acceptodds
Under review as a conference paper at ICLR 2027

Matryoshka: Evidence-Gated Nested Channel Mixing for Time Series Forecasting

Abstract

In recent years, channel mixing has become a standard component of multivariate time series forecasting. Most existing methods either mix all channels or restrict mixing to a fixed or learned interaction structure, without testing whether the other channels carry predictive information. As a result, a module that helps on datasets with hundreds of channels can hurt on datasets with only a few. We propose **Matryoshka**, a plug-in module that wraps a forecasting backbone without modifying it. Inspired by the classical Granger causality test, Matryoshka formulates channel mixing as a selection problem, in which held-out evidence of predictive gain decides where and at what scale channels are mixed. The module arranges channels into nested neighborhoods built from a diffusion geometry of the training data, and keeps a neighborhood only if its shared history improves held-out prediction beyond each channel's own history. Each retained neighborhood receives a residual mixer, a “doll”, whose capacity grows with the neighborhood's size. The dolls start at zero, so training begins exactly at the backbone. Our theoretical analysis shows that the optimal neighborhood balances omitted cross-channel information against an estimation cost that grows linearly with its size. On datasets with hundreds of channels, Matryoshka reduces the MSE of its backbone by 22% on average on the unified benchmark and by 45% on the CI/CD benchmark.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Multivariate Time Series, Channel Mixing, Channel Dependence, Nested Representations, Diffusion Maps

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_N5EqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Matryoshka is a plug-in for multivariate time series forecasting that selects channel interactions using predictive evidence. Training-only calibration constructs a landmark diffusion embedding and a binary hierarchy of channels. Ridge pilots compare neighborhood information against each channel's own lagged history on two chronological evaluation blocks. Accepted fine, coarse, and global scopes receive nested-width, zero-initialized residual mixers, together with a channel-wise self mixer. The paper gives Gaussian-reference risk and independent-evaluation concentration results and evaluates the plug-in across twelve datasets, five backbones in a unified benchmark, and Linear/Transformer backbones in a CI/CD benchmark.

Critical strengths
  1. The paper identifies an important limitation of existing channel interaction strategies in multivariate time series forecasting. Previous approaches typically determine channel interactions based on predefined correlations, learned structures, or clustering mechanisms, without explicitly verifying whether the selected channels provide additional predictive benefits. In contrast, this paper formulates channel mixing as an evidence-based selection problem and introduces a predictive gain criterion to determine whether incorporating a channel neighborhood improves forecasting beyond each channel's own history.
  2. The plug-in design is practically appealing, as the proposed module can be integrated with different forecasting backbones without redesigning the entire architecture.
  3. The experimental evaluation is extensive, covering multiple forecasting benchmarks with different numbers of channels and various backbone architectures. The consistent improvements on high-dimensional datasets demonstrate the effectiveness of adaptive channel interaction, especially when the number of variables is large.
  4. Although the theoretical results rely on simplified assumptions, they offer useful intuition for why channel interaction should not always be expanded globally.
Critical weaknesses
  1. The provided implementation includes an additional validation-driven pruning procedure that removes doll modules according to validation performance. While this procedure itself is reasonable, it is not described in the paper and introduces an additional model selection stage beyond the proposed evidence gate. Therefore, the reported improvements are not fully attributable to the evidence-based qualification mechanism alone, and further ablation studies separating the effects of evidence gating and validation-based pruning would be helpful. Moreover, this additional validation-driven selection may introduce extra model selection overhead compared with baselines that do not perform such post-training structural optimization.
  2. The effectiveness of Matryoshka appears to depend strongly on the underlying backbone. Since channel-dependent architectures already perform channel interaction internally, the additional dolls provide limited gains on models such as iTransformer and DUET.
  3. Since the evidence gate already evaluates whether different channel neighborhoods provide additional predictive benefits, a natural alternative is to select only the most beneficial interaction range for each channel rather than retaining multiple qualified dolls in a nested manner. However, the paper does not provide sufficient theoretical or empirical analysis explaining why combining multiple qualified channel interaction granularities is preferable to selecting a single optimal one. An ablation comparing nested doll usage with single-doll selection would help clarify the motivation of the Matryoshka design.
Questions for authors
  1. The evidence gate is used to determine which channel neighborhoods should receive additional interaction capacity. Could the authors clarify why multiple qualified neighborhoods are combined in a nested doll structure rather than selecting only the most beneficial interaction range for each channel? Have the authors compared the proposed nested design with a single-doll selection strategy?
  2. The implementation appears to include an additional validation-driven pruning procedure after training. Could the authors clarify whether the reported results are obtained before or after this pruning step? If pruning is applied, how much does it contribute to the final performance improvement compared with the evidence gate alone?
5 comments

Response to Reviewer N5Eq

We thank the reviewer for the valuable and constructive feedback.

W1 / Q2. Validation-driven pruning.

  • Why the paper does not describe it. The pruning step was an option that we tried during development. It had almost no effect, so we decided not to make it part of the method, and the paper does not describe it. But we forgot to remove it from the code, so it was still active in the reported runs. This was our mistake, and we apologize.
  • Before or after? The 240 results of the unified benchmark (Tables 1 and 10–14) are therefore after pruning. The step switched off a unit in 3 of the 240 runs. The other 237 are identical with and without it ([Experiment 7]).
  • What the step does. After training, each built doll unit (a level, or one size bucket of a level) is switched off in turn. A unit is removed only if the validation loss falls by more than two standard errors. No weight changes. The step reads only the validation split, which every method already uses for early stopping, and never the test split.
  • What we do now. We remove the step from the code, and the revision reports the results without it.

