acceptodds
Under review as a conference paper at ICLR 2027

MET-C3: Mixed Exogenous–Temporal Dependency Modeling for Non-Stationary Time Series Forecasting

Abstract

Nonstationary time series forecasting is challenging, as the dependency structure among historical observations, temporal lags, and exogenous variables varies over time. Existing models capture temporal patterns implicitly, but often lack an explicit mechanism for distinguishing temporal inertia from exogenous impacts. This paper proposes MET-C3, a mixed node dependency graph framework for graph guided nonstationary forecasting. MET-C3 represents lagged target variables and exogenous variables as a unified node set, estimates dependency matrices in rolling windows, and introduces autocorrelation constrained coupling to weight dependencies among lag nodes, between lag and exogenous nodes, and among exogenous nodes. It further applies TMFG sparse topology construction and LoGo partial correlation inference to extract compact structural features, which are integrated with multiple forecasting backbones. Experiments on a proprietary Shandong Electricity Market dataset and three public datasets from electricity, mobility, and finance show that MET-C3 consistently reduces forecasting error across backbones and domains. Robustness analyses demonstrate its advantages under stronger structural nonstationarity, while runtime comparisons show comparable computational cost to the original backbones. These results indicate that explicit mixed dependency modeling offers an effective and interpretable approach to nonstationary forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
nonstationary time series forecasting, mixed node dependency graph, exogenous impact modeling, sparse graphical inference, graph-guided forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zmcrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The core innovation of this paper lies in proposing a method that simultaneously models the dependencies among lag nodes, between lag and exogenous nodes, and among exogenous nodes. By integrating TMFG topology construction and LoGo partial correlation inference, the proposed framework demonstrates strong forecasting performance across multiple datasets.

Critical strengths

The problem addressed by the authors represents a valid and worthwhile research direction.

Critical weaknesses

(1) Relegating all mathematical formulations in Section 3 (Methodology) to the appendix is highly unconventional for standard academic writing. Similarly, moving all experimental results (tables) and visualizations (figures) from Sections 4 and 5 entirely to the appendix is strongly discouraged. This structural flaw severely hampers the readability of the manuscript and constitutes the primary reason for my rejection recommendation. I believe this significantly violates standard ICLR submission guidelines regarding the proper division of content between the main text and the appendix.

(2) The paper fails to provide sufficient in-depth experimental analysis or visualizations to demonstrate the actual effectiveness of the proposed modeling approach for mixed temporal and exogenous dependencies. This critical omission significantly weakens the paper's motivation and undermines its claimed novelty.

(3) The mathematical formulas presented in Figure 1 are cluttered and do not conform to standard academic formatting conventions. The authors are strongly advised to carefully refine and reformat this figure.

(4) The abstract fails to effectively highlight the core novelty of the research. Additionally, the use of abbreviations is non-standard. For instance, in the sentence “It further applies TMFG sparse topology construction and LoGo partial correlation inference to extract compact structural features,” the acronyms (TMFG and LoGo) are introduced without their full expansions. Furthermore, the abstract lacks a clear explanation of the underlying purpose and significance of applying these specific techniques.

Questions for authors

no further questions

Reviewer_TnjoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes MET-C3, a mixed node dependency graph framework for graph-guided nonstationary forecasting. Specifically, MET-C3 represents lagged target variables and exogenous variables as a unified set of nodes and introduces autocorrelation-constrained coupling to weight dependencies among lag nodes. Then, it applies TMFG sparse topology construction and LoGo partial correlation inference to extract compact structural features. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths

The evaluation datasets are appropriate for the forecasting setting considered in the paper.

Critical weaknesses

This paper lacks innovation, as the key modules (e.g., autocorrelation-constrained coupling and sparse topology construction) are well-established techniques. The proposed method primarily integrates these techniques without introducing task-specific designs tailored to non-stationary time series forecasting.

Presentation and organization need major improvement. Several issues make the paper unnecessarily difficult to follow:

  1. Key figures, equations, and experimental results are scattered across different parts of the main text and appendix, forcing readers to flip back and forth to locate the relevant content.
  2. The related work section devotes considerable space to background material with limited relevance to the central contribution. The manuscript also includes a Keywords line below the abstract, which appears inconsistent with the standard conference paper format.
  3. Terminology is used inconsistently (e.g., "non-stationary" vs. "non stationary"), and the method name is not uniform throughout the paper (e.g., "MET-C3" vs. "MET-Gated").

