acceptodds
Under review as a conference paper at ICLR 2027

MUSE: Dependency-Aware Adaptation of a Frozen Vision Backbone for Multivariate Time Series Forecasting

Abstract

Multivariate time-series forecasting is essential to many real-world applications. Recent large vision models (LVMs) offer a promising paradigm by transferring cross-domain visual priors to time-series forecasting. However, existing LVM-based methods face two key challenges: balancing independent visual representation spaces with cross-variable dependency modeling, and adapting vision backbones pretrained on natural images to the distinct temporal semantics of time-series images. To address these challenges, we propose MUSE, a dependency-aware adaptation framework built on a fully frozen pretrained MAE. First, the Variable Context Refinement Module (VCR) aggregates shared temporal information within each variable and models cross-variable contextual dependencies while preserving independent visual spaces. Second, the Temporal-Periodic Refinement Module (TPR) performs lightweight refinement at different encoder depths and explicitly models across-period temporal dependencies and within-period periodic dependencies. The two modules independently produce forecasts, which are fused through a learnable prediction-level gate. Experiments on 10 real-world datasets demonstrate that MUSE achieves state-of-the-art performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_J9p9OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses multivariate long-term forecasting with a frozen ImageNet MAE, following the VisionTS imaging pipeline. It identifies two problems: channel-independent imaging ignores cross-variable dependencies, and the two image axes (across-period and within-period) carry temporal meaning that a natural-image MAE does not model. MUSE adds two trainable branches around the frozen MAE. VCR uses a TimeXer-style channel token per variable, inter-variable self-attention, and an additive per-variable correction before the encoder. TPR inserts AdaptFormer-style bottleneck adapters with axial temporal and circular periodic attention after the encoder blocks. The two branch forecasts are fused by a scalar gate. Experiments cover 10 TFB datasets at 4 horizons against 8 baselines (Table 1, Table 4), with an ablation on 4 datasets at H = 720 (Table 2) and a sensitivity study (Fig. 3).

Critical strengths
  1. The trade-off between channel-independent imaging (VisionTS, DMMV) and joint imaging (VisionTS++, Time-VLM) is stated clearly, and adding cross-variable context outside the frozen backbone is a reasonable way to keep the pre-trained representation intact.
  2. TPR's directional bias on the temporal axis and circular bias on the phase axis (Eqs. 9-10) are a sensible structural prior for period-folded images.
  3. The method is specified in enough detail to reimplement (Alg. 1, App. B.2), code is provided, the tables are internally consistent, and the evaluation spans 10 TFB datasets.
Critical weaknesses
  1. The "structural semantic gap" is asserted but not demonstrated. The paper's motivation is that a frozen natural-image MAE cannot model the temporal meaning of the two image axes, but no analysis (attention maps, probing, or a failure case of plain VisionTS) shows this gap or shows that TPR closes it. A related conceptual question is left open: once trainable adapters are added inside the backbone, what does the frozen vision prior still contribute compared with a strong forecaster trained from scratch?
  2. The cross-variable claim is not tested directly. VCR adds one additive vector per variable, broadcast over all patches (Eq. 5). It is not clear that this static correction can represent lagged or time-varying dependencies, which are what multivariate methods usually target. On Electricity (V = 321), where dependencies should matter most, MUSE is roughly on par with DMMV-A. A channel-shuffle test or an analysis of the inter-variable attention would make the claim concrete.
  3. Positioning and novelty are limited. VCR follows TimeXer, TPR follows AdaptFormer with axial attention, and the method combines them with a prediction-level gate. The closest related methods (VisionTS++, Time-VLM, TimeXer) are not compared, and only full-shot results are reported even though transfer is the usual reason for using a frozen pre-trained backbone. The modules are called "lightweight", but each branch runs the MAE, and no cost numbers are given.
Questions for authors
  1. What evidence shows that the frozen MAE fails to model the across-period and within-period axes, and that TPR fixes this? Can you show attention patterns or a probing result?
  2. What does the frozen vision backbone contribute beyond VCR + TPR? For example, how does MUSE perform with a randomly initialized ViT of the same size, or compared with TimeXer?
  3. Can VCR's per-variable additive correction capture lagged cross-variable dependencies? How does performance change if the variable order is shuffled or the inter-variable attention is removed?
7 comments
AuthorsOpenReview ↗

Response to Reviewer J9p9 (Part I)

We thank Reviewer J9p9 for the careful review and specific suggestions. In response, we conducted additional experiments and analyses.

W1.1 & Q1: Lack of direct evidence for the “structural semantic gap” in the frozen MAE’s modeling of the temporal and phase axes and for TPR’s ability to bridge it.

