acceptodds
Under review as a conference paper at ICLR 2027

Risk-Controlled Adaptive Native-Scale Fusion: A Backbone-Agnostic Forecasting Framework

Abstract

Multi-scale modeling is widely used in time series forecasting to capture complementary temporal structures, from fine-grained variations to slower, long-range dynamics. However, temporal aggregation changes the information available at each scale: coarse views suppress high-frequency variation while losing local detail. Requiring coarse branches to reconstruct the finest-grid future, or allowing coarse predictions to modify fine-scale forecasts without sufficient control, can therefore introduce mismatched supervision and harmful cross-scale corrections. To address these, we propose a backbone-agnostic multi-scale forecasting framework that separates native-scale forecasting from prediction-space fusion. Each branch first predicts the future on its native temporal grid, and cross-scale interaction is deferred until these forecasts are completed. We instantiate the framework with two complementary fusion mechanisms: Safe-linear, which combines deterministic alignment with learned participation gates, and Risk-Controlled Adaptive Native-Scale Fusion (RC-ANSF), which introduces bounded adaptive alignment for greater flexibility. A two-stage training procedure further separates scale-specific forecasting from cross-scale fusion. Across three backbone families, eight datasets, four forecasting horizons, the proposed framework consistently improves aggregate forecasting performance. These results support native-scale prediction followed by controlled prediction-space fusion as a reusable strategy for multi-scale forecasting. Our framework is available at https://anonymous.4open.science/r/RC-ANSF-Time-Seires-Forecasting-Framework.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Multi-Scale Forecasting, Adaptive Fusion

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_iAFNOpenReview ↗
Rating
3(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes a backbone agnostic multi‑scale time series forecasting framework separating native scale prediction and prediction‑space fusion. It develops Safe linear and RC‑ANSF fusion modules and adopts a two stage training pipeline. Evaluations across three backbones, eight datasets and four horizons verify consistent performance gains from this reusable multi‑scale design.

Critical strengths
  1. It decouples native‑scale forecasting and fusion, compatible with diverse existing backbones, showing good generalizability for multi‑scale forecasting research.
  2. It provides bounded risk controlled fusion mechanisms and structural theoretical bounds, constraining harmful cross scale corrections.
Critical weaknesses
  1. Risk controlled bounds only guarantee parameter space constraints rather than test‑set error improvement; performance gains are task dependent without universal superiority.
  2. It lacks real world case studies beyond standard public multivariate forecasting benchmarks.
Questions for authors
  1. Section 3.4 states that "risk control refers to these structural restrictions together with the empirical excess-loss penalty below, rather than a guarantee of non-increasing test error". Is this risk control only a constraint in parameter space, and does it not provide a guarantee of test error reduction in a statistical sense?
  2. In the experiments, the Scaleformer comparison only reuses the official coarse-to-fine procedure without targeted tuning for the three types of backbones. Is it possible that the actual performance of this comparison method is underestimated?
  3. Existing multi-scale models such as N-HiTS and TimeMixer have already realized cross-scale fusion. What is the specific core innovation difference of the design in this paper that decouples native prediction and fusion compared with similar methods?
  4. The discussion section mentions that "the evidence supports task-dependent control of cross-scale corrections, not unconditional improvements from tighter constraints". What specific data characteristics or scenarios will tighter constraints lead to performance degradation?
  5. Proposition 2 in Appendix A shows that the lower bound of the base model weight is 0 when \alpha=1. Does the fusion mechanism have the potential risk of completely discarding the fine-scale base prediction at this time?
  6. In the ablation study, "No gate" removes both the participation gate and the candidate-quality supervision at the same time, which cannot verify the role of the gating network alone. Why is a control experiment that only removes the supervision term not designed?
  7. Safe-linear has already realized deterministic alignment and gated fusion. On this basis, the bounded adaptive residual added by RC-ANSF brings limited gains. Where does the necessity of this module lie?
  8. The trust radius sensitivity experiment shows completely opposite change trends on the ETTh2 and Traffic datasets. What is the reason for the sequence characteristics behind this difference?
10 comments
AuthorsOpenReview ↗

Part1 for W1 and Q1

Thank you for your questions.