[Experiment 7]: Validation pruning: effect, cost, and interaction with the gate. We checked the stored pruning mask and the log of every reported run.

(i) Effect on the reported results (12-dataset MSE):

BackboneAlone+ Matryoshka, no pruning+ Matryoshka, as reportedRuns with a unit removed
DLinear0.3560.2620.2612/48
PatchTST0.2860.2460.2460/48
TSMixer0.2700.2460.2460/48
iTransformer0.2620.2510.2510/48
DUET0.2560.2460.2461/48

Pruning switched off a unit in 3 of the 240 runs: DLinear on Traffic at H = 192 and 336, and DUET on PEMS07 at H = 48. In the other 237 runs the pruned and unpruned models are the same model. The three affected cells, without -> with pruning, are 0.465 -> 0.451, 0.483 -> 0.468, and 0.091 -> 0.090. So the step matters in one place, DLinear on Traffic: 0.487 without it and 0.480 with it, against 0.625 for the backbone.

(ii) Gate against pruning, on the four ETT datasets and Weather (the setting of Table 7):

MSEBackboneGate + pruning (reported)Gate onlyPruning onlyNeither
DLinear0.3870.3790.3790.4160.421
TSMixer0.3480.3490.3490.3640.370
PatchTST0.3500.3470.3470.3610.366
iTransformer0.3610.3590.3590.3700.374

Pruning removed nothing in any gated run on these datasets, so the gate alone gives the reported numbers. Pruning cannot replace the gate.

(iii) Cost. The step needs at most 13 forward passes over the validation split and no retraining. It took 7% of the training time over 60 runs. The revised code no longer has this stage.

Summary. The pruning step was left in the code by mistake, and the reported results include it. Its effect is small: it changed 3 of the 240 runs and one four-horizon average, and it cannot replace the gate. The revision removes the step from the code and reports the results without it.

W2. Dependence on the backbone.

We agree, and Section 5 names it as the main limitation. The 4% average on iTransformer and DUET mixes three different cases (Tables 13–14):

MSE change with MatryoshkaSolarPEMS03PEMS04PEMS07PEMS08ElectricityTraffic
iTransformer−15%−29%−16%−15%−30%−3%+10%
DUET−13%−23%−25%−25%−21%+1%+5%
  • On Traffic the dolls hurt. The gate tests against a channel's own history and not against what the backbone already captures. Pruning does not catch it either: it removed nothing in these runs.
  • A gate on the backbone's residuals flags it, but does not fix it ([Experiment 1], qbj4-W1). Before any doll is trained, it accepts 1–2% of the Traffic channels with iTransformer and DUET, against 87% with the default gate. Trained with it, iTransformer on Traffic still gives 0.493, against 0.409 alone.
  • The loss comes from joint training. With the backbone trained alone and then frozen, the dolls give 0.409 with iTransformer and 0.437 with DUET on Traffic, the level of the backbones (0.409 and 0.434).

Summary. The dependence is real, and it is expected: a plug-in that mixes channels has less to add to a backbone that already mixes them. On the seven datasets with hundreds of channels, Matryoshka lowers the MSE by 47% and 28% with the channel-independent DLinear and PatchTST, and by 20%, 9%, and 9% with TSMixer, iTransformer, and DUET. But where the cross-channel structure is strong, these backbones still gain. On PEMS08, iTransformer improves by 30% [Experiment 6]. The revision states this in Section 4 and Section 5.

W3 / Q1. Why nested dolls instead of the single best range per channel?

  • The scales are not interchangeable. The fine doll sends pairwise lead-lag messages inside a group of at most 16 channels. The coarse doll adds the mean of a neighborhood of up to 32. The global doll adds an attention summary of all channels. A single range forces each channel to choose between pairwise detail and wide context.
  • The theory points the same way. Information never decreases as the neighborhood grows (Lemma 2), while in the linear reference of Theorem 1 the cost of estimating pairwise interactions grows with its size. Nesting takes pairwise capacity where it is cheap, and a fixed-capacity summary for the wide ranges.
  • Existing evidence. In Table 6, the fine level alone gives 0.199 on DLinear. Adding the coarse level gives 0.193, and adding the global level gives 0.178.

[Experiment 8]: Nested dolls against a single scale. We ran the comparison you ask for. In "best single scale", each channel keeps only the accepted scale with the largest held-out pilot gain. The innermost doll is kept in every row.

MSEDLinear (7 datasets with hundreds of channels)iTransformer (Solar, Electricity, PEMS08)
Backbone0.3330.195
Fine only0.1990.175
Coarse only0.2270.180
Global only0.1920.183
Best single scale per channel0.2120.179
Nested (ours)0.1780.165

On DLinear, the nested model is better than the best single scale on all seven datasets, and better than every fixed single scale on average. The global doll alone, the strongest single scale, matches the nested model on Electricity and Traffic (0.181 and 0.489 against 0.182 and 0.487). On Solar and the four PEMS datasets the nested model is 3–24% below it. On iTransformer, the nested model is again better than every single scale on average.