We thank the reviewer for helping us strengthen the paper by asking us to clarify the structural semantic gap and provide further evidence for the role of TPR. The structural semantic gap refers to the different meanings of the two axes in natural images and time-series images. During natural-image pretraining, MAE learns spatial relationships among image patches. In a time-series grid formed by folding the sequence into periods, the horizontal axis tracks the same phase across successive historical periods, while the vertical axis indexes different phases within a period. After resizing and patch embedding, the historical time and phase axes retain these meanings at a coarser resolution. TPR models temporal and phase dependencies separately along the two axes.

To test whether this explicit modeling improves forecasting, we compare MUSE with only the TPR branch against VisionTS. Table 2 (lines 397–407 of the original paper)‾\underline{\text{Table 2 (lines 397–407 of the original paper)}} already reports the results for MUSE with TPR alone on ETTh2, Weather, AQWan, and ZafNoo at a forecast horizon of H=720H=720. The table below reproduces those results alongside the VisionTS results for the same datasets and forecast horizon.

Comparison of MUSE with only the TPR branch and VisionTS. Each cell reports MSE / MAE. Lower is better, and the better result for each metric is in bold.

DatasetMUSE (w/o VCR)VisionTS
ETTh20.380 / 0.4220.387 / 0.426
Weather0.309 / 0.3300.366 / 0.368
AQWan0.899 / 0.5350.865 / 0.524
ZafNoo0.571 / 0.4710.575 / 0.477
Average0.540 / 0.4400.548 / 0.449

As shown in the table, MUSE with TPR alone achieves lower MSE and MAE than VisionTS on ETTh2, Weather, and ZafNoo, with the largest improvements on Weather. Across the four datasets, mean MSE decreases from 0.548 to 0.540 and mean MAE from 0.449 to 0.440. This comparison provides evidence at the forecasting level that explicitly modeling temporal and phase dependencies along the two axes helps adapt the frozen MAE’s visual representations to time-series forecasting. We will include this comparison and analysis in the final version of the paper.

AuthorsOpenReview ↗

Response to Reviewer J9p9 (Part II)

W1.2 & W3 & Q2: The added value of the frozen visual prior remains unclear, comparisons with VisionTS++, Time-VLM, and TimeXer are missing, and the full-shot setting and the cost of running both branches need clarification.

We thank the reviewer for raising specific questions about the contribution of the frozen visual prior, comparisons with related methods, and MUSE’s training setting and parameter overhead.

Through masked reconstruction pretraining on a large collection of natural images, MAE learns rich visual priors about local shapes, spatial layouts, and recurring patterns. It learns to use structural cues from visible patches to reconstruct masked regions. When a time series is arranged into a 2D image by period, local correlations, trends, and recurring patterns can appear as spatial structures. MUSE transfers MAE’s pretrained reconstruction capability to forecasting by reconstructing the masked future region from the visible historical region. VCR and TPR adapt the pretrained representations to cross-variable context and temporal–periodic dependencies, respectively.

PatchTST, for example, learns forecasting mappings directly in the numerical domain. As shown in Table 1 of the original paper‾\underline{\text{Table 1 of the original paper}} and Table 4 in Appendix C.2‾\underline{\text{Table 4 in Appendix C.2}}, MUSE achieves lower average MSE and MAE than PatchTST on all 10 datasets. These results support the effectiveness of combining a frozen visual backbone with adaptation tailored to time-series dependencies.

We further compared MUSE with the three methods suggested by the reviewer, VisionTS++, TimeXer, and Time-VLM, on six shared datasets. Each cell in the two tables below reports MSE / MAE averaged over four forecast horizons (96, 192, 336, and 720). Lower is better, and the best value for each metric is in bold.

VisionTS++ targets general-purpose time-series forecasting. It continues pretraining on LOTSA, which contains over 231 billion observations, and updates all backbone parameters. MUSE focuses on adapting a frozen visual backbone and uses the target training set to update only the newly introduced adaptation parameters. Given these differences in research objectives and training protocols, we omitted VisionTS++ from the main comparison in the original paper. The table below compares the two methods using MAE-Base in both cases.

Comparison of MUSE and VisionTS++ on six shared datasets.

DatasetMUSEVisionTS++
ETTh10.384 / 0.4050.402 / 0.416
ETTh20.325 / 0.3720.333 / 0.370
ETTm10.330 / 0.3620.360 / 0.372
ETTm20.245 / 0.3090.244 / 0.298
Electricity0.157 / 0.2510.184 / 0.265
Weather0.216 / 0.2610.222 / 0.241

As shown in the table, MUSE achieves lower average MSE on five of the six datasets and lower values for both metrics on ETTh1, ETTm1, and Electricity. The next table presents the comparison with TimeXer and Time-VLM on the same six datasets.

Comparison of MUSE, TimeXer, and Time-VLM on six shared datasets.