The risk control acts directly in prediction space. It bounds the alignment correction and the total coarse-scale contribution to the fused forecast, independently of projector parameter norms. Describing it as only a parameter-space constraint therefore overlooks the prediction-space guarantees. As explicitly stated in Section 3.4, these structural guarantees do not constitute a statistical guarantee of reduced test error. Their empirical effectiveness is supported by the broad improvements below.

Applicability across datasets and backbones. The same fusion framework works with linear, attention-based, and MLP-based predictors across all eight benchmarks. Relative to each single-scale backbone (Vanilla), its four-horizon average MSE is lower on 8/8 datasets for DLinear, 8/8 for iTransformer, and 6/8 for TimeMixer-Core at the precision reported in the main table. Averaged across backbones, RC-ANSF improves MSE on all eight datasets. Across the 96 evaluated settings, the macro MSE reductions for the three backbones are 14.6%, 5.6%, and 1.4%, respectively. These results demonstrate consistent aggregate gains across diverse forecasting domains and all three predictor families, providing direct empirical support for the broad applicability of RC-ANSF.

AuthorsOpenReview ↗

Part2 for W2

The all eight main benchmarks we used in our paper contain real-world measurements, including transformer temperatures and loads, weather, electricity usage, road occupancy, and solar output. We extend this offline evaluation beyond multivariate benchmarks to all 414 M4 Hourly univariate series and 30,490 M5 item–store sales series with covariates. M5 forecasts 28 days of sales using iTransformer with historical sales, prices, calendar/event features, and SNAP indicators.

Task and backboneMetricRC-ANSF (Ours)Safe-linearBase
M4 Hourly / DLinearMASE1.2081.2431.243
M4 Hourly / TimeMixer-CoreMASE1.3591.3851.529
M5 / iTransformerWRMSSE0.9230.9260.928

The eight main benchmarks cover diverse real-world forecasting domains, supporting the broad applicability of our framework. The supplementary M4 results also demonstrate strong performance on univariate forecasting. Gains on M5 are limited, suggesting less additional benefit in settings with substantial covariate-modeling demands. This is consistent with our design: the fusion framework introduces no dedicated module for modeling relationships between variables, facilitating adaptation to different backbone predictors.

AuthorsOpenReview ↗

Part3 for Q2

Regarding your concern that insufficient tuning may underestimate Scaleformer's performance, we clarify that the Scaleformer results in our paper already incorporate task-specific learning-rate tuning. For each backbone, dataset, and forecast horizon, we trained the adaptation at 11, 0.30.3, and 0.10.1 times the inherited learning rate, with all scale-specific backbones trainable. We selected the learning rate with the lowest validation MSE within this grid for each of the 96 settings and used that setting's selected rate in the subsequent repeated experiments reported in the paper.

The table illustrates why learning-rate search matters for a fair Scaleformer comparison. Entries are test MSE averaged over eight datasets and four horizons for each backbone; Val.-selected uses the best learning rate selected separately for each setting by validation MSE.

Backbone1×1\times0.3×0.3\times0.1×0.1\timesVal.-selected
DLinear0.6400.4540.3410.339
iTransformer0.3900.3910.4090.354
TimeMixer-Core0.3360.3190.3180.315
Overall0.4560.3880.3560.336

Without learning-rate search, Scaleformer can perform substantially worse: DLinear's average MSE varies from 0.341 to 0.640 across the tested fixed learning-rate multipliers. Selecting by validation MSE lowers overall test MSE from 0.456 at the inherited rate to 0.336, a 26.21% reduction. Our paper uses the best validation-selected learning rate within the search grid for every setting in the subsequent repeated experiments. Thus, the reported Scaleformer performance is not underestimated through omission of learning-rate tuning.

Reviewer_Wj78OpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper proposes RC-ANSF, a multi-scale time series forecasting framework that separates native-scale forecasting from prediction-space fusion. Each scale-specific branch first generates forecasts at its corresponding temporal resolution. The predictors are then frozen, and the final forecast is produced through alignment and fusion modules. The authors further introduce bounded residual alignment and gating mechanisms to control the influence of different scales on the final prediction.