The authors are encouraged to consult high-quality published papers (e.g., those appearing at AAAI, NeurIPS, ICLR, ICML, and KDD) and revise the manuscript accordingly.

Questions for authors
  1. Most of the baselines are published before 2024 and no longer state-of-the-art (SOTA). The authors should include comparisons with more recent SOTA methods from 2025 and 2026.

  2. How are the rolling-window length, lag set, graph update frequency, and coupling parameters selected? In particular, are the autocorrelation decay factor and other data-dependent quantities estimated exclusively from information available at each forecasting origin? Please also clarify whether Figure 7 is based on validation or test performance.

  3. What is the exact difference between M5 and M6? Table 3 describes M5 only as mixed features “without the full MET-C3 design.” Could you specify which components differ and provide a controlled comparison that isolates the contribution of autocorrelation-constrained coupling while keeping the node set, graph updates, and feature construction unchanged?

Reviewer_p4LgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors focus on the problem of efficiently modeling time series historical dependencies, temporal lags, and exogenous variables that exhibit non-stationary evolution while forecasting. The proposed MET-C3 method introduces a mixed node dependency graph framework for non-stationary forecasting. Specifically, lagged target variables and exogenous variables are represented as a unified node set, where dependency matrices are estimated based on Pearson autocorrelation from rolling windows, and then reweighted with an autocorrelation-constrained coupling. The proposed reweighted scheme down-weights long lags, sets exogenous edge weights, and controls coupling between lags and exogenous variables based on specific weight values. Additionally, TMFG sparse topology construction and LoGo partial correlation inference are applied to extract compact structural features that are integrated with several forecasting backbones. The experimental evaluation focuses on a proprietary Shandong electricity market dataset and three additional public datasets. Adding MET-C3 features is shown to achieve lower MAE values compared to vanilla backbones with comparable computational costs. A robustness analysis shows advantages under stronger non-stationarity, while runtime comparisons show comparable computational cost to the original backbones, which can be attributed to explicit mixed dependency modeling.

Critical strengths

The main strengths of this work are the following:

  • The proposed method works with any backbone, while the added cost is low. A few low-dimensional input features are added, whereas the backbone can differ (e.g., in the experiments, backbones range from gradient boosting to transformers).
  • Interpretability is implicitly achieved via the proposed modelization. Lagged targets and exogenous variables share one graph and different roles. Bootstrap stability analysis shows how dependencies change between regimes, which directly supports the interpretability claims of the paper.
Critical weaknesses

The main weaknesses of this work are summarized as follows:

  • The presentation of the work and its methodological novelty are very limited. There are no important formulations (dependency estimation, coupling weights, LoGo, features, objective) provided in the main paper’s methods section, but rather in Appendix B. Similarly, results’ Tables and Figures are provided in Appendix C. This makes the method’s core understanding and its significance very hard to understand and quantify. Similarly, the syntax and expression in the text are in several places very complex. Additionally, several methodological components, such as rolling correlations, dependency graphs, TMFG, and LoGo, pre-exist in the literature. The novelty part is limited to the proposed re-weighting scheme and the feature vector, which remain rather conceptually simple, while no theoretical justifications for their importance are given.
  • The experimental evaluation is questionable. The chosen baselines are very few, while there are no comparisons with forecasting backbones focusing on exogenous modeling (e.g., TimeXer, NBeatsX), and non-stationarity (e.g., RevIN, Non-stationary transformer). Additionally, many variances exceed or match means of gains (MTGNN - Table 4, LSTM, UCI -Table 5), and performance gains are non-consistent in most cases. The datasets chosen are very few, not following the standard forecasting benchmarks (https://github.com/thuml/Time-Series-Library), while results on the proprietary dataset are not reproducible.
Questions for authors
  • Authors should tackle concerns about methodological novelty, presentation, and experimental rigor presented above. Consider adding baselines and datasets, and better highlighting the theoretical insights behind the method's main components.
  • It is unclear how successful runs are selected, and significance tests should be provided. Also, it is unclear why neural backbones perform worse than LightGBM and why variances exceed means in several cases.
  • How are weighting parameters (alpha, beta, and gamma) tuned, and what splits are used for hyperparameter tuning in general?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.