acceptodds
Under review as a conference paper at ICLR 2027

Channel-Dependent State Space Model for Multivariate Time Series Forecasting

10 views

Abstract

Multivariate time series forecasting (MTSF) is critical across many real-world domains. Existing deep learning approaches fall into two paradigms with distinct limitations: channel-independent (CI) methods unconditionally ignore cross-variable dependencies and model only temporal dynamics, while channel-dependent (CD) methods consider both but typically rely on architectural compromises to mitigate overfitting and computational overhead. We therefore propose Chameleon, a specialized CD state space model (SSM) that enables data-dependent, fine-grained interactions across variables while scaling linearly with their number. By connecting selective SSMs with the Kalman filter, we leverage the missing measurement update in the former for cross-variable modeling while preserving the SSM backbone for robust temporal modeling. We further identify favorable inductive biases of GatedDeltaNet for time series, adapt it as our backbone, and improve generalization through additional techniques, including a previously unexplored stochastic perturbation of reversible instance normalization. On strongly dependent ODE and PEMS datasets, Chameleon achieves the best MSE and MAE across all settings, while its CI ablation and prior CD methods incur 61-178% higher MSE on average. Across 28 standard benchmark settings, Chameleon also achieves better MSE and MAE than each baseline in at least 27 and 22 cases, respectively. Training-time and peak-memory analyses on Traffic and ETT further demonstrate competitive efficiency and favorable memory scalability across different variable counts.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, State Space Models, Kalman Filter

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EQnFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Chameleon, a novel Channel-Dependent (CD) State Space Model (SSM) designed for Multivariate Time Series Forecasting. To overcome the scalability and performance trade-offs of existing models, Chameleon draws inspiration from the Kalman filter, decoupling the forecasting process into a "time update" and a "measurement update". The "time update" leverages GatedDeltaNet to model temporal dynamics independently for each variable. The "measurement update" introduces a data-dependent, cross-variable state calibration mechanism that utilizes the Woodbury matrix identity to achieve linear complexity with respect to the number of variables. Furthermore, the authors introduce a stochastic perturbation technique for RevIN during training to enhance generalization. Extensive experiments on strongly dependent physical systems (ODE datasets), PEMS traffic datasets, and 7 standard long-horizon benchmarks demonstrate that Chameleon achieves SOTA performance with competitive computational efficiency and favorable memory scalability

Critical strengths
  1. Connecting selective SSMs with the Kalman filter's recursive two-step cycle to neatly factorize intra-series evolution (time update) and inter-series interaction (measurement update) is conceptually elegant and theoretically sound.

  2. By applying the Woodbury matrix identity to the measurement update, the model avoids explicitly materializing the N×NN \times N correlation and gain matrices. This reduces the inversion complexity, ensuring strictly linear scaling with the number of variables NN.

  3. Chameleon consistently outperforms strong CI and CD baselines in terms of MSE and MAE across almost all 28 evaluated settings.

Critical weaknesses
  1. The choice to dynamically apply or disable RevIN and decomposition relies on highly heuristic thresholds (e.g., Trend > 0.2, Std. Shift > 1.2) calculated over the training and validation splits. Such hard-coded thresholds may lack robustness when applied to unseen domains or streaming data.

  2. The paper evaluates performance on spatial-temporal traffic datasets (PEMS03, PEMS08), but compares mostly against generic multivariate baselines rather than state-of-the-art Spatial-Temporal Graph Neural Networks (STGNNs), which are specifically designed to leverage explicit sensor structures in these datasets.

  3. The evaluation on the heavily coupled ODE datasets is restricted to shorter forecasting horizons (F∈{48,96,192}F \in \{48, 96, 192\}) due to error accumulation in chaotic systems. The paper lacks an analysis of how the proposed measurement update degrades over long-horizon rollouts in these challenging scenarios compared to standard CD models.