DatasetMUSETimeXerTime-VLM
ETTh10.384 / 0.4050.437 / 0.4370.405 / 0.420
ETTh20.325 / 0.3720.367 / 0.3960.341 / 0.391
ETTm10.330 / 0.3620.382 / 0.3970.350 / 0.377
ETTm20.245 / 0.3090.274 / 0.3220.248 / 0.311
Electricity0.157 / 0.2510.171 / 0.2700.172 / 0.273
Weather0.216 / 0.2610.241 / 0.2710.224 / 0.263

As shown in the table, MUSE achieves lower average MSE and MAE than both TimeXer and Time-VLM on all six datasets.

We use the full-shot setting to evaluate how well pretrained visual representations transfer to time-series forecasting when the entire target training set is available for adaptation. The MAE encoder and decoder remain frozen throughout training, and we update only the newly introduced adaptation parameters. The frozen MAE contains 111.908M parameters, and MUSE adds 1.147M trainable parameters, equivalent to 1.03% of the MAE parameter count. Approximately 98.99% of the full model’s parameters are frozen. In the paper, “lightweight” refers specifically to the trainable parameter overhead of adaptation. We will include these baseline comparisons and parameter counts in the final version and clarify the role of the frozen visual prior, the full-shot training setting, and what we mean by “lightweight.”

AuthorsOpenReview ↗

Response to Reviewer J9p9 (Part III)

W2 & Q3: It remains unclear whether VCR can capture lagged or time-varying cross-variable dependencies. Its contribution to cross-variable modeling needs direct validation through channel shuffling, removal of inter-variable attention, or analysis of attention weights.

We thank the reviewer for pointing out a limitation of VCR and suggesting that we directly test the role of cross-variable attention. We apologize for the lack of clarity about this limitation in the original paper. As shown in Eq. (5)‾\underline{\text{Eq. (5)}}, VCR generates one correction vector per variable and adds it to all patches of that variable within the same historical window. Consequently, VCR cannot capture lagged or time-varying cross-variable dependencies.

On Electricity, which has 321 variables, MUSE and DMMV-A perform comparably. Table 1 of the original paper (line 388)‾\underline{\text{Table 1 of the original paper (line 388)}} reports average MSE / MAE of 0.157 / 0.251 for MUSE and 0.158 / 0.248 for DMMV-A. We interpret these results using two dataset characteristics (stationarity and transition) from TFB [1].

Definition 1 (Stationarity). In a weakly stationary series, the mean and variance are constant, the variance is finite, and the covariance between any two observations depends only on their time lag. For any shift (rr) that keeps both indices within the series,

cov⁡(xi,xj)=cov⁡(xi+r,xj+r).\operatorname{cov}(x_i,x_j)=\operatorname{cov}(x_{i+r},x_{j+r}).

TFB assesses stationarity using the augmented Dickey–Fuller (ADF) test and defines the dataset-level Stationarity value as the mean of the ADF (p)-values across variables.

Definition 2 (Transition). Transition is the trace of the covariance matrix associated with the transition matrix for a three-symbol alphabet. It characterizes identifiable temporal patterns such as trends and periodicity. Lower values correspond to clearer temporal regularities, while higher values correspond to less pronounced patterns.

Electricity has a Stationarity value of 0.00515, indicating strong evidence of stationarity, and a Transition value of 0.01049, indicating that trends and periodic patterns are readily identifiable. These clear within-variable temporal and periodic patterns are well suited to TPR’s design. By modeling within-variable dependencies along the historical time and phase axes, respectively, TPR plays a more important role in MUSE’s forecasts on Electricity.

To directly assess the contribution of cross-variable attention to forecasting, we conducted an intra-variable-only VCR ablation. Table 2 (lines 397–407 of the original paper)‾\underline{\text{Table 2 (lines 397–407 of the original paper)}} reports ablation results on ETTh2, Weather, AQWan, and ZafNoo at a forecast horizon of H=720H=720. We evaluated this control on the same datasets and forecast horizon. In the full model, VCR uses cross-variable interaction to generate a correction vector for each variable. In this ablation, we remove the inter-variable multi-head self-attention in Eq. (3)‾\underline{\text{Eq. (3)}}.

Comparison of the full model and the intra-variable-only VCR control. Each cell reports MSE / MAE. Lower is better, and the best result for each metric is in bold.

DatasetFull ModelVCR intra-variable only
ETTh20.373 / 0.4160.376 / 0.419
Weather0.299 / 0.3260.303 / 0.328
AQWan0.883 / 0.5260.888 / 0.530
ZafNoo0.564 / 0.4680.570 / 0.471

As shown in the table, the full model achieves lower MSE and MAE than the intra-variable-only control on all four datasets. This indicates that the historical representations of other variables can provide additional predictive information for each variable. VCR uses cross-variable attention to incorporate this information into the variable-level correction vectors, improving forecasting performance.

