acceptodds
Under review as a conference paper at ICLR 2027

N-ring: Redefining Multivariate Time-Series Forecasting with Length-N Directed Ring Pairs

Abstract

Multivariate time-series forecasting is widely used in energy management, traffic monitoring, finance, and environmental sensing, yet many high-performing methods rely on deep architectures, complex correction modules, or temporal downsampling, increasing computational cost or risking information loss. We propose N-ring, a compact shallow model that decomposes long-horizon forecasting into structured local components without reducing the original sampling resolution. The model represents temporal dynamics with reusable directional rings and jointly learns their length, direction, phase, and local waveform, transforming monolithic value regression into coordinated dynamical-state composition. Across 32 input-output configurations and seven mainstream baselines, including PatchTST, DLinear, and MixLinear, the 3294–7244-parameter N-ring model achieves first place in 20/32 MAE and 6/32 MSE comparisons, and second place in 5/32 MAE and 6/32 MSE comparisons; relative to the two best lightweight baselines, MixLinear and SparseTSF, it reduces MAE by 3.02%–11.97% and MSE by 2.46%–9.34% across multiple datasets, with average reductions of 6.40% and 5.74%, respectively. These results establish structured dynamical-state decomposition and coordinated local specialization as an effective paradigm for compact and efficient long-horizon sequence modeling.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time-Series Forecasting, Structured Temporal Decomposition, Directed Temporal Dynamics, Conditional Expert Routing, Lightweight Long-Horizon Forecasting.

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_CmmVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes N-ring, a lightweight and shallow model for multivariate time-series forecasting. Instead of using deep monolithic architectures or temporal downsampling, it decomposes the time series into reusable, directed, and monotonic "rings" at the native resolution. The model routes these structured local components (which explicitly model length, direction, phase, and amplitude) to shared local shape experts, and it incorporates a periodic and short-window state auxiliary path for contextual calibration. With an extremely small footprint of only 3,249 to 7,244 parameters, N-ring demonstrates highly competitive MAE performance against state-of-the-art models across 8 datasets, though its MSE performance is less dominant due to its limited capacity for handling extreme outliers.

Critical strengths
  1. The directed-ring decomposition is an innovative way to conceptualize time-series dynamics, preserving native resolution while effectively extracting local structured components like length, direction, and phase.

  2. The model achieves highly competitive results with a remarkably small parameter footprint, significantly lowering deployment, inference, and storage costs for edge devices.

  3. Extensive experiments across 32 configurations show that N-ring achieves Rank-1 in 20/32 and Rank-2 in 5/32 MAE comparisons, proving its high robustness and accuracy in absolute error terms.

  4. The paper provides detailed ablations of the periodic, state, amplitude, and routing components. Furthermore, the authors transparently diagnose and explain the model's extremum error limitations regarding MSE, making the empirical claims highly credible.

Critical weaknesses
  1. As acknowledged by the authors, the model struggles with MSE compared to deeper models (e.g., iTransformer, PatchTST) because it lacks dedicated mechanisms or deep residual corrections for high-amplitude, rare extremum events.

  2. The hard-coded threshold for ring length (N≤18N \le 18) and the overlapping block subdivision for N>18N > 18 appear somewhat arbitrary. For datasets with long monotonic trends (e.g., Exchange), this clipped expert categorization becomes an approximation and may bottleneck the model's extrapolation capability.

  3. The core ring-routing mechanism heavily relies on a separate periodic calibration and state auxiliary branch to provide global context. Removing these leads to performance degradation on periodic datasets, which makes the overall architecture feel slightly fragmented rather than unified.

  4. While inference is fast, the two-stage training process (backbone then calibration) takes around 618.2 seconds on Electricity, which is comparable to or slower than some deep baselines (e.g., PatchTST takes 656.1s, iTransformer takes 175.8s). The computational advantage is mainly in inference/storage rather than training speed.