  4. The ablation studies successfully show that removing the entire "measurement update" degrades performance. However, the paper lacks an architectural ablation comparing this specific Kalman-inspired O(N)\mathcal{O}(N) update against a simpler baseline, such as standard linear attention across the channel dimension, to prove its representational superiority.

Questions for authors
  1. Please provide a sensitivity analysis or further justification for the hard thresholds (Trend > 0.2, Std. Shift > 1.2). What happens to the model's performance if these heuristics are marginally misaligned or if a fully learnable, threshold-free module is used instead?

  2. To validate that the Kalman-inspired measurement update is not just computationally efficient but representationally superior, please provide an ablation where the measurement update block is replaced by a standard O(N)\mathcal{O}(N) linear cross-channel attention mechanism.

  3. While PEMS03 and PEMS08 showcase strong dependencies, the comparisons lack spatial-temporal specific models (e.g., STGCN, DCRNN, AGCRN). Including at least one modern STGNN benchmark would strengthen the claim that your implicit modeling matches or exceeds models with explicit graph structures.

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

This paper proposes Chameleon, a channel-dependent state space model for multivariate time series forecasting. It combines a GatedDeltaNet backbone for temporal modeling with a Kalman-inspired measurement update for cross-variable state correction. By leveraging structured covariance parameterization and the Woodbury identity, the proposed method achieves linear complexity with respect to the number of variables. Experiments on standard forecasting benchmarks and strongly dependent datasets demonstrate competitive predictive performance and computational efficiency.

Critical strengths

1.Well-motivated and unified framework

The paper draws a connection between state space models and Kalman filtering, naturally integrating temporal modeling and cross-variable interactions. The overall design is clear and provides a reasonable inductive bias.

2.Efficient cross-variable modeling

Through structured covariance parameterization and the Woodbury identity, the proposed method avoids explicitly constructing and inverting large cross-variable matrices, achieving linear complexity with respect to the number of variables.

3.Strong empirical performance

The method achieves competitive results on standard forecasting benchmarks and particularly strong performance on datasets with significant cross-variable dependencies. The ablation studies and efficiency analyses further support its effectiveness.

Critical weaknesses

1. The current ablation study does not fully isolate the contribution of cross-variable interactions.

The paper constructs Chameleon (CI) by removing the entire Measurement Update module and attributes the performance improvement mainly to cross-variable dependency modeling. However, this removes both cross-variable interactions and channel-wise state correction, making their individual contributions unclear.

Specifically, when γ=0\gamma=0, the Kalman gain becomes diagonal, disabling cross-variable interactions while preserving channel-wise state correction through a data-dependent gating mechanism. Such correction may itself improve performance.

I suggest adding an ablation with γ=0\gamma=0 fixed during training. Comparing this variant with Chameleon (CI) and the full model would better isolate the contribution of cross-variable interactions.

2. The advantage of the Kalman-inspired parameterization over general Diagonal + Low-Rank Mixing remains unclear.

The Kalman-inspired Measurement Update provides a well-motivated and computationally efficient approach to cross-variable state correction. However, the resulting mechanism can be expressed in a diagonal-plus-low-rank form, and it remains unclear whether the specific Kalman-inspired parameterization offers advantages over more general mixing mechanisms.

For example, a baseline could use K=diag⁡(g)+AB⊤K=\operatorname{diag}(g)+AB^\top, where gg, AA, and BB are dynamically generated from the input and prior states, with appropriate normalization and stability constraints. Such a design can also achieve data-dependent interactions with linear complexity.

I suggest comparing these approaches under matched backbones, parameter counts, and computational budgets to clarify whether the gains arise from low-rank mixing itself or the specific inductive biases of the Kalman-inspired design.

3. The look-back window selection procedure requires clarification.

Section 4.2 states that multiple look-back window lengths TT are evaluated and the best results are reported. However, it is unclear how the optimal TT is selected.

If test performance is used for selection, this could introduce test-set selection bias and lead to overly optimistic results, especially since different models may use different optimal windows.