We further visualize VCR’s cross-variable attention on ZafNoo in Figure 5 (Appendix C.3, page 18, lines 918–942 of the revised paper)‾\underline{\text{Figure 5 (Appendix C.3, page 18, lines 918–942 of the revised paper)}}. In Figure 5, rows represent query variables, columns represent the variables they attend to, and colors indicate attention weights.

At both forecast horizons, most off-diagonal entries in the swc_deep column have relatively high weights, indicating that multiple variables draw on the historical representation of swc_deep when generating their correction vectors. The swc_shallow column is also more prominent at H=720H=720 than at H=336H=336, illustrating how cross-variable attention weights differ between the two horizons. We will include this ablation and visualization in the final version of the paper and clarify the scope of the cross-variable dependencies that VCR models.

[1] TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods. Proceedings of the VLDB Endowment, 17(9), pp. 2363–2377, 2024.

Reviewer_M5acOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework for multivariate time-series forecasting based on a fully frozen pretrained vision backbone. It addresses two challenges: modeling cross-variable dependencies while preserving independent visual representations, and adapting visual features to the specific structures of time-series data. It introduces a Variable Context Refinement module for intra- and inter-variable dependency modeling and a Temporal–Periodic Refinement module for structure-aware adaptation. Experiments demonstrate competitive performance against representative baselines, with ablation studies supporting the effectiveness of the proposed components.

Critical strengths
  1. This paper explores how to effectively adapt a fully frozen pretrained vision backbone to multivariate time-series forecasting, an important problem with broad real-world applications. The motivation is clear, particularly in addressing the challenges of transferring visual representations to time-series data while preserving the pretrained backbone.

  2. The paper conducts extensive experiments on 10 public datasets with representative baselines, supported by ablation and sensitivity studies. The results demonstrate competitive performance and validate the proposed components. The paper is well-organized and clearly written, with clear figures and tables that effectively illustrate the motivation, methodology, and experimental findings

Critical weaknesses
  1. The choice of periodicity is insufficiently explained. The model first organizes time series into images according to the period length F, then processes information along the across-period and within-period axes. However, the paper does not specify the F values used for each dataset or explain how they are selected.

  2. Table 1 reports competitive forecasting results, but the paper does not provide results from repeated runs with different random seeds or report standard deviations.

  3. The paper adopts a fully frozen pretrained vision backbone and introduces lightweight modules for time-series adaptation. However, the motivation for keeping the entire backbone frozen, rather than allowing partial fine-tuning, is not sufficiently discussed.

Questions for authors
  1. How are the periodicity values F determined for different datasets?

  2. Could the authors provide results from repeated runs with different random seeds to demonstrate the stability and reliability of the improvements reported in Table 1?

  3. What is the motivation for keeping the entire pretrained vision backbone frozen rather than allowing partial fine-tuning? Could the authors further clarify the considerations behind this design choice?

4 comments
AuthorsOpenReview ↗

Response to Reviewer M5ac (Part I)

We thank Reviewer M5ac for the thoughtful review and constructive feedback. We appreciate the opportunity to further explain details on MUSE.

W1 & Q1: The paper does not report the period length F used for each dataset or explain how these values were selected.

We thank the reviewer for pointing out this omission. In Algorithm 1 of Appendix A (lines 712–713 and 729–734 of the original paper‾\underline{\text{lines 712–713 and 729–734 of the original paper}}), FF determines the number of samples in each column of the two-dimensional grid formed by folding the time series. MUSE uses a 24-hour folding period across all datasets in the main experiments. Specifically, MUSE infers the sampling interval from the training timestamps and sets FF to the number of samples in one day. For a sampling interval of Δt\Delta t minutes, F=1440/ΔtF = 1440/\Delta t. Each dataset therefore uses the same FF across all forecast horizons.

Table 3 in Appendix B.1 (lines 757–767 of the original paper‾\underline{\text{lines 757–767 of the original paper}}) reports the sampling intervals for all datasets. ETTh1, ETTh2, Electricity, and AQWan are sampled hourly. ETTm1, ETTm2, and Wind are sampled every 15 minutes, Weather every 10 minutes, and ZafNoo and CzeLan every 30 minutes. The table below lists the corresponding FF values. In the final version, we will explain how FF is determined and report the value used for each dataset.

DatasetSampling intervalPeriod length FF
ETTh11 hour24
ETTh21 hour24
ETTm115 minutes96
ETTm215 minutes96
Electricity1 hour24
Wind15 minutes96
Weather10 minutes144
AQWan1 hour24
ZafNoo30 minutes48
CzeLan30 minutes48

W2 & Q2: The paper does not report results from repeated runs with different random seeds or the corresponding standard deviations, leaving the stability of the improvements unclear.

We thank the reviewer for suggesting this evaluation. We ran MUSE with two additional random seeds 2022 and 2023 on six datasets: ETTh1, ETTh2, ETTm1, ETTm2, ZafNoo, and CzeLan. The seed 2021 results are from Tables 1 and 4 of the original paper (lines 379–394 and 811–862, respectively)‾\underline{\text{Tables 1 and 4 of the original paper (lines 379–394 and 811–862, respectively)}}.