Questions for authors
  1. Provide more justification for why N=18N=18 was chosen as the cutoff. Is this primarily a computational constraint to keep the expert bank small, or does the empirical representation power naturally degrade beyond this length?

  2. Could a lightweight robust loss function (e.g., Huber loss, or an asymmetric loss) or a very simple auxiliary extremum head help bridge the MSE gap without destroying the parameter efficiency?

  3. Clarify the training process overhead. Since training takes comparable time to deeper models, can the two-stage training (backbone + calibration) be unified into a single end-to-end optimization step in future iterations to reduce training time?

  4. Please elaborate on the detector smoothing process. How sensitive is the model's performance to the smoothing hyperparameter, and how exactly does high-frequency noise affect the ring decomposition step?

2 comments

Thank you for your valuable feedback In the past two months, we have been committed to improving the MSE performance of the model, and ultimately we have implemented a new calibration method, horizon_calib, To improve the MAE and MSE performance of the model. In the latest results, the model calibrated with horizon_calib has achieved the lowest average MAE and MSE rankings among multiple baselines. Although the MAE performance of the model is still better than MSE, the model has also gained considerable competitiveness in terms of MSE indicators. We have made four major changes based on the feedback from all experts. Firstly, in order to reduce MSE, we proposed horizon_calib to modify the Nring backbone, which slightly increases model parameters and computation time. Across the 32 configurations, 29 showed lower MAE and 24 showed lower MSE. Improvements in multivariate datasets such as electricity and transportation are substantial. Secondly, we added an appendix-only accuracy comparison with the official WPMixer implementation on eight 96-to-96 tasks using three seeds. N-ring achieves lower mean MAE on six datasets and lower mean MSE on four. We could not verify an official implementation of WaveletMixer and therefore do not report results for it. Thirdly, we reported the use of MASE, sMAPE, ND, and NRMSE as metric extensions in GIFT Eval. Fourthly, we added 72 complete test set evaluations and checkpoint level component interventions under clean and damaged inputs for Electricity 192 to 192 and Weather 192 to 96. New evidence suggests that the model is sensitive to recent continuous interruptions. We also clarified statistical uncertainty, sensitivity control, implementation details, and conclusion scope.The following are responses to specific questions:

RQ1: Thank you for your valuable suggestion. The maximum forecasting stride of 18 controls bank capacity, while the encoder retains natural observed-run lengths. We tested 34, 32, 24, and 16 experts with three seeds. Reducing 34 to 16 lowers parameters from 3,558 to 1,980 (44.35%); MAE changes range from a 3.95% improvement to a 2.28% degradation, and MSE from a 9.22% improvement to a 4.09% degradation. These results show a dataset-dependent capacity trade-off. Figure 2 now uses the actual unsmoothed detector, excluding constant runs and runs containing long flat intervals.

RQ2: Thank you for your valuable suggestion. We specify the pointwise MAE, temporal SmoothL1, and first-stage length/direction objectives. Calibration reduces MSE in 24/32 settings, by 4.95% on average, while high-amplitude regions account for 21.92%–71.37% of squared error. The Discussion proposes controlled tests of amplitude-sensitive objectives and a compact extremum head under matched training conditions.

RQ3: Thank you for your valuable suggestion. We report both training stages on eight 96-to-96 tasks. Calibration accounts for 18.69%–52.33% of total training-loop time, averaging 33.02% across datasets. Freezing the same backbone allows separate comparison of refinement methods. Joint end-to-end optimization can be evaluated for cost and convergence under matched conditions in future work.

RQ4: Thank you for your valuable suggestion. Input segmentation uses unsmoothed observations, and first-stage routing labels use fixed width-three smoothing. At 10% input noise, 28.25%–55.22% of interior boundary indicators change, while MSE increases by 1.88%–6.20%. The experiment measures broadband input-noise effects. Smoothing-width and isolated-frequency perturbations are further experiments.

Thank you again for your valuable suggestion. We are always committed to improving model performance from a structural perspective rather than parameter optimization. If you have any new ideas or suggestions for our model, you can also provide them, which may become our future research direction.