I suggest clarifying the selection procedure.

Questions for authors

see above

Reviewer_23uXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Chameleon, a channel-dependent (CD) state-space model for multivariate time-series forecasting (MTSF). The main idea is to use GatedDeltaNet as a temporal backbone and to add a Kalman-filter-inspired cross-variable measurement update. Each variable is first processed independently over time. Then a data-dependent observation signal, a diagonal prior covariance, and a low-rank correlation structure are used to construct a Kalman gain that calibrates hidden states across variables. The authors use a Woodbury-form computation to avoid explicitly materializing N×N matrices and claim linear complexity in the number of variables. They also introduce a stochastic perturbation of reversible instance normalization (RevIN) during training and optionally use decomposition and patching. The method is evaluated on strongly dependent systems, including Double Pendulum, Lorenz Coupled, PEMS03, and PEMS08, as well as seven standard long-horizon benchmarks (ETT, Electricity, Solar, Traffic).

Critical strengths
  1. The paper does not simply stack attention or an SSM along the variable dimension. Instead, it assigns temporal modeling to a GatedDeltaNet-style time update and cross-variable calibration to a measurement-update-like module. The low-rank/Woodbury formulation is a reasonable attempt to make cross-variable interaction scalable.
  2. Chameleon achieves the best MSE and MAE on the ODE and PEMS datasets across the reported settings. The ablation is also informative: removing the measurement update increases MSE and MAE by about 61% and 34% on average, with much larger degradation on Lorenz Coupled.
  3. The paper compares against several major modeling paradigms, covers both strongly dependent and standard benchmarks, and reports parameter counts, training time, and peak memory as the number of variables changes. The ablations on RevIN perturbation and decomposition also provide some insight into stability and preprocessing choices. Overall, the empirical section is more complete than a minimal proof-of-concept.
Critical weaknesses
  1. The update is performed once after the full temporal recurrence, sets H_l=I, uses diagonal prior covariance, and is closer to a low-rank gated calibration than a faithful Kalman measurement update. No direct ablation against a generic low-rank cross-variable correction isolates the claimed novelty.
  2. The authors report results from the best look-back T per model-dataset-horizon but do not clearly state whether selection uses validation or test data, nor whether all baselines receive the same tuning budget and look-back grid. This materially affects the SOTA claim.
  3. On weakly dependent ETT datasets, removing the measurement update often changes MSE by only about 0.04%–1%. Strong evidence comes mainly from ODE/PEMS systems, so the broad superiority claim across diverse MTSF settings is not fully supported.
  4. Woodbury avoids explicit N×N matrices, but observation construction and low-rank factors may still be costly. Scalability is mainly shown on Traffic up to 862 variables; larger-N experiments and component-wise FLOPs/memory breakdowns are needed.
  5. Corrected states are not fed back into the GatedDeltaNet recurrence, and all calibration is applied after the temporal pass. This weakens the Kalman-style “recursive correction” narrative and leaves unclear why this route is preferable to post-hoc cross-variable calibration.
  6. The authors omit the original convolutional layers for k and v and rely on patch embeddings for local details. This is a technical-route choice that may affect both accuracy and efficiency, but its individual contribution is not separated from the measurement update.
Questions for authors
  1. The authors should add a direct ablation comparing the proposed measurement update against a generic low-rank cross-variable linear correction, variable-wise gated attention, or a variable-dimension SSM under matched parameter and compute budgets. This would clarify whether the specific Kalman-style formulation is necessary.
  2. The authors should clearly state whether the best look-back T is selected on validation or test data, and should confirm that all baselines use the same tuning budget, look-back grid, and early-stopping criteria. Per-setting mean±std and statistical significance should also be reported clearly.
  3. On the same backbone, how does the Kalman update compare with variable attention, a low-rank linear mix, and a variable-axis SSM on the ODEs, PEMS08, and Traffic?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.