We report MSE and MAE as mean ± sample standard deviation (SD) across the three seeds. Each Avg row summarizes the three seed-level averages over the four forecast horizons. Across the six datasets, the largest sample SDs in the Avg rows are 0.003606 for MSE and 0.002517 for MAE.

DatasetHorizonMSE (mean ± sample SD)MAE (mean ± sample SD)
ETTh1960.339 ± 0.0023090.371 ± 0.000577
ETTh11920.388 ± 0.0000000.403 ± 0.000000
ETTh13360.405 ± 0.0005770.417 ± 0.000577
ETTh17200.407 ± 0.0005770.432 ± 0.000577
ETTh1Avg0.385 ± 0.0011550.406 ± 0.000577
ETTh2960.265 ± 0.0010000.324 ± 0.000577
ETTh21920.320 ± 0.0005770.364 ± 0.001155
ETTh23360.344 ± 0.0032150.390 ± 0.005033
ETTh27200.372 ± 0.0011550.416 ± 0.000577
ETTh2Avg0.325 ± 0.0005770.373 ± 0.001000
ETTm1960.287 ± 0.0020000.330 ± 0.000577
ETTm11920.321 ± 0.0036060.354 ± 0.001732
ETTm13360.337 ± 0.0017320.369 ± 0.000577
ETTm17200.379 ± 0.0005770.398 ± 0.001000
ETTm1Avg0.331 ± 0.0011550.363 ± 0.001000
ETTm2960.165 ± 0.0025170.253 ± 0.001155
ETTm21920.222 ± 0.0020000.292 ± 0.001155
ETTm23360.270 ± 0.0005770.323 ± 0.002000
ETTm27200.335 ± 0.0092920.377 ± 0.006028
ETTm2Avg0.248 ± 0.0036060.311 ± 0.002000
ZafNoo960.423 ± 0.0010000.390 ± 0.002000
ZafNoo1920.480 ± 0.0045830.425 ± 0.004163
ZafNoo3360.514 ± 0.0032150.447 ± 0.001732
ZafNoo7200.563 ± 0.0011550.468 ± 0.000000
ZafNooAvg0.495 ± 0.0010000.433 ± 0.000577
CzeLan960.174 ± 0.0055080.230 ± 0.005568
CzeLan1920.204 ± 0.0032150.257 ± 0.004041
CzeLan3360.232 ± 0.0020820.282 ± 0.001000
CzeLan7200.242 ± 0.0037860.298 ± 0.003000
CzeLanAvg0.213 ± 0.0020000.267 ± 0.002517
AuthorsOpenReview ↗

Response to Reviewer M5ac (Part II)

W3 & Q3: The paper provides limited justification for fully freezing the pretrained vision backbone.

We thank the reviewer for raising this question. The central motivation behind MUSE is to test whether an MAE pretrained on natural images can serve as the sole forecasting backbone and achieve competitive multivariate time series forecasting performance through adaptation, with all its pretrained weights kept frozen. Throughout training, we freeze every parameter in the MAE encoder and decoder, including the LayerNorm parameters, and train only the newly introduced time series adaptation modules and the fusion gate. This setup allows us to test whether the reconstruction capability learned from natural images transfers to time series forecasting with the original MAE weights fixed.

During masked reconstruction pretraining on natural images, MAE learns to infer the contents of masked regions from local shapes, spatial layouts, and recurring patterns in visible patches. When a time series is folded by period into a two-dimensional grid, the horizontal axis orders observations at the same phase over time, while the vertical axis orders observations by phase. Both changes over time at a fixed phase and relationships between phase positions can thus appear as spatial structure in the grid. The observed history forms the visible region, and the future to be forecast forms the masked region. MAE can therefore apply its pretrained reconstruction capability to forecasting by using structural cues in the visible history to reconstruct the masked future.

As shown in Table 1 of the original paper (lines 379–394)‾\underline{\text{Table 1 of the original paper (lines 379–394)}}, MUSE achieves lower average MSE and MAE than VisionTS on all 10 datasets with the MAE fully frozen. These results show that, with adaptation tailored to time series structure, a fully frozen MAE pretrained on natural images can serve as a competitive backbone for multivariate time series forecasting.

I appreciate the authors’ response and additional experiments, which have satisfactorily addressed my concerns. I will raise my score.

Thank you for your positive feedback and recognition of our efforts. We sincerely appreciate your support.

Reviewer_9aMzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MUSE, a multivariate time-series forecasting framework built on a frozen pretrained MAE. It introduces two trainable branches: the Variable Context Refinement Module (VCR), which aggregates variable-level context and models cross-variable dependencies, and the Temporal–Periodic Refinement Module (TPR), which applies bottleneck adapters with axis-specific attention throughout the frozen encoder. The two forecasts are combined through a learnable prediction-level gate.