Sorry for overlooking the response to the weakness in the review comments earlier. We have expanded the squared-error diagnostics, clarified the treatment of ring lengths, quantified the calibration components’ contributions, and reported training costs separately for both stages. The corresponding changes are detailed below.

RW1: Original Section 4 and Appendix A.7 discussed extremum and boundary errors using an Electricity diagnostic. Revised Section 4 and Appendix A.6, Table A3, extend this analysis to eight datasets: high-amplitude regions account for 21.92%–71.37% of squared error. Revised Section 2.2 specifies horizon-dependent routing and amplitude-aware periodic correction, while Section 3.4 and Table 4 quantify their combined contribution relative to the same backbone. Average relative MSE reductions across the four short-window configurations are 13.79% on Electricity, 12.67% on Traffic, and 8.18% on Weather; the remaining dataset-dependent behavior is discussed in Section 4.

RW2: Original Section 2.1 and Appendix A.4 described overlapping subdivision for long rings, while Section 4 acknowledged approximation from clipped expert categories. Revised Section 2.1 clarifies that the complete model retains natural observed-run lengths and separately uses forecasting strides from 2 to 18; Appendix A.4 specifies the subdivision rule used in the encoding-cap controls. Section 3.4 and Appendix B.3, Tables A6–A7, report three-seed controls on ETTh1 and Exchange. Within the feature-only control configuration, reducing the encoding cap from 18 to 8 changes mean relative MAE/MSE by −0.93%/−0.23%, with the allocated parameter count unchanged at 3,558. These results quantify sensitivity within the tested configuration.

RW3: Original Section 2.2 assigned repeated-pattern context to periodic calibration and short-history context to the state auxiliary branch, but their relationship to prediction could be explained more directly. Revised Section 2.2 presents horizon-dependent expert weighting, periodic references, and amplitude correction within the combined horizon_calib module; Appendix B.1 specifies its complete computation. Section 3.4 and Appendix B.4 quantify the components’ contributions: disabling periodic correction or future routing increases relative MSE by averages of 9.68% and 4.64%, respectively, across the two tested settings. These results support their complementary roles in the fitted forecasting pipeline.

RW4: Original Table 5 already reported the complete two-stage training time, including 618.2 seconds for Electricity, but the discussion needed a clearer separation of model size and training cost. Revised Section 2.4 specifies the backbone and frozen-backbone calibration stages, and Section 3.5 with Table 5 reports their measured times separately. For Electricity 96-to-96, the revised results average 302.3 seconds for backbone training and 331.9 seconds for calibration across three runs. Section 3.5 also reports calibration’s share of training-loop time, while Section 4 identifies matched-total-compute evaluation as further work.

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

This paper proposes N-ring, a lightweight framework for multivariate time-series forecasting that decomposes each channel into directed monotonic segments called rings. Each ring is represented by its length and other information to allow the model to separate local waveform modeling from structural continuation. N-ring uses experts for different ring lengths and directions, and a router to predict future ring structure. The method is evaluated on eight standard forecasting datasets. The paper additionally reports ablations, convergence analysis, representation diagnostics, and efficiency measurements.

Critical strengths
  1. Computational simplicity and efficiency. N-ring reduces forecasting to using only a few thousand parameters.
Critical weaknesses

1 Figure 1 and figure 2 are difficult to understand. Figure 1 contains many components, including ring decomposition, shared experts, the router, periodic calibration, and the state auxiliary branch, but their interactions are not sufficiently clear from the figure alone. Figure 2 seems related to the design choice, but I don’t get how the distribution supports the design choice of using N=2,..., 18, especially the large N>18 for weather and exchange datasets.

2 The theoretical contributions are limited. I don’t see how they support the advantage of the N-ring representation to improve forecasting accuracy over other local decompositions.