Critical strengths

1,The motivation is clear, and separating scale-specific forecasting from cross-scale fusion is a reasonable design choice.

2,The overall framework is simple, easy to understand and implement, and can accommodate different forecasting backbones.

Critical weaknesses

W1. Limited methodological novelty and insufficient validation of the independent contribution of the core design. The proposed method mainly combines existing techniques, including multi-scale sampling, independent predictors, interpolation, gated fusion, and residual constraints. Overall, it appears to be an incremental combination of existing approaches. Although the authors emphasize the separation of native-scale supervision and prediction-space fusion, the experiments do not sufficiently demonstrate that this design provides distinct advantages over conventional multi-scale forecasting and ensemble methods.

W2. Limited backbone selection and insufficient comparisons with recent competitive forecasting models. The paper aims to introduce a backbone-agnostic forecasting enhancement framework. However, the experiments only consider DLinear, iTransformer, and a simplified TimeMixer-Core, without evaluating the framework on more competitive forecasting models from 2025–2026. This limits the evidence supporting its general applicability. The paper mainly demonstrates improvements over the original backbones but lacks direct comparisons with recent methods such as GTR and SOFTS. Moreover, the additional computational overhead is not adequately analyzed.

W3. Insufficient evidence for the necessity and effectiveness of the risk-control mechanisms. The ablation results show that removing the residual magnitude constraint can actually improve forecasting accuracy in a considerable number of experimental settings, while removing the projection router has little impact on overall performance. These observations raise questions about the necessity of the individual components in the complete risk-control design. Furthermore, some ablation studies modify multiple factors simultaneously, and the auxiliary losses are not independently evaluated, making it difficult to identify the actual sources of performance improvement.

Questions for authors

Q1. What is the key technical novelty of the proposed method compared with existing multi-scale forecasting and ensemble approaches? Does native-scale supervision combined with two-stage fusion still provide independent benefits when model capacity and training budgets are matched?

Q2. Could the authors evaluate the proposed framework on more recent and competitive forecasting backbones, and compare its performance against GTR, SOFTS, the full TimeMixer, and TimeMixer++ under consistent experimental settings? Could the authors also report the additional computational costs?

Q3. Could the authors provide more rigorous ablation studies to demonstrate the necessity of the proposed risk-control components?

5 comments
AuthorsOpenReview ↗

Part1 for W1 and Q1

Our methodological contribution is the decoupling of native-scale future learning and risk-controlled cross-scale forecast adaptation. Each branch learns a distinct forecasting task on its own future grid, fs(Ds(X))≈Ds(Y)f_s(D_s(X))\approx D_s(Y). These native forecasts create a specific alignment problem: their temporal resolutions differ, and learned alignment can introduce excessive corrections. We address this by retaining deterministic interpolation as a reference and learning magnitude-bounded residual adaptations. Forecast-dependent gates then incorporate the adapted candidates as corrections to the frozen fine-scale forecast under a coarse-contribution budget. The residual bound controls how far each candidate can depart from its alignment reference, while the fusion budget controls its influence on the final prediction. Together, these mechanisms separate learning scale-specific future structure from deciding how that structure should modify the final forecast.

Comparison with multi-scale forecasting and ensemble models. TimeMixer and Scaleformer are multiscale forecasting methods; TimeFuse is an adaptive ensemble framework. RC-ANSF and Scaleformer use TimeMixer-Core. Full TimeMixer and the Core used by RC-ANSF share the same per-scale hyperparameter settings. RC-ANSF uses the Core without PDM/FMM, supplying native-scale supervision and risk-controlled fusion instead.

Entries average H=96,192,336,720H=96,192,336,720. TimeFuse values come from its original paper; NR denotes an unreported result. Bold marks the lowest value in each row.