Summary. Nesting is better than any single scale overall. Against the alternative you propose, the best single scale per channel, the nested model has a 16% lower MSE with DLinear (0.178 against 0.212) and an 8% lower MSE with iTransformer (0.165 against 0.179). Against the strongest fixed scale it is 7% and 6% lower (global only with DLinear, fine only with iTransformer). The gain is largest where cross-channel structure is strong: 15–24% on the PEMS datasets.

Overall.

  • The results do not depend on pruning. 237 of the 240 runs are identical without it, and no backbone average changes by more than 0.001. The step never read the test split. The revised code does not contain it, and the revision reports every result without it.
  • Backbones that mix channels still gain where the structure is strong. On Solar and the four PEMS datasets, Matryoshka lowers the MSE of iTransformer and DUET by 13–30%. The clear losses are on Traffic.
  • Nesting is better than a single scale. The nested model has a 16% lower MSE than the best single scale per channel with DLinear and an 8% lower MSE with iTransformer. It is also better than every fixed single scale on average.

We hope that the full account of the pruning step and the two new experiments address your concerns, and we would be grateful if you considered them in your final assessment.

New experiments referenced: [Experiment 7] Validation pruning: effect, cost, and interaction with the gate · [Experiment 8] Nested dolls against a single scale

I appreciate the additional experiments, but I remain unconvinced by the rebuttal. The authors have partially addressed my questions, but several concerns remain unresolved:

  1. The proposed method still lacks consistent effectiveness on models that already incorporate channel interactions

  2. The experimental evidence remains insufficient to establish the distinct contribution of the proposed core mechanisms, much of the performance improvement may come from introducing additional channel interactions into the original backbone, while the proposed evidence gating and hierarchical structure provide relatively limited additional gains.

The current evidence is still insufficient to justify acceptance, and I will maintain my original rating.

Reviewer_VJtgOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Matryoshka, a plug-in channel-mixing module for multivariate time series forecasting. The method constructs nested channel neighborhoods from training data and uses held-out predictive-gain tests to determine whether and at which scales cross-channel interactions should be introduced. Qualified neighborhoods are equipped with residual mixing modules whose capacities adapt to neighborhood size.

The authors also provide theoretical analysis and evaluate Matryoshka across multiple forecasting datasets, backbones, and existing channel-interaction plug-ins. The paper reports reductions in forecasting error and includes ablation studies to investigate the contributions of its main components.

Critical strengths

S1. A meaningful research problem. The paper addresses the observation that cross-channel interaction is not always beneficial and proposes using predictive evidence to determine whether channel mixing should be applied.

S2. An interesting architectural framework. Matryoshka combines predictive-gain testing, nested channel neighborhoods, and residual mixing modules into a unified framework for modeling channel interactions at different scales.

S3. Relatively comprehensive analysis. The paper includes theoretical results, experiments across multiple forecasting backbones, and component-level ablations that help explain the motivation and behavior of the proposed method.

Critical weaknesses

W1. Insufficient hyperparameter tuning undermines the fairness of the main experimental comparisons. The paper explicitly states that LIFT and CCM are evaluated using their published default configurations without per-cell hyperparameter tuning. Although these configurations originate from official implementations, default hyperparameters may not be optimal under the current unified evaluation protocol or across different forecasting backbones. Therefore, the experiments do not establish that these baselines received reasonable optimization opportunities. Since the main conclusions rely heavily on Matryoshka's reported advantages over existing plug-in methods, this limitation directly undermines the reliability of the performance comparisons.

W2. The independent contributions of the core design choices require stronger validation. Matryoshka combines evidence-based gating, nested neighborhoods, and residual mixing modules with varying capacities. Although the paper provides ablation studies, the current evidence does not fully distinguish whether the reported improvements arise from the predictive-evidence selection mechanism or from the additional residual modeling capacity. More direct controlled comparisons with alternative channel-selection mechanisms under comparable parameter budgets would help establish the necessity of the proposed design.

W3. A gap remains between the theoretical analysis and the practical model. The main expected-risk decomposition is derived under a linear Gaussian setting and corresponding independence assumptions, whereas the practical method involves data-driven neighborhood construction and nonlinear residual mixing modules. Consequently, the theoretical results primarily motivate the design rather than directly guaranteeing the generalization advantages of the implemented Matryoshka framework. The scope and limitations of these theoretical conclusions should be more clearly established.

Questions for authors

Q1. Section 4.1 explicitly states that LIFT and CCM use their published default configurations without per-cell hyperparameter tuning. Why were these baselines not given reasonable hyperparameter searches under the current experimental protocol? How can the authors establish that the reported performance advantages are not influenced by insufficient baseline tuning?

Q2. Could the authors provide more direct controlled comparisons between predictive-evidence-based neighborhood selection and alternative channel-selection mechanisms under comparable parameter budgets and training protocols? This would help isolate the independent contribution of the evidence-gated design.

Q3. The theoretical analysis is primarily developed under linear Gaussian assumptions, while the practical method uses a data-driven hierarchy and nonlinear mixing modules. Could the authors further clarify the extent to which the theoretical results apply to the implemented model and discuss their limitations?

3 comments

We thank the reviewer for the detailed and constructive feedback.

