acceptodds
Under review as a conference paper at ICLR 2027

Forecast Privately, Correct Minimally:Bounded Relational Interventions for Multivariate Time Series

Abstract

Cross-variable information can improve multivariate forecasting, yet an unrestricted mixer can also replace reliable channel-private dynamics and propagate local faults to healthy channels. We ask a narrower question: how much relational intervention is sufficient? We introduce the Minimum-Intervention Relational Forecaster (MIRF), which first produces a complete channel-private temporal representation and then applies one identity-start residual correction. A C-r-C factorization limits the correction to r channel directions, while spectral caps and a bounded scalar gain limit its magnitude. This design uses one ordinary MAE objective and no auxiliary routing or decomposition losses. Under a fixed look-back of 96 and four forecast horizons, MIRF ranks first in 12 of 16 dataset-average MSE/MAE cells across eight benchmarks and a compact recent baseline set, including two 2026 methods. Controlled relation misalignment and sensor-outage tests show why the constraint matters. On Traffic with 50% channel outage, the selected bounded correction incurs 14.11% untouched-channel MSE spillover, compared with 41.90% for an unbounded low-rank correction and 111.72% for an unbounded dense correction. At 862 channels, rank four uses 207.8 fewer relational parameters and runs 37.7 faster at inference than the matched dense correction. These results support a minimum-intervention principle: relational structure is most useful as a controlled amendment to a private forecast, rather than as its unrestricted replacement.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time-series forecasting, long-term forecasting, cross-variable modeling, low-rank relational correction, robust forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ABy5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

MIRF augments channel-independent temporal representations with a low-rank relational correction initialized at zero and constrained by spectral caps and a bounded gain. Experiments across eight benchmarks and controlled sensor faults show competitive accuracy and highlight the trade-off between cross-channel information sharing and error propagation.

Critical strengths

Controlled channel perturbations and matched comparisons directly examine the trade-off between forecasting accuracy and error propagation, providing useful evidence for bounded relational corrections.

Critical weaknesses
  • The approach combines established low-rank updates, zero initialization, and norm constraints. The theoretical properties largely follow from this construction, offering limited new insight into when minimal relational intervention is preferable.
  • It is unclear whether the private control evaluates the trained MIRF with its correction disabled, making it difficult to assess whether a strong private forecast is preserved after training.
Questions for authors
  • Is the private control trained separately or obtained by disabling the correction in the trained MIRF? If trained separately, what are MIRF’s MSE and MAE with the correction disabled and no retraining?
  • With the same spectral caps and gain ceiling, how do low-rank and dense corrections compare in clean accuracy and error propagation on Traffic and Electricity?
Reviewer_boffOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MIRF, which adds an identity-initialized, low-rank, spectrally bounded correction to a channel-independent forecaster. Evaluation covers eight forecasting benchmarks, channel-corruption tests, architectural ablations, and computational scaling.

Critical strengths

The architecture achieves competitive accuracy across eight datasets while substantially reducing channel-mixing costs.

Measuring corruption spillover on untouched channels provides useful evidence about error propagation, including clearly disclosed failure cases.

Critical weaknesses

The prescribed gain-selection rule does not reliably control propagation: on Electricity, the selected model incurs 211.84% spillover under 50% outage versus 40.03% for dense mixing. This limits the practical robustness benefit.

The benefit of low-rank beyond norm control is unclear.

Questions for authors

NA

1 comment

My main concern regarding the effect of low-rank control is still unresolved. Therefore, I will maintain my rating.

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

This paper proposes the Minimum-Intervention Relational Forecaster (MIRF) for multivariate time series forecasting. The model first constructs channel-private temporal representations and then applies a low-rank residual correction across channels. Zero initialization makes the correction vanish at the start of training, while spectral normalization and a bounded scalar gain constrain its strength. The authors evaluate forecasting accuracy on eight benchmarks and study the propagation of channel misalignment and sensor outages through the relational module.

Critical strengths
  1. The paper addresses a relevant limitation of cross-variable forecasting: information from other channels can improve clean prediction while spreading errors when a subset of sensors becomes unreliable.

  2. The proposed architecture is simple and its components have clear intended roles. The low-rank factorization, bounded gain, and zero initialization make the relational amendment easy to inspect.

  3. The controlled fault experiments provide useful evidence beyond ordinary forecasting tables, and the paper explicitly reports settings where the proposed method does not provide the best robustness.

Critical weaknesses
  1. The main design consists of channel-independent forecasting, low-rank cross-variable interaction, and bounded residual correction, but these components are all based on existing general design principles. For example, CPiRI first performs channel-independent modeling and then introduces cross-variable interactions, while low-rank cross-variable interaction resembles existing Transformer-based variable routing approaches, such as Crossformer and PhaseFormer. The main contribution of this paper appears to lie in combining these components into a minimum-intervention relational forecasting framework and constraining the number of directions and magnitude of cross-variable corrections to control error propagation. However, the paper does not sufficiently clarify the substantive differences between the proposed design and existing routing-based variable interaction and residual correction methods, leaving its methodological novelty limited.

  2. The paper proposes to first construct a complete channel-independent forecaster and then introduce cross-variable information through low-rank residual correction, with the aim of reducing interference with the original predictions. However, the existing experiments mainly compare the channel-independent forecaster with variants of relational correction using different ranks and gain bounds, and do not establish that residual-based correction reduces interference more effectively than direct cross-variable interaction. The paper lacks a controlled comparison under the same temporal modeling backbone between residual-based correction and direct variable interaction. Such an experiment would help establish the advantages of the proposed architecture over existing approaches that separate channel-independent modeling from cross-variable interaction, such as CPiRI and TimeBridge.

  3. The paper reports that, with 862 variables and a relation rank of 4, the low-rank relational module uses 207.8 times fewer parameters than the dense correction. However, according to the parameter-count formulas provided in the paper, 2Cr+12Cr+1 for the low-rank module and C2C^2 for the dense mapping, the parameter reduction ratio is approximately 107.7, which is inconsistent with the reported figure. The paper does not sufficiently explain the parameter-counting convention underlying this discrepancy, and the quantitative claims regarding computational efficiency therefore require further clarification.

Questions for authors
  1. What substantive advantage distinguishes minimum-intervention relational forecasting from existing router-based interaction and low-rank residual correction methods?

  2. How does residual-based correction compare with direct cross-variable interaction when the same temporal backbone and comparable training conditions are used?

  3. Could the authors reconcile the reported parameter-reduction ratio with the parameter-count formulas given in the paper and clarify the counting convention?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.