DatasetMetricRC-ANSFScaleformerTimeMixerTimeFuse
ETTh1MSE0.4210.4430.4700.430
ETTh1MAE0.4370.4560.4700.433
ETTh2MSE0.3450.3920.3710.365
ETTh2MAE0.3940.4200.4070.395
ETTm1MSE0.3440.3570.3690.369
ETTm1MAE0.3790.3860.3950.388
ETTm2MSE0.2560.3090.2720.274
ETTm2MAE0.3210.3520.3310.319
WeatherMSE0.2240.2250.2290.240
WeatherMAE0.2610.2790.2660.270
ElectricityMSE0.1720.1690.1690.163
ElectricityMAE0.2690.2680.2640.260
TrafficMSE0.4320.4310.4150.419
TrafficMAE0.2990.2980.2910.272
Solar-EnergyMSE0.2130.2170.218NR
Solar-EnergyMAE0.2840.2800.284NR

RC-ANSF achieves the best reported MSE on six of the eight datasets, attaining the largest number of dataset-level MSE wins among the compared multiscale forecasting and ensemble methods.

Capacity and training budget. Under these matched settings, the following controlled comparison averages all eight datasets and four horizons.

Mean across all datasetsRC-ANSFFull TimeMixer
MSE0.3010.314
MAE0.3310.338
Total parameters (rounded)335,0001,449,000
Peak training GPU memory (MiB)7791326
Batch forward latency (ms)17.3518.73
Total training time (s)1389905

RC-ANSF uses 76.89% fewer parameters and reduces mean batch forward latency from 18.73 to 17.35 ms (7.37%), while lowering macro MSE by 3.89%. This demonstrates improved forecasting with a smaller model and lower measured inference latency. The total epoch cap is matched, although two-stage training takes longer in wall-clock time.

AuthorsOpenReview ↗

Part2 for W2 and Q2

Thank you for your suggestions to add more recent backbone to make our method more universal.

We evaluated GTR+MLP, SOFTS, and TimeFilter on eight datasets and four forecast horizons, using the source code released by the authors of each respective paper. Each experiment compares the single-scale Base with Safe-linear and RC-ANSF. The two fusion methods share frozen Stage-I predictions.

RC-ANSF: relative MSE reductions and wins

PredictorRC-ANSF vs. Base: MSE reductionRC-ANSF vs. Base: WinsRC-ANSF vs. Safe-linear: MSE reductionRC-ANSF vs. Safe-linear: Wins
GTR+MLP2.27%28/320.93%31/32
SOFTS1.59%28/320.59%28/32
TimeFilter2.18%30/320.61%32/32

Reductions average within-task relative changes; wins count settings where RC-ANSF has lower MSE than the indicated comparator.

Generality across predictors. RC-ANSF improves over the single-scale Base in 28/32 settings for GTR+MLP, 28/32 for SOFTS, and 30/32 for TimeFilter: 86/96 settings (89.6%) overall. Against Safe-linear, it improves in 31/32, 28/32, and 32/32 settings, respectively, totaling 91/96 (94.8%). The positive mean MSE reductions for all three predictors, together with gains in the large majority of settings, demonstrate that the benefits of RC-ANSF generalize across different backbone architectures.

Observed cost. Entries give Base →\to RC-ANSF, averaged across all datasets and forecasting horizons.

PredictorParameters (millions)Train (h)Batch latency (ms)
GTR+MLP1.15→3.101.15\to3.100.134→0.4960.134\to0.49612.24→44.3312.24\to44.33
SOFTS2.45→8.922.45\to8.920.187→0.8260.187\to0.8269.21→31.359.21\to31.35
TimeFilter2.00→6.542.00\to6.540.169→0.7190.169\to0.71919.39→81.5119.39\to81.51

Experiments ran on shared NVIDIA RTX 3090 GPUs, with matched batch sizes within each task for timing comparisons. As we expected, training time includes both stages of RC-ANSF. We acknowledge that the complete multiscale pipeline increases total parameter count and inference latency relative to the single-scale Base, as well as training time.

AuthorsOpenReview ↗

Part3 for W3 and Q3

You raise two related concerns in W3 and Q3: whether the risk-control mechanisms are effective and necessary, given that removing the residual bound can improve accuracy and removing the router has little effect; and which components actually produce the gains, since some original ablations change multiple factors and do not evaluate the auxiliary losses independently. We address these concerns through a direct comparison of correction-control mechanisms, followed by single-factor structural and loss ablations.