W1 / Q1. LIFT and CCM were run with published defaults.

  • What we ran, and why. LIFT and CCM were proposed on DLinear, PatchTST, and TSMixer, which is why these are three of our five backbones. Their published numbers cannot be placed in our tables: LIFT uses a lookback of 336, and CCM does not state its lookback and does not cover Solar or PEMS. We therefore re-ran both under our protocol with the defaults of their official code, one configuration per plug-in. We chose the defaults because their authors set them for the same backbones and for datasets that overlap with ours, so we expected them to stay close to tuned at our lookback of 96, since they were tuned by the authors of the original works. [Experiment 5] now tests this. For xCPD, whose plug-in code is not released, we compare with its published numbers (Table 16).
  • Your concern is justified for CCM and for one LIFT setting. We compared the gain each baseline reaches in our tables with the gain its own paper reports, on every backbone–dataset pair that both cover:
MSE change against the backbonePairsPublished gainGain in our tablesMatryoshka, same pairs
LIFT (published at lookback 336)6−3.9%−2.7%−9.8%
CCM14−1.5%+1.3%−4.6%

LIFT fails in one pair, PatchTST on Traffic (+19% in our tables, −2.6% in its paper). CCM does not reach its published gains with the code defaults: its paper selects β\beta and the number of clusters per dataset.

[Experiment 5]: Tuned LIFT and CCM. Both baselines are searched over the spaces of their own papers and official scripts and selected on validation loss, per backbone and dataset, at the first horizon. The selected setting is then run at the other horizons. LIFT: K∈1,4,8,16K \in \\{1, 4, 8, 16\\}, N∈1,4,8,16N \in \\{1, 4, 8, 16\\}, temperature ∈1,0.1\in \\{1, 0.1\\}, giving 9 or 10 settings per dataset, including the default and the values of its official scripts. CCM: clusters ∈2,4,8\in \\{2, 4, 8\\}, β∈0.3,0.5,0.9\beta \in \\{0.3, 0.5, 0.9\\}, giving 6 settings, including the default and the values of its Table 10. Six datasets: ETTh1, ETTm1, Weather, Solar, Electricity, and PEMS08. Matryoshka is not searched: it uses one configuration for all datasets.

(i) Tuned baselines (6-dataset MSE):

BackboneLIFT defaultLIFT tunedCCM defaultCCM tunedMatryoshka
DLinear0.3320.2990.2950.3350.3330.263
TSMixer0.2820.2770.2750.2880.2850.259

The defaults were already close to tuned. The search selects a setting other than the default in 23 of the 24 cases (two backbones, two plug-ins, six datasets), but it changes the six-dataset MSE by at most 0.004. The largest change on one dataset is 0.017 (LIFT with DLinear on Solar, 0.268 -> 0.251), and in 5 cases the selected setting is worse on the test split than the default. Our protocol, with its lookback of 96, therefore did not leave the baselines untuned. The baselines have the larger search in this comparison: 9 or 10 settings for LIFT and 6 for CCM, for every backbone and dataset. Matryoshka has no per-dataset setting at all.

(ii) LIFT with the protocols of its official PatchTST scripts: backbone pre-trained and frozen, learning rate ∈{0.001,0.005}\in \{0.001, 0.005\}, and on Traffic also joint training with K = 8 or 4 and N = 1. Best on validation, first horizon:

MSE at H = 96BackboneLIFT, default (Table 11)LIFT, official protocolMatryoshka
Weather0.1730.1610.1690.161
Electricity0.1600.1520.1570.138
Traffic0.4160.5240.4250.419

On Weather and Electricity, LIFT is better in our tables than with its official protocol. On Traffic you are right: its official settings give 0.425. The 0.524 of Table 11 comes from joint training with the same default for every dataset (N = 8) and not from LIFT itself. The revision reports 0.425.

  • Size of the gap. Over the 240 cells of Tables 10–14, Matryoshka has a lower MSE than LIFT in 196, ties in 9, and loses in 35. The average over the five backbones and 12 datasets is 0.276 for LIFT, against 0.250 for Matryoshka.
  • Where a baseline failed, we corrected it in its favor. In the submission, CCM on iTransformer/PEMS04 uses a learning rate of 1e-4 instead of the backbone's 5e-4, because at 5e-4 it diverges and predicts the mean (MSE 0.534 at H = 12, against 0.077 at 1e-4). The revision does the same for LIFT on PatchTST/Traffic (0.524 -> 0.425, part (ii) above).

Summary. Your concern was justified in part. The defaults of LIFT and CCM were close to tuned in our protocol: a search over 6 to 10 settings per backbone and dataset moves them by at most 0.004. Against the tuned baselines, Matryoshka stays 0.032 ahead on DLinear with one configuration for all datasets. Two things change. LIFT's PatchTST/Traffic cell becomes 0.425. And on TSMixer, Matryoshka's lead over tuned LIFT is 0.016. The revision reports the tuned baselines and the corrected cells.

W2 / Q2. Evidence-based selection or extra residual capacity?

  • Direct answer. Both of them contribute. Mixing still provides much gain further over capacity.

[Experiment 6]: Selection rules at a matched budget, and a capacity control. These are the two controls that were missing. The dolls, widths, and training are identical in every row.