The evaluation covers ten datasets and four forecasting horizons, with favorable dataset-average results against the selected baselines. The approach is clearly motivated and potentially useful. However, the current experiments do not sufficiently distinguish the proposed dependency-aware mechanisms from generic adaptation and ensemble effects, and the computational cost and comparison protocol require further clarification.

Critical strengths
  1. The paper identifies two relevant limitations of visual forecasting: the absence of explicit cross-variable interaction in variable-wise processing and the different semantics of the two axes in period-folded images.
  2. VCR exchanges information through compact variable summaries, while TPR shares adapters within encoder stages. These are reasonable choices for adapting a frozen backbone without updating all pretrained parameters.
  3. Results cover ten datasets across several application domains and four horizons. MUSE achieves strong dataset-average performance, with particularly encouraging MSE results on Wind and CzeLan.
Critical weaknesses
  1. VCR combines learnable summary tokens, intra-variable attention, inter-variable attention, and residual write-back. TPR combines bottleneck adaptation with axis-wise attention and relative-position biases. These components are reasonably motivated, but the experiments do not establish how much benefit comes specifically from the proposed dependency-aware organization rather than adding trainable capacity to a frozen backbone.
  2. VCR and TPR independently produce forecasts and receive separate supervision, while the fusion gate is trained on detached predictions. Consequently, the full system also functions as a learned ensemble of two predictors. Removing either branch reduces both model capacity and ensemble diversity. The current branch-removal ablations therefore do not establish that the gains require the proposed complementary dependency structures. Useful controls include equal-weight averaging, two independently initialized copies of the same branch, and a single branch combining both refinement mechanisms. Reporting learned gate values and correlations between branch errors would also help explain when the branches are genuinely complementary.
  3. The VCR ablation removes the entire branch rather than isolating cross-variable interaction. It therefore cannot determine whether improvements come from inter-variable dependencies, intra-variable aggregation, or simply learned input refinement. An intra-variable-only VCR control is particularly important because the variable-level correction is broadcast identically to every historical patch. Similarly, replacing temporal–periodic attention with mean pooling changes both adaptivity and trainable capacity. Stronger controls should separately remove temporal attention, periodic attention, relative-position biases, and circular indexing. There is also a semantic qualification that needs attention: after resizing and patch embedding, a historical patch column may summarize multiple original periods rather than one period. TPR operates on a compressed temporal–phase grid, so the interpretation of its attention as modeling individual periods should be justified more carefully.
Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9aMz (Part I)

We would like to sincerely thank Reviewer 9aMz for the thorough review and specific suggestions. In response, we have conducted additional ablation experiments and analyses.

W1.1 & W3.1: The lack of an intra-variable-only VCR control makes it difficult to isolate the contribution of cross-variable interaction.

We thank the reviewer for this suggestion. To isolate the contribution of cross-variable interaction, we added an intra-variable-only VCR ablation. Table 2 in lines 397–407 of the original paper‾\underline{\text{Table 2 in lines 397–407 of the original paper}} reports ablation results on ETTh2, Weather, AQWan, and ZafNoo at a forecast horizon of H=720H=720. We evaluated this control on the same datasets and forecast horizon.

In the full model, VCR uses cross-variable interaction to generate a correction vector for each variable. In the intra-variable-only ablation, we remove the inter-variable multi-head self-attention in Eq. (3) of the original paper‾\underline{\text{Eq. (3) of the original paper}}.

Comparison of the full model and the intra-variable-only VCR control. Each cell reports MSE / MAE. Lower is better, and the best result for each metric is in bold.

DatasetFull ModelVCR intra-variable only
ETTh20.373 / 0.4160.376 / 0.419
Weather0.299 / 0.3260.303 / 0.328
AQWan0.883 / 0.5260.888 / 0.530
ZafNoo0.564 / 0.4680.570 / 0.471

As shown in the table, the full model achieves lower MSE and MAE than the intra-variable-only control on all four datasets. This indicates that the histories of other variables provide predictive information beyond what each variable’s own history offers. Cross-variable interaction allows VCR to incorporate this information into the variable-level correction vectors, improving MUSE’s forecasting performance.

W1.2 & W3.2: The study lacks separate ablations for temporal attention, periodic attention, relative-position biases, and circular indexing.

We thank the reviewer for suggesting separate ablations of the TPR components. Table 2 in lines 397–407 of the original paper‾\underline{\text{Table 2 in lines 397–407 of the original paper}} reports ablation results on ETTh2, Weather, AQWan, and ZafNoo at a forecast horizon of H=720H=720. We evaluated four additional ablations on the same datasets and forecast horizon.

To isolate the contribution of each component, we separately remove temporal attention, periodic attention, relative-position bias, and circular indexing in periodic attention.