Bounded corrections versus unconstrained adaptation and L2 regularization. To test whether explicit amplitude control adds value beyond unrestricted learning or a conventional soft penalty, we compare Bounded, using ρqstanh⁡(Zs/qs)\rho q_s\tanh(Z_s/q_s); Unbounded, using ZsZ_s; and L2, using ZsZ_s with an additional penalty λmean⁡(Zs2)\lambda\operatorname{mean}(Z_s^2). The table below averages all four horizons (96,192,336,72096,192,336,720), with all three seeds included at each horizon on ETTh2.

BackboneMetricBounded (Ours)L2Unbounded
DLinearMSE0.3520.3960.462
DLinearMAE0.4000.4260.455
iTransformerMSE0.3960.4330.557
iTransformerMAE0.4350.4570.513
TimeMixer-CoreMSE0.3450.3810.427
TimeMixer-CoreMAE0.3950.4170.440

On ETTh2, Bounded outperforms L2, and L2 outperforms Unbounded, in both horizon-averaged MSE and MAE for all three backbones. Bounded reduces MSE by 8.64–11.01% relative to L2 and 19.09–28.98% relative to Unbounded. This ordering demonstrates the value of controlling learned corrections: the tested L2 penalty improves over unconstrained adaptation, while an explicit amplitude bound provides a further gain.

Single-factor structural ablation. Separately, to isolate saturation from local scaling, we replace ρqstanh⁡(Zs/qs)\rho q_s\tanh(Z_s/q_s) with ρZs\rho Z_s, preserving the slope at zero. Across the completed paired settings covering all eight datasets and three backbones, removing saturation raises macro MSE from 0.327 to 0.333 and tail excess which is defined below from 0.038 to 0.087: the bound reduces tail excess by 56.60%. This isolates protection against large harmful corrections; the router has a smaller incremental effect on mean accuracy.

Independent auxiliary-loss ablations. We remove each loss separately, fixing the frozen predictors, fusion initialization, data, architecture, training budget, and validation-MSE selection rule. MSE-only training measures their joint contribution. Full RC uses utility/risk/projection weights 0.1/0.1/0.010.1/0.1/0.01.

VariantΔ\Delta MSE (%) ↓\downarrowΔ\Delta MAE (%) ↓\downarrowHarm (%) ↓\downarrowΔ\Delta Tail excess (%) ↓\downarrow
Full RC0.0000.00039.7210.000
Without risk loss+0.052+0.10041.396+7.956
Without utility loss+0.222+0.21241.448+40.550
Without projection loss+0.013+0.01240.398+1.637
MSE only+0.496+0.55943.869+66.970

Scope: TimeMixer-Core, averaged across all eight datasets and four horizons (96,192,336,72096,192,336,720), averaging over three seeds.

Relative changes in MSE, MAE, and tail excess are 100×(ablation/Full RC−1)100\times(\text{ablation}/\text{Full RC}-1), using unrounded means; positive values indicate degradation. Harm is the percentage of windows with higher MSE than the frozen fine-scale predictor. Tail excess averages the largest 10% of positive-part window MSE differences, including zeros. Metrics are averaged over the three seeds within each dataset–horizon setting and then equally over all 32 settings; bold marks the best value in each column.

The single-factor ablations identify complementary contributions from the auxiliary objectives. Removing any loss increases mean MSE, MAE, and tail excess across the full evaluation. Utility supervision has the largest individual effect: its removal increases MSE by 0.222% and tail excess by 40.550%. Removing risk loss raises tail excess by 7.956% and harm by 1.674 percentage points. Removing projection loss also harm the all metrics on average.Especially, removing all auxiliary losses increases MSE by 0.496%, MAE by 0.559%, and tail excess by 66.970%, while raising the harmful-window rate from 39.721% to 43.869%. These controlled interventions directly address the attribution concern in your W3 and Q3 by separating each loss's contribution from their joint effect.

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

The paper studies how to combine forecasts from different temporal scales. It proposes a two-stage framework: each branch first predicts at its native resolution, then a separate fusion module combines the forecasts. The authors introduce Safe-linear and RC-ANSF, which adds bounded adaptive corrections.

