acceptodds
Under review as a conference paper at ICLR 2027

DMCnet: Multi-Scale Trend and Patch Coordination for Online Time Series Forecasting under Concept Drift

Abstract

Online time series forecasting must adapt to evolving data distributions while limiting the propagation of errors across sequential updates. This paper presents DMCnet, a multi-scale coordination framework for online multivariate time series forecasting under concept drift. DMCnet employs two complementary predictors: a trend-oriented branch that extracts multi-scale temporal patterns using smoothed moving-average representations and dilated temporal convolutions, and a patch-oriented branch that models local dynamics from segmented input patches. The forecasts are combined online through convex weight integration, allowing the model to adjust the relative contributions of long-term and local predictors as the stream evolves. To reduce the effect of unstable short-term updates, we further introduce the Online Noise Sensitivity Protector (ONSP), which uses the discrepancy between the incoming sequence and its smoothed trend as a data-dependent signal to regulate short-term weight updates. The resulting mechanism is designed as a practical stabilization strategy rather than as a direct statistical estimator of observation noise. We evaluate DMCnet on five real-world datasets, including the newly released PCpower dataset, which records personal-computer power consumption under diverse usage scenarios. Across the evaluated datasets and forecasting settings, DMCnet reduces the average cumulative mean squared error (MSE) and mean absolute error (MAE) by 24.67% and 19.32%, respectively, compared with the strongest competing methods. These results indicate that coordinating multi-scale trend representations with local patch representations can improve online forecasting robustness under distributional change.

Primary area:
learning on time series and dynamical systems
Keywords:
Online time series forecasting, Concept drift, Online noise sensitivity protector

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vxCFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DMCnet for online multivariate time series forecasting under concept drift and error accumulation. It combines a multi-scale trend predictor and a local patch predictor via online convex optimization-based dynamic weighting, and introduces the Online Noise Sensitivity Protector (ONSP) to stabilize short-term weight updates using a DTW-based confidence signal. The authors also release the PCpower dataset. Experiments on five real-world datasets show consistent improvements in MSE and MAE over strong baselines, with ablations validating the main components.

Critical strengths
  1. This paper focuses on online learning for time series forecasting, which is a promising research direction.
  2. Figure 2 intuitively illustrates the motivation of DMCnet, where energy and rhythm misalignment are common issues in the time series forecasting domain.
Critical weaknesses
  1. The novelty of the proposed method is incremental. The idea of ensembling two predictors is inherited from OneNet, the design of the Dynamic Weight Integration module closely resembles that of OneNet, and the encoder block is identical to those used in FSNet and OneNet. The authors' contribution is mainly confined to the short-term forecasting weights of OneNet, which is rather limited.
  2. The authors emphasize the error accumulation problem in the introduction; however, Figure 1 presents cumulative error curves, whose computation inherently gives rise to error accumulation. This is therefore not convincing. Plotting point-wise error curves would be necessary to substantiate the authors' claim.
  3. The evaluation protocol adopted in this paper is outdated and suffers from information leakage [1], which substantially undermines the confidence in the experimental results. The authors are encouraged to adopt a fairer evaluation methodology (e.g., Proceed [1]) and to re-run the experiments.
  4. The most recent method compared is OneNet (NeurIPS 2023), which is no longer timely in ICLR2027. The authors should compare against more recent and stronger online time series forecasting baselines (e.g., Proceed [1], DSOF [2], Adaptz [3], DynaME [4]) to demonstrate the advantages of the proposed method.
  5. The public datasets used in this paper contain a limited number of variables, with the largest, Weather, having only 21 variables. High-dimensional online time series forecasting tasks are often more challenging. The authors should supplement experiments on higher-dimensional datasets (e.g., Traffic, ECL).
  6. The main text contains a non-anonymous github link, which I am concerned may violate the double-blind review policy.

[1] Zhao, L.; and Shen, Y. 2025. Proactive model adaptation against concept drift for online time series forecasting. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, 2020–2031

[2] Lau, Y. Y. A., Shao, Z., and Yeung, D. Y. 2025. Fast and slow streams for online time series forecasting without information leakage. In International Conference on Learning Representations (Vol. 2025, pp. 28736-28762).

[3] Huang, X., Qiu, S., Du, J., and Yang, C. 2026. Online time series prediction using feature adjustment. In International Conference on Learning Representations (Vol. 2026, pp. 150718-150744).

[4] Hong, S., Chae, S., Kim, S., Jang, S., and Yu, H. 2026. Dynamic Multi-period Experts for Online Time Series Forecasting. In Proceedings of the ACM Web Conference 2026 (pp. 7669-7678).