Comparison of the full model and the individual TPR component ablations. Each cell reports MSE / MAE. Lower is better, and the best result for each metric is in bold.

DatasetFull Modelw/o Temporal Attentionw/o Periodic Attentionw/o Relative-Position Biasw/o Circular Indexing
ETTh20.373 / 0.4160.376 / 0.4200.375 / 0.4180.374 / 0.4160.373 / 0.416
Weather0.299 / 0.3260.299 / 0.3260.299 / 0.3260.299 / 0.3260.299 / 0.326
AQWan0.883 / 0.5260.890 / 0.5310.889 / 0.5310.885 / 0.5270.885 / 0.528
ZafNoo0.564 / 0.4680.573 / 0.4720.574 / 0.4730.565 / 0.4680.567 / 0.471

As shown in the table, we have the following observations: 1) Removing either temporal attention or periodic attention increases both MSE and MAE on ETTh2, AQWan, and ZafNoo. These increases are larger than those observed when removing relative-position bias or circular indexing, indicating that each attention mechanism contributes to forecasting performance. 2) Relative-position bias and circular indexing also provide additional gains on some datasets. 3) On Weather, all four component ablations match the full model in both MSE and MAE at the reported precision. As shown in the w/o TPR column of Table 2 in lines 397–407 of the original paper‾\underline{\text{Table 2 in lines 397–407 of the original paper}}, removing the entire TPR branch increases MSE by only 0.001, from 0.299 to 0.300. This suggests that forecasting performance on Weather is relatively insensitive to these TPR components in this setting.

AuthorsOpenReview ↗

Response to Reviewer 9aMz (Part II)

W2: Removing either branch of MUSE reduces both model capacity and ensemble diversity, obscuring the contribution of VCR–TPR complementarity.

W2.1: Lack of four ablations that could isolate the benefit of placing VCR and TPR in separate branches.

Thank you for suggesting these controls to clarify the benefit of placing VCR and TPR in separate branches. Table 2 in lines 397–407 of the original paper‾\underline{\text{Table 2 in lines 397–407 of the original paper}} reports ablation results on ETTh2, Weather, AQWan, and ZafNoo at a forecast horizon of H=720H=720. We evaluated the suggested controls using the same datasets and forecast horizon.

We compare the full model with four variants. 1) Equal-weight fusion: retains the VCR and TPR branches and averages their forecasts with a fixed weight of 0.5 for each branch. 2) VCR × 2: uses two independently initialized VCR branches and fuses their forecasts. 3) TPR × 2: uses two independently initialized TPR branches and fuses their forecasts. 4) Single branch (VCR + TPR): applies both refinement mechanisms in the same branch and produces a single forecast.

Each cell reports MSE / MAE. Lower is better, and the best results are in bold.

DatasetFull ModelEqual-weight FusionVCR × 2TPR × 2Single Branch (VCR + TPR)
ETTh20.373 / 0.4160.373 / 0.4160.388 / 0.4310.378 / 0.4190.387 / 0.429
Weather0.299 / 0.3260.300 / 0.3260.302 / 0.3310.309 / 0.3300.301 / 0.331
AQWan0.883 / 0.5260.884 / 0.5270.899 / 0.5300.899 / 0.5350.892 / 0.531
ZafNoo0.564 / 0.4680.565 / 0.4680.584 / 0.4810.568 / 0.4730.571 / 0.475

As shown in the table, we have the following observations: 1) the full model slightly improves on equal-weight fusion in four of the eight comparisons and ties in the other four. This indicates that learning the fusion weights provides a modest gain over equal-weight averaging. 2) VCR × 2, TPR × 2, and the single-branch perform substantially worse than the full model across the four datasets. Both VCR × 2 and TPR × 2 retain a learnable two-branch ensemble, yet the full model outperforms both variants in every comparison. This shows that the distinct roles of VCR and TPR, which model cross-variable context and temporal and periodic structures, respectively, provide forecasting gains beyond those achieved by ensembling two branches of the same type. 3) The single-branch control further shows that when both refinement mechanisms are retained, generating forecasts in separate branches and then fusing them yields lower forecast errors.

W2.2: The learned gate weights and correlations between branch forecast errors are unreported, leaving it unclear when the two branches complement each other.

We thank the reviewer for suggesting a further analysis to improve the quality of our paper. In MUSE, the fusion gate controls the contribution of each branch to the final forecast. We denote the scalar gate learned for each dataset and forecast horizon by gg. VCR receives weight gg, and TPR receives weight 1−g1 - g. The table below reports the learned gate values at four forecast horizons for the 10 datasets described in lines 360–366 of the original paper‾\underline{\text{in lines 360–366 of the original paper}}.

Learned gate values gg across datasets and forecast horizons HH.