The problem is meaningful, and the design is clean. However, the novelty is not fully established. Existing work such as Scaleformer, TimeFuse, and AdaMixT already studies multi-scale forecasting or adaptive fusion. The authors need to clarify what is fundamentally new about native-scale supervision and bounded fusion.

The experiments cover eight datasets, four horizons, and three backbone configurations. This is a reasonable start, but the evidence is not yet sufficient to support the broad claims of risk control and backbone-agnostic generality.

Critical strengths

Clear motivation and modular design. Separating native-scale prediction from fusion is intuitive and makes the framework easier to analyze and reuse.

Reasonable experimental coverage. The paper evaluates multiple datasets and backbones and includes ablation studies, rather than only comparing against single-scale baselines.

Critical weaknesses
  1. Novelty is unclear. The paper needs stronger comparisons with related multi-scale and adaptive fusion methods, especially TimeFuse and AdaMixT. The difference from existing approaches is not sufficiently demonstrated.

  2. Risk control is not guaranteed. Bounding the correction magnitude does not guarantee lower test error. The paper should show directly whether the bound reduces harmful corrections compared with unbounded fusion or simple regularization.

  3. Contribution is not well isolated. Improvements may come from extra branches or model capacity, not necessarily native-scale supervision. Also, TimeMixer-Core removes the original model's internal multi-scale components, so it does not establish effectiveness with the full TimeMixer architecture.

  4. Evaluation on time-series foundation models. The paper claims to provide a backbone-agnostic forecasting framework, but the current experiments cover only DLinear, iTransformer, and a simplified single-scale TimeMixer-Core. I suggest evaluating the framework with at least one or two pretrained time-series foundation models, such as TimesFM and Moirai or Chronos. This would test whether native-scale prediction and controlled fusion remain effective when the underlying predictors have been pretrained on large and diverse time-series corpora. In particular, it would be valuable to compare the original foundation-model forecasts against Safe-linear and RC-ANSF, and to examine whether bounded adaptive alignment provides consistent benefits over simpler fusion. Such experiments would strengthen the generality claim and clarify whether the proposed method adds value beyond the representations learned through large-scale pretraining. The authors should also report any adaptations required to support different input lengths, forecast horizons, or output formats, along with the associated computational costs.

  5. Limited diversity of forecasting settings. The current experiments cover several established multivariate forecasting benchmarks, but it remains unclear whether the proposed framework generalizes to other common forecasting settings. I suggest adding a small number of targeted experiments on (1) univariate time series and (2) panel time series with covariates, such as retail or demand forecasting. These settings would test whether the proposed native-scale fusion mechanism remains effective without cross-variable dependencies and when forecasts must account for additional covariates. Please compare the base model, Safe-linear, and RC-ANSF under consistent protocols, and clarify any adaptations required to handle covariates or heterogeneous series. These experiments would provide stronger evidence for the framework’s generality beyond the current benchmark suite.

Questions for authors

see the weakness section

7 comments
AuthorsOpenReview ↗

Part1 for W1

RC-ANSF trains predictors on native-scale targets, then freezes them and learns bounded alignment and forecast-dependent fusion. TimeFuse learns input-dependent weights to combine forecasts from independently trained heterogeneous forecasting models. AdaMixT adaptively combines multi-patch features from general pretrained and domain-specific experts before a prediction head.

Compared with TimeFuse (ICML 2025), both methods decouple predictor training from fusion; our distinction lies in the native-scale prediction interface and bounded cross-scale alignment. RC-ANSF differs in defining each branch's input and target on its native grid, Ds(X)↦Ds(Y)D_s(X)\mapsto D_s(Y), then interpolating its forecast to the target grid and applying a bounded adaptive correction. TimeFuse weights completed, common-grid forecasts using historical meta-features; RC-ANSF uses forecast agreement/discrepancy for channel-wise gating.

Compared with AdaMixT (IJCAI 2025), RC-ANSF differs in changing both input and target resolution and supervising each native forecast separately before fitting bounded fusion with frozen predictors. AdaMixT varies input patch lengths and fuses expert features before final-target supervision.

AuthorsOpenReview ↗

Part2 for W2

Thank you for your suggestion.In your W2, you suggest comparing our bounded correction with an unbounded correction and simple regularization to assess its empirical benefit. We address this with two controls using the same frozen predictors and fusion architecture.