(i) Capacity control. All mixing dolls are removed, and a channel-independent residual MLP with their number of parameters is added:

MSEBackboneCapacity only, no mixingMatryoshkaMatryoshka against capacity only
DLinear, Electricity0.2100.1850.182−2%
DLinear, Traffic0.6250.5060.487−4%
DLinear, Solar0.3260.2200.205−7%
DLinear, PEMS030.2640.1340.095−29%
DLinear, PEMS040.2640.1470.091−38%
DLinear, PEMS070.3110.1430.077−46%
DLinear, PEMS080.3320.1560.110−30%
iTransformer, Solar0.2430.2100.207−1%
iTransformer, PEMS080.1650.1490.116−22%

You are right that capacity explains some part of the gain. However, capacity alone is not enough: it is worse than Matryoshka in all nine rows.

(ii) Selection rules. Only the rule that accepts neighborhoods changes. Each count-matched rule accepts as many neighborhoods as the gate, so the parameter count is equal:

MSE, DLinearETT + WeatherElectricityTraffic
Backbone alone0.3870.2100.625
Evidence gate (ours)0.3790.1820.487
Random acceptance, same count (3 seeds; range)0.382 (0.380–0.385)0.183 (0.181–0.184)0.504 (0.494–0.513)
Correlation ranking, same count0.3810.1830.489
Lead-lag ranking (LIFT-style), same count0.3820.1820.524
Learned end-to-end gate0.4140.1820.490
No gate (all candidates)0.4210.1870.503

Which neighborhoods are mixed matters on Traffic: the MSE of the gate is 4% below random acceptance, 8% below lead-lag ranking, and 1% below a learned gate. How much is mixed matters on the few-channel datasets: a learned gate (0.414) and no gate (0.421) are worse than the backbone alone (0.387), while the rules that inherit the gate's count stay within 1% of it. The gate's most reliable contribution is therefore the budget.

  • More mixing capacity does not help (Tables 6–7, DLinear, large datasets). Building every candidate gives 0.181, full widths give 0.181, and a random tree with the same group sizes gives 0.184, against 0.178.
  • The gate matters less on the other backbones. On the few-channel datasets, the no-gate control gives 0.370 against 0.349 with the gate on TSMixer, 0.366 against 0.347 on PatchTST, and 0.374 against 0.359 on iTransformer ([Experiment 7], N5Eq-W1/Q2).

Summary. Both of your alternatives are partly right. Some part of the gain is added per-channel capacity. Among selection rules of equal size, the evidence gate is best or tied, but it is clearly better only on Traffic, and its main contribution is deciding how much to mix. The revision adds both controls to Table 6.

W3 / Q3. How far does the theory apply to the implemented model?

  • Direct answer. We agree with your reading. The theory applies to one part of the implemented model, the gate, and it motivates the design. It does not guarantee the error of the trained model, and we do not claim that it does:

  • Lemma 2 (Appendix B.1)

    • What it says: With the best possible predictor, more channels never raise the risk
    • What it assumes: Finite second moments only
    • Does it apply to the implemented model: Only as a statement about the optimum. It says nothing about a fitted model
  • Theorem 1

    • What it says: For a fitted linear predictor, the expected risk is the omitted information plus an estimation cost. As the neighborhood grows, the first can only fall, while the factor p/(n−p−1)p/(n − p − 1) of the second grows (p predictors, n fitting origins)
    • What it assumes: Jointly Gaussian data, i.i.d. fitting origins, OLS, a chain fixed before fitting
    • Does it apply to the implemented model: No, it is a reference case. It does not cover the learned hierarchy, the ridge pilots, dependent origins, or the neural dolls
  • Theorem 2

    • What it says: The held-out gain that the gate measures is close to its true value for all candidates at once, and the error depends on the number of channels only through its logarithm
    • What it assumes: Conditions on the fitting segment, held-out origins i.i.d., sub-gamma losses, own-history risks bounded away from zero. No Gaussian or linear assumption on the data
    • Does it apply to the implemented model: In part. It covers the data-driven hierarchy, the ridge pilots, and the top-down search of the tree, because all three are fixed once the fitting segment is fixed. It does not cover serial dependence between held-out origins, the forecast-horizon test (Appendix B.3), the margin τ\tau (the bound exceeds 0.3 at our block sizes, against τ=0.01\tau = 0.01), or the trained dolls
  • Why a linear Gaussian reference. It is the setting in which the trade-off can be written exactly (Equation 9), and the gate's pilot is also linear. Its purpose is to explain why a fitted model can lose from more channels, although the optimal risk never rises (Lemma 2). This is our reason to test every neighborhood instead of mixing by default.

  • Theorem 2 applies to the gate as implemented at the pilot horizon: the learned hierarchy and the ridge pilots are the ones that the code runs. We did not make this clear. It justifies the form of the gate, a tree of candidates tested on held-out data, and not its margin.

  • Nothing covers the nonlinear dolls. We claim no generalization guarantee for the trained model. For the trained model the evidence is experimental ([Experiment 6] above).

  • The gate was checked where the assumptions fail ([Experiment 4], qbj4-W4). On independent synthetic channels, the unchanged calibration accepts none of 32,400 fine candidates. With planted leader–follower groups it recovers 53–55% of the followers. On real data, the held-out pilot error first falls and then rises with the neighborhood on Electricity and Traffic, and keeps falling on PEMS03. That is the trade-off of Theorem 1 in direction. In size, the OLS error grows much faster than the Gaussian formula. The trained model shows the same direction: with DLinear on the few-channel datasets, mixing every candidate is worse than the backbone alone (last row of [Experiment 6] (ii)).

