My main concerns focus on the generalizability of the proposed framework and the clarity of its methodological design. Several architectural choices and implementation details require further explanation to better establish the reproducibility of the approach. The details are summarized in the table below. I hope the authors will consider addressing these points to improve the quality of the paper.

Questions for authors

Could the authors evaluate the proposed context selection mechanism on other backbone architectures—such as MLP-based or Transformer-based models—to assess its generalization capability?

How does the "Cross-Variable Representation Learning" module model the relationships between variables? Does it involve explicit variable interactions? Furthermore, would employing a GNN or a Transformer architecture yield better performance?

The proposed framework uses a cross-variable interaction matrix to capture dependencies among different variables. is this matrix dynamically updated for each input sample or remains fixed during inference?

The experimental analysis shows that the selected historical context lengths vary across prediction horizons. And separate context selectors are trained for different prediction horizons or whether the same selector is shared across them?

The learnable residual adjustments δin\delta_{\mathrm{in}} and δout\delta_{\mathrm{out}} are introduced in Eqs. (11) and (14). What are their exact dimensions, and are these parameters shared across variables or learned independently?

Comments on the paper

1 comment

Dear Reviewers M48z and vdGA,

The authors have responded to your questions. Please review their responses and consider whether your concerns have been addressed. You may also take a look at the other reviewers' comments and the revised manuscript and share any further feedback if you have any.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.