Questions for authors

Please refer to Weaknesses.

Reviewer_oA7fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DMCnet for online multivariate time-series forecasting under concept drift. DMCnet combines two complementary predictors: a trend-oriented predictor for multi-scale temporal patterns and a patch-oriented predictor for local dynamics, with their predictions dynamically weighted during online learning. To reduce error accumulation from noisy online updates, the paper introduces ONSP, which uses the difference between the input sequence and its smoothed trend to regulate short-term weight updates. The paper also introduces PCpower, a new dataset for component-level laptop power consumption across different hardware configurations and usage scenarios. The authors conducted experiments on five datasets to show the effectiveness of DMCnet.

Critical strengths

Well-designed PCpower dataset - The paper introduces a detailed component-level laptop power dataset covering multiple hardware configurations and real usage scenarios, with fine-grained behavior annotations. This provides a useful resource beyond the forecasting task studied in the paper.

Clear motivation for the two-branch design - The paper identifies two concrete forecasting errors, amplitude and temporal misalignment, and uses them to motivate complementary trend-oriented and patch-oriented predictors. This provides an intuitive connection between the observed problems and the proposed architecture.

Critical weaknesses

Limited evaluation of concept-drift adaptation - Although the paper is motivated by adapting to concept drift, the experiments mainly report cumulative MSE/MAE on fixed datasets and forecasting horizons. There is little analysis of how DMCnet behaves around identifiable distribution shifts, such as adaptation speed, performance before/after a shift, or robustness under different types or strengths of drift. This makes it difficult to determine whether the gains specifically come from better adaptation to concept drift or simply from a stronger forecasting architecture

Limited effectiveness of ONSP - ONSP provides only modest improvements over DMCnet without ONSP, with several settings showing negligible gains. Given that ONSP is presented as a main contribution, the experiments provide limited evidence that it materially contributes to DMCnet’s overall performance gains

Weak theoretical justification for ONSP - Proposition 1 in Appendix shows that noise can accumulate during online gradient updates, but it does not explain why the DTW distance between the input and its smoothed trend is an appropriate signal for controlling this noise. Thus, I believe that the connection between the theoretical analysis and ONSP remains heuristic.

Questions for authors

Please refer to the weaknesses.

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

The paper proposes DMCnet, a method for online forecasting of multivariate time series under concept drift. DMCnet keeps the online ensembling framework of OneNet but uses two different experts. The first is a trend-oriented predictor: moving averages at several scales followed by a TCN-like convolutional encoder. The second is a patch-oriented predictor: a PatchTST-like encoder applied to each channel independently. The two forecasts are combined with long-term weights updated by exponentiated gradient descent and short-term offsets produced by a small MLP. The main new component, ONSP (Online Noise Sensitivity Protector), scales the short-term offset by the sigmoid of the DTW distance between the input and its smoothed trend. Its goal is to limit the accumulation of gradient noise in the short-term weights, which Proposition 1 analyzes. The authors also release PCpower, a new dataset of component-level laptop power consumption. Experiments on five datasets with horizons 24 and 48 show lower MSE and MAE than the compared baselines.

Critical strengths
  • Extremely relevant problem choice. Nonstationarity in time series is a real-world problem that needs new solutions.
  • The idea is intuitive and simple. Combining multiple forecasts based on trend and patches is a solid proposal.
  • Besides a new approach, a new dataset is also proposed, which enriches the value of the paper.
Critical weaknesses
  • My largest worry is that the method itself is maybe too simple. The idea of combining the two forecasts from two parallel model branches comes directly from OneNet, and both the multi-scale and the patch approach are standard in forecasting models. The largest contribution seems to come from ONSP, but I would say that even here the theoretical contribution of Proposition 1 is very small. Dropping the signal-driven term as higher order and assuming independent noise around the “ideal signal” basically makes the whole proof a simple computation of the variance of a random walk.
  • In the experiments, the only methods designed specifically for online forecasting are FSNet and OneNet (with its variants). The other baselines are generic continual learning methods or offline architectures. More recent methods, such as PROCEED (Zhao & Shen, 2024) and NatSR (Urettini & Carta, 2026), are not compared, which makes the experiments weaker. PROCEED is even cited in the paper.
  • At some point, without any clear explanation, the terms Rhythm and Energy pop up.
Questions for authors
  • Online learning for time series depends greatly on the hyperparameters used along the entire stream of data. A wrong pick from the beginning would break the entire learning. Can you please describe which strategy you used to select them?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.