Summary. The theory is motivation for the gate, and only Theorem 2 concerns the gate as implemented. The submission stated these limits in Section 3.5 and in Appendix B.2–B.3. The revision renames Section 3.5 to "Theoretical motivation".

Overall.

  • The comparison holds against tuned baselines. A search over 6 to 10 settings per backbone and dataset moves LIFT and CCM by at most 0.004. Matryoshka stays 0.032 ahead of tuned LIFT on DLinear and 0.016 ahead on TSMixer, with one configuration for all datasets.
  • Mixing adds to capacity. Against a channel-independent module with the same number of parameters, Matryoshka is better in all nine cases, by 29–46% on the four PEMS datasets with DLinear and by 22% with iTransformer on PEMS08.
  • The gate decides how much to mix, and this protects the backbone. Without it, Matryoshka is worse than the backbone on the few-channel datasets (0.421 against 0.387 with DLinear, 0.370 against 0.348 with TSMixer, Table 7). With it, the MSE is 0.379 and 0.349. No selection rule of the same size is better than the evidence gate in [Experiment 6].

We hope that the tuned baselines, the two controls, and the stated scope of the theory address your concerns, and we would be grateful if you considered them in your final assessment.

New experiments referenced: [Experiment 5] Tuned LIFT and CCM · [Experiment 6] Selection rules at a matched budget, and a capacity control

Reviewer_qbj4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Matryoshka, a plug-in channel-mixing module for time-series forecasting. It constructs a nested hierarchy of channel groups using correlation and lead-lag geometry, then employs held-out linear pilot models to determine whether each group provides predictive information beyond a channel’s own history. Accepted groups define fine-to-global residual channel mixers that are attached to different forecasting backbones. The paper evaluates the method with five backbones across twelve datasets and reports particularly strong improvements on high-dimensional datasets. The evidence-based selection of channel neighborhoods is interesting and the evaluation is unusually broad, although the statistical gate and theoretical analysis do not fully correspond to the claims or the final nonlinear forecasting system.

Critical strengths
  1. The paper tackles a meaningful problem in channel-dependent forecasting. Instead of assuming that full channel independence or unrestricted channel mixing is universally preferable, Matryoshka attempts to determine empirically when and at what scale cross-channel information is useful.
  2. The nested design provides a clear inductive bias. Organizing mixers from local channel neighborhoods to global groups offers a structured alternative to a single dense channel-mixing layer and avoids treating all channel dependencies as equally informative.
Critical weaknesses
  1. The evidence gate does not test incremental information relative to the actual forecasting backbone. It compares linear group-history pilots against linear own-history pilots. A group can therefore pass because it helps a simple linear predictor even when the same information is already captured by a channel-dependent backbone. This substantially weakens the interpretation of the gate as selecting information that is useful to the final model and is consistent with the much smaller gains on stronger channel-dependent backbones.
  2. The acceptance statistic is group-averaged rather than channel-specific. A neighborhood is retained based on its average improvement across member channels, even if it harms or provides no benefit to some individual channels. The resulting construction therefore does not fully support the paper’s language of evidence-based selection beyond each channel’s own history.
  3. The hierarchy and accepted groups are static for the entire dataset. They cannot accommodate regime-dependent, seasonal, or window-specific changes in inter-channel relationships. This is a substantial restriction for non-stationary forecasting settings, where the usefulness and direction of cross-channel information may change over time.
  4. The theoretical results have limited correspondence to the implemented method. The analysis assumes a fixed hierarchy, Gaussian linear regression, OLS estimation, and effectively independent validation origins. The actual pipeline estimates the hierarchy from the data, uses ridge pilots, operates on strongly overlapping time-series windows, and trains nonlinear residual mixers jointly with the backbone. The theoretical guarantees therefore do not directly explain the behavior of Matryoshka as implemented.
  5. The empirical benefit is concentrated in high-channel spatial datasets and is not universal. Improvements on several low-dimensional datasets and strong channel-dependent backbones are small, with occasional degradations. Thus, the experiments support Matryoshka as a useful specialized augmentation when substantial exploitable cross-channel structure exists, but not the broader suggestion of a generally effective plug-in across forecasting settings.
Questions for authors

See Weaknesses.

6 comments

We thank the reviewer for the thoughtful and constructive feedback.

W1. The gate tests against own history, not against the backbone.

  • We agree. It is the price of the design choice. Section 5 names it as the main limitation. We test against own history so that the gate is backbone-agnostic, cheap, and fast. It reads only the training split, runs on a CPU in few seconds per dataset (0.1-2 seconds, even for high-channel datasets), and finishes before any training. One result serves all five backbones. A test against the backbone needs the backbone's out-of-sample errors, so the backbone must first be trained alone, for every backbone, dataset, and horizon.
  • Where it costs. On Traffic, Matryoshka is worse than its backbone with iTransformer (+10% MSE), DUET (+5%), and PatchTST (+1%).