3 The baselines are not sufficient to cover the current decomposition in temporal or frequency, for example, WaveletMixer and WPMixer also performed the decomposition, it is hard to tell the ring decomposition merits without such a baseline.

[1] Zhang, Zichi, et al. "WaveletMixer: a multi-resolution wavelets based MLP-mixer for multivariate long-term time series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 21. 2025.

[2] Murad, Md Mahmuddun Nabi, Mehmet Aktukmak, and Yasin Yilmaz. "Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025.

4 There are no standard deviations among the three seeds. Given some performances are very close (the difference between N-ring and iTransformer are 0.228 with 0.229 on electricity with L=192 and H=96), I encourage the author to show the statistical significance for such close comparisons.

5 There is no systematic evaluation under missing data, irregular sampling, noisy turning-point detection, different smoothing settings. These evaluations seem particularly important because the method depends on turning-point detection and the assumed local ring structure.

Questions for authors

1 Can the author provide some interpretations whether figure 2 should be related to the design choice or not?

2 The decomposition and recombination through the router seems to be very familiar. Do you have theoretical evidence that the Ring decomposition will be a better way than other temporal decomposition or frequency decomposition? Or how will the ring decomposition benefit the forecasting objects?

3 Can the author compare N-ring with more closely related work like WaveletMixer or WPmixer?

4 Can the author report the statistical significance for the main comparison, since the performances are really close to those baselines.

5 The current standard benchmark might be over-saturated, so I encouraged the author to report their results on the broader forecasting benchmark such as GIFT-eval.

Aksu, Taha, et al. "Gift-eval: A benchmark for general time series forecasting model evaluation." arXiv preprint arXiv:2410.10393 (2024).

6 Can the author evaluate the robustness of the N-ring under extreme conditions like missing observations, irregular sampling, or noisy turning-point detection?

4 comments

Thank you for your valuable feedback. Your proposed new model WPmixer and forecasting benchmarks like GIFT Eval have broadened our horizons. The existing modification time is not sufficient to support the full-scale evaluation of our model on the GIFT eval baseline. We will include the scores on the GIFT eval baseline in future journal expansion versions. In the revised version of the manuscript, we made four major revisions.Firstly, in order to reduce MSE, we proposed horizon_calib to modify the Nring backbone, which slightly increases model parameters and computation time. Across the 32 configurations, 29 showed lower MAE and 24 showed lower MSE. Improvements in multivariate datasets such as electricity and transportation are substantial. Secondly, we added an appendix-only accuracy comparison with the official WPMixer implementation on eight 96-to-96 tasks using three seeds. N-ring achieves lower mean MAE on six datasets and lower mean MSE on four. We could not verify an official implementation of WaveletMixer and therefore do not report results for it. Thirdly, we reported the use of MASE, sMAPE, ND, and NRMSE as metric extensions in GIFT Eval. Fourthly, we added 72 complete test set evaluations and checkpoint level component interventions under clean and damaged inputs for Electricity 192 to 192 and Weather 192 to 96. New evidence suggests that the model is sensitive to recent continuous interruptions. We also clarified statistical uncertainty, sensitivity control, implementation details, and conclusion scope. The following are responses to specific questions:

RQ1: Thank you for your valuable suggestion. We corrected the Figure 2 counting protocol. The original figure used smoothing and extrema selection; the revised figure uses the implemented unsmoothed, non-strict detector with shared endpoints, excluding constant runs and runs containing at least eight equal consecutive values. Percentages are pooled retained-run counts across training channels. Runs with N>18 account for 0.0007%–1.5112% across the eight datasets. The figure shows the predominance of short runs; 18 is the maximum forecasting stride, while natural input-run lengths are retained. The sensitivity studies quantify the capacity trade-off.

RQ2: Thank you for your valuable suggestion. We clarify how the representation benefits forecasting: raw values retain observed samples, point-aligned descriptors supply local shape and phase context, and shared experts with routing generate forecasts. The reconstruction identity establishes exact representation of observed data using full shape vectors. Its predictive value is evaluated empirically: removing descriptors increases mean relative MSE by 2.75% across the two tested settings. We report this representation effect and explain its role in the model design.