Our bound. Let ZsZ_s be the learned projected residual at scale ss, LsL_s the linearly interpolated native-scale forecast, and qs,c=max⁡(RMS⁡t(Ls,c),ϵ)q_{s,c}=\max(\operatorname{RMS}_t(L_{s,c}),\epsilon). We constrain each sample's correction by Δs,c=ρqs,ctanh⁡(Zs,c/qs,c),∣Δs,c,t∣≤ρqs,c.\Delta_{s,c}=\rho q_{s,c}\tanh(Z_{s,c}/q_{s,c}),\qquad |\Delta_{s,c,t}|\leq\rho q_{s,c}.

Controls. Unbounded directly uses Δs=Zs\Delta_s=Z_s, without radius scaling or tanh⁡\tanh. For simple regularization, we use the same unbounded correction and add λ,mean⁡(Zs2)\lambda\\,\operatorname{mean}(Z_s^2) to the training loss, penalizing the residual before routing and averaging over scales, samples, time, and channels.

The table presents a selected subset of TimeMixer-Core results on the four ETT datasets and Weather: test MSE and MAE, averaged over four horizons (96,192,336,72096,192,336,720) and three seeds per dataset. We use λ=0.1\lambda=0.1 for L2; apart from the correction constraints described above, all experimental settings are identical across the three methods.

DatasetMetricBounded (Ours)L2Unbounded
ETTh1MSE0.4210.4250.436
MAE0.4370.4420.449
ETTh2MSE0.3450.3810.427
MAE0.3950.4170.440
ETTm1MSE0.3440.3460.364
MAE0.3790.3810.392
ETTm2MSE0.2560.2590.265
MAE0.3210.3200.326
WeatherMSE0.2240.2220.222
MAE0.2610.2650.268

Across the five displayed datasets, the average MSE and MAE rank the methods as Bounded >> L2 >> Unbounded in predictive performance. Bounded reduces mean MSE by 2.59% relative to L2 and 7.25% relative to Unbounded, with corresponding MAE reductions of 1.78% and 4.44%. This ordering supports the importance of constraining adaptive corrections: simple regularization improves on unconstrained correction, while explicit amplitude bounds provide a further benefit in these settings, supporting our risk-control design.

AuthorsOpenReview ↗

Part3 for W3

To solute your concern about our contributions.We directly compare RC-ANSF + TimeMixer-Core (Ours) against the full official TimeMixer, which retains PDM and FMM. We match the data protocol, lookback, forecast horizon, scale count, downsampling, per-scale width and layer count, batch size, learning rate, and weight decay. Each Core branch uses the default single-scale configuration, with residual MLP blocks and a native-horizon prediction head. Our framework supplies the downsampled inputs to these branches using the matched scale count and downsampling settings. The Core omits PDM's cross-scale seasonal/trend mixing; RC-ANSF supplies native-scale supervision and replaces the original multi-predictor summation with bounded alignment and adaptive fusion.

The table displays six datasets, with MSE and MAE averaged over four horizons per dataset.

DatasetMetricOursTimeMixer
ETTh1MSE0.4260.470
MAE0.4420.470
ETTh2MSE0.3480.371
MAE0.3930.407
ETTm1MSE0.3450.368
MAE0.3800.395
ETTm2MSE0.2530.272
MAE0.3200.331
WeatherMSE0.2250.229
MAE0.2610.266
Solar-EnergyMSE0.2160.218
MAE0.2890.284

Efficiency comparison. Resource metrics below average all 24 settings; total training time includes both RC stages.

ResourceOursTimeMixer
Parameters (M)0.3351.449
Peak training GPU memory (MiB)7791326
Batch forward latency (ms)17.3518.73
Total training time (s)1389.0905.0

Across the evaluation, RC-ANSF has lower average MSE, using 76.89% fewer parameters. Mean forward latency is 7.37% lower, although two-stage training takes longer.These results demonstrate a substantial performance improvement relative to TimeMixer, refute the explanation that “a larger parameter count is the cause of the performance gain,” and also show that our native-scale supervision handles multi-scale information in a superior way.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.