[Experiment 1]: Backbone-aware gate. We built the gate you describe. The backbone is trained alone and run on the validation split, then the gate's two-block test is applied to its out-of-sample residuals. It asks what a neighborhood adds beyond the backbone. The hierarchy, the margin, and the dolls are unchanged.

(i) What the gate accepts before any doll is trained (share of channels in accepted fine groups, first horizon):

Default gate (any backbone)Backbone-aware: DLinearTSMixerPatchTSTiTransformerDUET
Traffic87%35%18%2%1%2%
Electricity84%6%2%1%3%1%
Solar100%53%0%6%0%0%
PEMS0399%100%67%96%86%63%
PEMS04100%98%68%97%63%95%
PEMS07100%100%42%94%72%89%
PEMS0895%97%68%88%71%74%

The gate flags the failures before training: the four pairs in which Matryoshka loses to its backbone (Traffic with PatchTST, iTransformer, and DUET, and Electricity with DUET) receive 1–2%. Over all pairs, the accepted share ranks the MSE change of the reported runs with a Spearman correlation of −0.85, against −0.43 for the default gate. The gate also agrees with the capacity control of [Experiment 6] (VJtg-W2/Q2). iTransformer gains from mixing on PEMS08 (0.149 with the same parameters and no mixing, 0.116 with mixing), where the gate accepts 71%.

(ii) Training with the backbone-aware gate, and (iii) training the dolls on a frozen backbone:

Backbone aloneDefault gate (reported)(ii) Backbone-aware gate(iii) Backbone frozen, default gate
iTransformer, Traffic0.4090.4500.4930.409
DUET, Traffic0.4340.4580.4470.437
PatchTST, Traffic0.4420.4470.4470.442
iTransformer, Electricity0.1780.1720.1760.173
iTransformer, Solar0.2430.2070.2310.218
iTransformer, PEMS080.1650.1160.1110.151

The backbone-aware gate does not fix the problem (ii). With iTransformer on Traffic the result is worse, although only 10 of the 862 channels are still mixed, so the loss is not caused by the accepted neighborhoods. It comes from joint training (iii): in the two losing cases the joint model fits the training data worse than the backbone alone (training MSE +16% with iTransformer, +3% with DUET). With the backbone frozen, the result returns to the level of the backbone alone for all three (within 0.003), so the dolls add nothing there, as you predict. Freezing is a diagnostic only: on Solar and PEMS08, much of the gain needs joint training.

Summary. The own-history test is a design choice: it keeps the gate backbone-agnostic and costs seconds on a CPU. Its price is the one you name: it cannot see what the backbone already captures. [Experiment 1] measures this price. A gate on the backbone's residuals flags every losing case of the reported runs before training, but it does not remove the Traffic losses, because they come from joint training and not from the accepted groups. The revision states the limitation in Section 3.3, adds the backbone-aware gate as a diagnostic, and reports the Traffic losses next to the averages.

W2. The statistic is averaged over the group, not channel-specific.

  • Correct. Appendix B.3 says that the score "does not ensure a gain for every channel". We move this to Section 3.3 and replace "beyond each channel's own history" with "on average over the neighborhood's channels".
  • Why the group is the unit. A channel with no gain of its own can be the source of the others' gain: a leading sensor gains nothing from its followers.

[Experiment 2]: Channel-level evidence. On the seven datasets with hundreds of channels, we counted the channels in accepted fine groups that gain individually, and we trained a channel-specific gate.

(i) Share of the channels in accepted fine groups that gain individually:

SolarElectricityTrafficPEMS03PEMS04PEMS07PEMS08
Held-out pilot gain of their own, in both blocks (calibration)100%67%73%97%98%98%98%
Lower test MSE than the backbone, DLinear100%80%99%99%100%100%99%
Lower test MSE than the backbone, iTransformer96%76%7%92%87%81%90%

Your concern is real on Electricity and Traffic: only 67–73% of the member channels have a held-out gain of their own there. On the test split, most of them still improve with DLinear (80% and 99%). With iTransformer on Traffic the loss is not a selection effect: the MSE rises by 11% for the channels inside accepted fine groups and by 8% for those outside, so it follows from W1.

(ii) Channel-specific gate. A channel receives a doll's residual only if its own gain passes the margin in both blocks, while every member still sends:

SolarElectricityTrafficPEMS03PEMS04PEMS07PEMS08Average
Group gate0.2050.1820.4870.0950.0910.0770.1100.178
Channel-specific gate0.2050.1850.5030.1010.0950.0840.1160.184

The channel-specific gate is worse on six datasets and equal on Solar. A channel without a measured gain of its own still benefits from the trained doll.

Summary. Yes, the statistic is a group average, and our wording claimed more, so we correct it. The group is still a better unit: a per-channel rule removes receivers that do benefit and raises the MSE by 3%.

W3. The hierarchy and the accepted groups are static.

  • Correct. Membership is fixed for the dataset. Only the mixing weights change with the window (the tanh term of Equation 5 and the attention of Equation 8).
  • The test selects relations that persist. A gain must hold on two consecutive held-out blocks, each 20% of the training period: one to two weeks on PEMS, and seven weeks to five months elsewhere.

[Experiment 3]: Stability over time.

(i) Persistence. The structure and the pilots are fixed on the training split, and the pilot gain of each accepted neighborhood is measured on later data (twelve datasets):

Keep a positive pilot gainValidation periodTest period
Accepted fine groups (296)94%88%
Accepted coarse neighborhoods (109)96%90%

Most accepted relations persist beyond the data that selected them.