RQ3: Thank you for your valuable suggestion. We added an appendix comparison with the official WPMixer implementation on eight 96-to-96 tasks using three seeds. N-ring achieves lower mean MAE on 6/8 datasets and lower mean MSE on 4/8. The appendix reports all means and sample standard deviations. WPMixer uses at most 10 training epochs, and N-ring follows the stated two-stage protocol. The available training budget allowed completion of these eight tasks. We could not verify an official author implementation of WaveletMixer.

RQ4: Thank you for your valuable suggestion. We added standard deviations, 95% paired confidence intervals, two-sided paired t-tests, and mean ranks. Across 64 comparisons per metric against MixLinear and SparseTSF, lower errors with p<0.05 occur in 49 MAE and 44 MSE cases; higher errors at the same threshold occur in 1 and 3 cases. N-ring has the lowest average MAE/MSE ranks, 1.406/2.781. All tests and supplementary Holm-adjusted p values are provided in the appendix.

RQ5: Thank you for your valuable suggestion. We computed N-ring’s MASE, sMAPE, ND, and NRMSE from 96 verified checkpoints across all 32 local configurations. On Electricity 96-to-96, the three-seed means are 0.9955, 12.60%, 0.0814, and 0.8270. The appendix gives all means and standard deviations. Official GIFT-Eval specifies its own horizons, rolling origins, task collection, and aggregation, requiring additional data preparation and evaluations. The Discussion identifies official-protocol evaluation and matched multi-model metric comparisons for a journal extension.

RQ6: Thank you for your valuable suggestion. We completed 72 full-test evaluations across two settings, three seeds, and 12 input conditions. Gaussian noise at 10% of training standard deviation increases MSE by 1.88%–6.20%; 20% random missingness increases it by 0.21%–10.55%; losing the latest 10% of observations increases it by 26.35%–90.18%. Recent contiguous outages have the largest measured effect. Evaluation on irregular sampling is identified as further work in the Discussion.

Sorry for overlooking the response to the weakness in the review comments earlier. We have revised the figures and methodological explanation, added a related decomposition baseline, and expanded the statistical and robustness analyses. Below, we identify what the original manuscript reported and where the corresponding revisions appear.

RW1: Original Figure 1 combined decomposition, prediction, routing, and calibration, while Figure 2 presented ring-length statistics whose relationship to the implementation was insufficiently explained. Revised Figure 1 and Sections 2.1–2.2 explicitly describe the input pathways, complete forecast candidates, fusion, and calibration. Figure 2 and Table 1 have been recomputed using the model’s unsmoothed detector and explicit constant-run exclusions; Appendix A.3 documents the corrected counting procedure. The retained (N>18) proportions are now 1.51% for Weather and 0.0398% for Exchange. Section 2.1 also distinguishes observed ring length from forecasting stride, while Section 3.4 and Appendix B.3 provide sensitivity evidence for the design choices.

RW2: Original Section 2.1 and Appendices A.2 and A.5 established exact reconstruction from complete observed-ring representations and conditional bounds on shape-approximation error. We interpreted this comment as requesting a clearer connection between these properties and forecasting performance. Revised Section 2.1 explains how observed-ring descriptors enter the predictor, and Section 4 clarifies the scope of the theoretical results. Section 3.4 and Appendix B.4 add empirical attribution evidence: disabling the descriptor pathway increases MAE by 0.24%–4.26% and MSE by 0.44%–5.06% across the two tested settings. Comparative forecasting advantages are supported by these experiments and the baseline comparisons.

RW3: Original Section 3.1 and Table 2 included seven mainstream baselines but did not include the suggested wavelet-based models. Revised Appendix B.5, Table A11, adds the official WPMixer implementation on eight 96-to-96 tasks with three seeds, reporting mean MAE/MSE and sample standard deviations. N-ring achieves lower mean MAE on six datasets and lower mean MSE on four. WPMixer uses a ten-epoch maximum, and N-ring follows the disclosed two-stage protocol; the available training budget permitted these eight tasks. We could not verify an official author implementation of WaveletMixer, so its comparison remains unavailable.