DatasetH=96H = 96H=192H = 192H=336H = 336H=720H = 720
ETTh10.510.500.510.51
ETTh20.520.500.510.51
ETTm10.530.540.520.52
ETTm20.460.530.510.52
Electricity0.440.400.350.38
Wind0.540.540.540.52
Weather0.550.540.540.51
AQWan0.520.500.510.55
ZafNoo0.480.530.500.54
CzeLan0.500.550.500.53

As shown in the table, gg differs from 0.50 in 34 of the 40 settings and equals 0.50 in the remaining six. The gate thus favors one branch over the other in most settings. Above results indicate that learning fusion weights for each dataset and forecast horizon provides a modest performance gain over fixed equal weighting.

AuthorsOpenReview ↗

Response to Reviewer 9aMz (Part III)

Furthermore, to address your question about when the two branches complement each other, we analyzed Electricity and CzeLan and added the results to Appendix C.3‾\underline{\text{Appendix C.3}} of the revised paper. Figure 4 (lines 892–911 of the revised paper)‾\underline{\text{Figure 4 (lines 892–911 of the revised paper)}} compares the validation MSE of VCR and TPR at all four forecast horizons on both datasets. The branch losses are defined in line 351 of the original paper‾\underline{\text{in line 351 of the original paper}}. TPR has lower validation MSE at all four horizons on Electricity, whereas VCR has lower validation MSE at all four horizons on CzeLan. This reversal is consistent with the learned gate weights reported in Part II. We interpret this difference using two dataset characteristics (stationarity and transition) from TFB [1].

Among the 10 datasets used in MUSE, CzeLan has the highest Stationarity value 0.15963 (values ≤ 0.05 indicate stationarity). Electricity has a lower value 0.00515, indicating stronger evidence of stationarity. CzeLan also has a higher Transition value 0.05317 than Electricity 0.01049 (Lower values correspond to clearer temporal regularities, while higher values correspond to less pronounced patterns), indicating that temporal patterns such as trends and periodicity are easier to identify in Electricity. These characteristics are consistent with the branch performance in Figure 4. The clearer within-variable patterns in Electricity favor TPR’s modeling of temporal dependencies across historical periods and periodic dependencies within each period. On CzeLan, within-variable temporal and periodic patterns are harder to capture consistently, and VCR provides more useful forecasting information by modeling cross-variable contextual dependencies. Therefore, the two branches complement each other across datasets by capturing different structures that are useful for forecasting. We will include this analysis in the final manuscript to clarify how VCR and TPR complement each other across datasets with different temporal characteristics.

Definition 1 (Stationarity). In a weakly stationary series, the mean is constant, the variance is finite, and the covariance between two observations depends only on the time lag between them. For any shift rr that keeps both indices within the series,

cov⁡(xi,xj)=cov⁡(xi+r,xj+r).\operatorname{cov}(x_i,x_j)=\operatorname{cov}(x_{i+r},x_{j+r}).

TFB assesses stationarity using the augmented Dickey–Fuller (ADF) test. The dataset-level Stationarity value is the mean of the ADF (p)-values across variables.

Definition 2 (Transition). Transition is the trace of the covariance of transition matrix between symbols in a 3-letter alphabet. It characterizes recognizable patterns such as trends and periodicity. Lower values correspond to clearer temporal regularities, while higher values correspond to less pronounced patterns.

[1] TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods. Proceedings of the VLDB Endowment, 17(9), pp. 2363–2377, 2024.

W3.3: A compressed historical patch column may summarize multiple original periods, leaving the single-period interpretation of TPR attention unclear.

We thank the reviewer for pointing out that our original wording could imply that each historical patch column corresponds to a single original period. We have clarified that TPR models temporal dependencies across historical patch columns at a fixed phase position and periodic dependencies among phase positions within each column. These changes appear in the revised paper’s Abstract (lines 24–25)‾\underline{\text{in the revised paper’s Abstract (lines 24–25)}}, Introduction (lines 104–105)‾\underline{\text{Introduction (lines 104–105)}}, Section 3.4 (lines 263–264, 295–298, 310–313, and 322–323)‾\underline{\text{Section 3.4 (lines 263–264, 295–298, 310–313, and 322–323)}}, and Conclusions (line 483)‾\underline{\text{Conclusions (line 483)}}.

Notably, the core motivation of our work still holds after this clarification. As shown in Algorithm 1 of Appendix A (lines 733–734 and 742 of the original paper‾\underline{\text{lines 733–734 and 742 of the original paper}}), period folding yields a grid in which each column corresponds to one original period and each row aligns the same phase across periods. After bilinear resizing and patch embedding, a historical patch column may aggregate information from multiple adjacent original periods. On the compressed grid, the horizontal axis retains chronological order and the vertical axis retains phase order, both at a coarser resolution. These distinct time and phase semantics differ from the spatial semantics of the axes in natural images and motivate TPR to model dependencies along the two axes separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.