acceptodds
Under review as a conference paper at ICLR 2027

DSSER: Dynamic Sparse Spatiotemporal Evidence Restoration for Drift-Robust Multivariate Time Series Forecasting

Abstract

Multivariate time series forecasting predicts future states from multiple interdependent historical series, yet real-world observations are often incomplete. Existing approaches typically address missingness by reconstructing unavailable values or learning representations robust under incomplete inputs. However, we observe that forecasts from incomplete observations drift away from those obtained with complete histories as missingness conditions vary. This drift arises from missingness-induced changes in the availability of historical evidence and the reliability of the remaining evidence. To mitigate such drift, we propose Dynamic Sparse Spatiotemporal Evidence Restoration (DSSER). The framework selectively leverages reliable evidence available under incomplete observations to improve forecasting stability. Reliability-Guided Evidence Routing (RGER) further regulates the contribution of restored evidence through learned routing confidence, introducing it as a bounded residual correction to the base forecast. To maintain reliability across training and inference, DSSER further introduces Cross-Stage Uncertainty Calibration (CSUC). Teacher predictive uncertainty and branch missingness jointly regulate teacher–student supervision during training, while learned routing confidence regulates restored evidence utilization at inference. Across three public traffic and air-quality benchmarks, DSSER consistently outperforms existing methods under random missingness, achieving the best results across all 36 dataset-rate-metric combinations. Under unseen-rate and dynamic-missing shifts, DSSER maintains robust performance, reducing MAE by 2.5%–3.6% on average relative to the best-performing non-DSSER baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Missing Data, Robust Forecasting, Spatiotemporal Modeling, Uncertainty Estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QuF2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DSSER addresses multivariate forecasting when historical observations are missing. It combines an STID-based forecasting model with a sparse evidence module that retrieves historical observations and uses them to make a bounded correction to the base forecast. During training, a teacher with access to the original history supervises student views with different missing rates. Teacher uncertainty adjusts the training losses, while a learned gate controls the evidence correction during inference.

Critical strengths
  • The paper is in good structure and well organized

  • Recovered information can hurt a forecast when it is unreliable. Maintaining a base prediction and limiting the size of the evidence-based correction is a sensible design. The teacher and its Monte Carlo uncertainty estimation are required only during training, which keeps the inference procedure relatively lightweight.

  • DSSER achieves the best results in all 36 main dataset–rate–metric comparisons. At 90% missingness, its MAE improvements over the strongest reported competing baseline are approximately 5.1% on PEMS04, 5.0% on METR-LA, and 4.7% on China-AQI.

Critical weaknesses
  • The paper adds considerable terminology and machinery to forecasting with missing observations. It is unclear what additional insight comes from framing prediction differences as “forecast drift,” particularly since lower drift does not necessarily correspond to better forecasting accuracy. The framework then combines multiple student views, Monte Carlo teacher uncertainty, several auxiliary losses, three evidence anchors, and a learned routing mechanism. Although the ablations show benefits from individual components, they do not establish whether this full framework is necessary compared with a simpler missingness-aware forecaster with a bounded residual correction. This makes the main contribution harder to identify and increases the training and tuning burden. A comparison against such a simpler alternative would help justify the complexity.

  • The meaning and evaluation of “uncertainty calibration” need clarification. As described, CSUC uses teacher uncertainty and missingness to adjust supervision weights, while RGER produces a learned gating score. It is unclear what statistical calibration property these quantities are intended to satisfy. The reported Spearman correlation and high-/low-risk MAE comparisons demonstrate risk discrimination, but do not establish calibration. The authors should first specify the calibration target and then provide an appropriate evaluation. For regression uncertainty, this could include predicted-error-scale versus empirical-RMSE reliability diagrams and ENCE, or coverage calibration and interval width if predictive intervals are available. If routing confidence represents the probability that a candidate correction improves the base forecast, an event-based reliability diagram and ECE would be appropriate. Otherwise, it should be described and evaluated as a gating weight rather than a calibrated probability. Although the paper cites calibration-related distillation work, its relationship to established regression calibration methods, such as Kuleshov et al. (2018) and Levi et al. (2022), remains insufficiently discussed. This distinction matters because the current evidence supports uncertainty-aware weighting more directly than statistical uncertainty calibration.

Questions for authors

See weekness above, the paper lacks the deep dive into the uncertainty calibration and complicates the whole process.

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

The paper proposes DSSER for multivariate time-series forecasting with missing observations. It combines sparse spatiotemporal evidence restoration, reliability-guided forecast correction, and uncertainty-aware teacher–student training to improve robustness under changing missing rates.

Critical strengths
  1. The reliability-guided correction offers a coherent approach to controlling how restored evidence influences predictions.
  2. The method shows consistent forecasting improvements across the evaluated datasets and missing rates, with ablations supporting the contributions of its main components.
Critical weaknesses

1.Experiments mainly use a fixed-length input window and forecasting horizon, leaving performance under varying history lengths and prediction horizons unexplored. 2. Missingness is only generated through random point masking. Testing structured patterns such as contiguous missing blocks or evaluating on datasets with naturally occurring missing values will strengthen the experimental analysis.

Questions for authors

Could the authors provide additional evaluation results addressing the concerns above? Experiments with varying input lengths would help make the evaluation more complete. Given the limited rebuttal time, however, I would prioritize experiments with different missingness patterns, or datasets with naturally occurring missing values. These would provide stronger empirical evidence for the method’s robustness in realistic settings.

1 comment

Thank you for the additional experiments. The results across different history lengths and forecasting horizons address my first concern. In particular, the tests with contiguous missing blocks and spatially correlated missingness address my main concern about robustness beyond random masking. I have no further questions and will raise my score.

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

This paper studies multivariate time-series forecasting under missing observations by explicitly con- sidering the availability and reliability of spatiotemporal evidence. The paper appears to be a modest extension over the SOA.

Critical strengths

The paper is very thorough and it message clearly articulated. The authors have done a good job in making their case.

Critical weaknesses

I am unconvinced that this is a significant contribution beyond the SOA. The authors really need to explain the "leap" in this paper in a better manner.

Questions for authors

What is the significance of this work over the prior art?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.