RW4: Original Section 3.1 and Table 2 primarily reported three-seed means without accompanying uncertainty estimates. Revised Appendix B.2, Tables A4–A5, reports mean and sample standard deviation for the complete backbone/full-model matrix; Appendix B.7, Tables A14–A17, reports paired differences, 95% confidence intervals, conventional (p)-values, and supplementary Holm-adjusted values against MixLinear and SparseTSF. Revised Section 3.1 summarizes these comparisons: at nominal (p<0.05), N-ring is favored in 27/32 MAE and 24/32 MSE comparisons against MixLinear, and 22/32 and 20/32 against SparseTSF. These tests cover the two lightweight baselines.

RW5: Original Appendix A.3 described missing-value and detector handling, but the manuscript lacked a systematic corruption evaluation. Revised Section 3.4 and Appendix B.4, Tables A9–A10, add 72 complete test-set evaluations across Electricity 192-to-192 and Weather 192-to-96, covering Gaussian noise, random missingness, recent contiguous outages, and spikes. Noise with standard deviation equal to 10% of the training-channel standard deviation increases MSE by 1.88%–6.20%, whereas losing the latest 10% of observations increases it by 26.35%–90.18%. Appendix A.3 distinguishes unsmoothed input segmentation from smoothed training-label construction. Irregular-sampling evaluation and a label-smoothing-width sensitivity study remain outstanding.

Dear N-Ring Authors,

Thank you for your response. I appreciate your efforts in revising the paper. However, some of my concerns remain unresolved. First, your comparisons mainly involve relatively old baselines, such as MixLinear, DLinear, and PatchTST, which were proposed before 2024. Even against these older baselines, your method achieves better MSE results in only 9 out of 32 cases. This makes me concerned about the claimed state-of-the-art performance. Nevertheless, I appreciate your efforts to include statistical significance tests.

Second, I do not find the two theoretical results sufficient to demonstrate that your proposed decomposition is superior to existing decomposition methods like wavelet decomposition.

Finally, I could not find the GIFT-Eval comparison results that I suggested in the revised manuscript.

Therefore, I prefer to maintain my current score.

Thank you for your suggestion. The current time only allows us to do the simplest sampling training (Aggregated across the 97 tasks, each sampled epoch processes approximately 0.121% of the eligible sliding windows) and conduct a complete GIFT evaluation, covering all 97 task configurations and 326490 rolling test instances. N-ring was trained from scratch using seed 2025. Training comprised two stages: backbone training, followed by training horizon_calib with the backbone frozen, with up to 30 epochs per stage and validation-based early stopping.

We report all 11 evaluation metrics below. To summarize datasets with different scales, we divide each task’s metric by the corresponding Seasonal Naive metric and compute the geometric mean across all 97 tasks. Values below 1 indicate improvements over Seasonal Naive. The MASE and CRPS entries reproduce the official normalized aggregate scores; the remaining entries provide descriptive summaries using the same normalization. All original per-task values are provided in the accompanying results table.

Evaluation metric Normalized geometric mean ↓

MSE — predictive mean 0.6328

MSE — predictive median 0.6328

MAE 0.7891

MASE 0.8140

MAPE 0.8593

sMAPE 0.9936

MSIS 2.1800

RMSE 0.7955

NRMSE 0.7955

ND 0.7891

CRPS — mean weighted quantile loss 0.7006

Using the official leaderboard scoring code and the leaderboard snapshot checked on November 19, 2026, N-ring ranks 2nd among the 12 entries categorized as conventional deep-learning models trained from scratch, based on average task rank for MASE. It is second only to xLSTM-Mixer (1st) and ranks ahead of PatchTST (3rd) and iTransformer (4th).