(ii) MSE reduction over the backbone in five consecutive fifths of the test split (seven large datasets):

1st2nd3rd4th5th
DLinear41%48%44%43%48%
iTransformer12%14%9%10%13%

The gain does not decay with the distance from the training period.

(iii) Without the window-adaptive weights (MSE of DLinear):

SolarElectricityTrafficPEMS03PEMS04PEMS07PEMS08Average
Default0.2050.1820.4870.0950.0910.0770.1100.178
Static mixing weights0.2050.1780.5020.0910.0910.0770.1160.180

The average is 0.178 against 0.180. Only Traffic clearly depends on these weights (0.502 against 0.487). The other differences are at most 0.006, in both directions. Almost all of the gain comes from relations that persist, so we do not present the window-adaptive weights as an answer to your concern.

Summary. The structure is static, and on these benchmarks that is sufficient: the accepted relations persist into the test period, and the gain is stable across it. Matryoshka cannot use a relation that appears only after training or that changes with the regime.

W4. The theory does not match the implemented pipeline.

We agree. Point by point:

  • Hierarchy estimated from the data. Covered by Theorem 2 at the pilot horizon: it conditions on the fitting segment, which fixes the hierarchy. Not covered by Theorem 1 (fixed chain), and not covered at the forecast horizon (Appendix B.3).
  • Ridge pilots instead of OLS. Covered by Theorem 2, which holds for any pilot fitted on the fitting segment. Theorem 1 is OLS only.
  • Dependent origins. Not covered. The pilot's origins are 4–18 steps apart and are not the overlapping 96-step windows, but serial dependence remains (Appendix B.3).
  • Nonlinear dolls trained jointly. Not covered by any result, and we do not claim any.

[Experiment 4]: The gate and Theorem 1 in the implemented regime. We checked the gate where the assumptions fail.

(i) Synthetic data, run through the unchanged calibration (learned hierarchy, ridge pilots, serially dependent origins):

SettingResult at the margin of 0.01
Independent autocorrelated channels (32 to 512 channels, 270 datasets)0 of 32,400 fine, 1,080 coarse, and 270 global candidates accepted. With the margin set to zero: 0.18% of the fine and 1.1% of the global candidates
Planted groups of one leader and seven followers (lead of 24 to 28 steps)53–55% of the followers accepted in a group with another member of their planted group
Independent channels next to the planted groups10–22% end up in an accepted group, always one that contains planted channels

The gate does not accept noise and is conservative on true structure. Its group average also lets in some unrelated channels, which is the effect of your W2.

(ii) Real data. We follow 64 channels up the hierarchy and report the held-out error of each channel's pilot relative to its own-history pilot (median: below 1 means the neighbors help, 1 means no gain, above 1 means it hurts):

Share of all channels in the neighborhood1/641/321/161/81/41/2all
Electricity, ridge (the gate's pilot)0.950.920.900.890.860.890.92
Traffic, ridge0.920.930.940.961.001.101.09
PEMS03, ridge0.770.660.540.440.380.350.32
Electricity, OLS0.890.890.890.961.183.31not defined
Traffic, OLS1.001.181.573.3021.1not definednot defined
PEMS03, OLS0.730.640.570.520.651.36not defined

On Electricity the error first falls and then rises with the neighborhood, as Theorem 1 describes. On Traffic it is lowest for small neighborhoods and exceeds the own-history error from half of the channels on. On PEMS03 it keeps falling up to the full set. The OLS cost is far larger than the Gaussian value p/(n−p−1), so Theorem 1 gives the direction and not the size.

Summary. The theory motivates the gate and guarantees nothing about the trained model. Therefore, the revision renames Section 3.5 to "Theoretical motivation". Empirically, the gate does not accept noise, and the pilot error follows the direction of the trade-off of Theorem 1.

W5. The benefit concentrates in high-channel datasets and is not universal.

  • We agree. Where the gate finds no evidence, it does not mix. On the ETT datasets it accepts at most two pairs of channels (Table 5). Matryoshka changes the MSE by −1% on the ETT datasets and Weather (LIFT: +4%), and by −22% on the seven datasets with hundreds of channels (LIFT: −9%).
  • This is expected for a plug-in that mixes channels. Its gain can only come from other channels, and seven channels offer little to use. But the number of channels does not decide it: Traffic has 862 channels and the smallest gain of the large datasets (−3% on average over the backbones), while PEMS08 has 170 and the largest (−43%). What decides it is the cross-channel structure and shared information in data, which the gate measures.
  • On the few-channel datasets, Matryoshka is still safer than the other plug-ins. It raises the MSE of its backbone in 9 of the 25 backbone–dataset pairs, by at most 2.6%. LIFT does so in 17 pairs, by up to 22%, and CCM in 19. Over the 100 cells of these datasets, Matryoshka has a lower MSE than LIFT in 71 and than CCM in 63.
  • The benefit is predictable before training. The share of channels in accepted fine groups ranks the twelve datasets by the actual MSE reduction with a Spearman correlation of 0.85.

Summary. We accept your characterization. The revision presents Matryoshka as a plug-in for data with usable cross-channel structure that does not mix elsewhere.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.