Model MASE rank within the 12 entries Normalized MASE ↓

xLSTM-Mixer 1st 0.7804

N-ring 2nd 0.8140

PatchTST 3rd 0.8486

iTransformer 4th 0.8932

N-ring achieves lower MASE than PatchTST on 60/97 configurations and iTransformer on 65/97 configurations. These positions were calculated offline using the reference leaderboard. The second-place result applies specifically to MASE; N-ring ranks 10th among these 12 entries for CRPS. Our current implementation produces deterministic point forecasts, so the predictive mean and median coincide, and the interval-based MSIS reflects a degenerate forecast distribution.

Each training epoch used series-balanced sampling of training windows, with an effective batch size of 64 and at least 50 batches. The batch count was increased where necessary to cover every eligible training series at least once per epoch. Evaluation covered all official rolling test instances. Exhaustively traversing all eligible overlapping windows during training would require substantially more computation. Based on measured throughput, a linear projection gives approximately 55 days on two RTX 4090 GPUs if the observed epoch counts were retained, or approximately 88 days if both stages completed all 30 epochs.

The full leaderboard provides a broader comparison: N-ring ranks 143rd of 171 entries for MASE. Among the 142 entries ranked ahead of N-ring, the official metadata identifies 57 zero-shot, 36 pretrained, 7 fine-tuned, 41 agentic, and one conventional deep-learning entry, xLSTM-Mixer. Representative leading systems include Taichu-TimeSeries-Agent-v2, STRIDE w/ Synapse, and RegLab-AH-2 in the agentic category, and PDM-Nano and Timestral-209M in the zero-shot category. The agentic label describes a system-level forecasting strategy and does not, by itself, establish whether its components are pretrained.

We emphasize the conventional deep-learning comparison because our study investigates a compact forecasting model trained from scratch on each task’s training data. The broader leaderboard includes different pretraining and adaptation regimes, together with systems that select or combine forecasts. We therefore report both the overall ranking and the category-specific ranking: N-ring ranks 2nd among the 12 conventional deep-learning entries, behind xLSTM-Mixer (1st) and ahead of PatchTST (3rd) and iTransformer (4th), based on MASE. This comparison clarifies N-ring’s performance within its intended training setting; it does not imply matched computational budgets across models. The stronger full-leaderboard entries remain relevant performance targets.

Furthermore, we have not claimed to have found a result that surpasses all other models. Our main contribution is that we propose an architecture that can achieve comparable or even slightly better performance than mainstream models such as WPMixer in a streamlined manner (parameters and structure).

I hope this answer can partially solve your confusion. We will include the results of this lightweight training in the manuscript and upload them to GIFT Eval as much as possible after completing the full data training.

Reviewer_qGbpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes N-ring, a lightweight model for multivariate time series forecasting based on directed monotonic segments. It combines shared low-rank experts, a router, and periodic and state modules to predict future values at the original sampling resolution. Experiments cover eight datasets, three random seeds, and seven baselines, with comparisons of accuracy, model size, and runtime. The paper reports strong MAE results, while MSE performance is less consistent.

Critical strengths
  1. The paper explores a useful direction for lightweight forecasting through local shape prediction and structural continuation.

  2. The model uses only 3,249–7,244 parameters and provides fast inference relative to several larger baselines.

Critical weaknesses
  1. The paper defines N-ring, but does not clearly explain how it differs from dividing a series into increasing and decreasing segments. It is unclear what this representation adds to forecasting.

  2. The method combines ring decomposition, low-rank experts, routing, and calibration, but the main contribution is unclear. The paper should explain which part is new and how it addresses a specific weakness of existing methods.

  3. Equations (4)–(5) do not explain how inputs are prepared or how predicted segment lengths, boundaries, and amplitudes form a full forecast. These missing details make the method hard to reproduce.

  4. The paper describes a ten-epoch comparison, but Table 5 lists ten backbone epochs followed by ten calibration epochs. The authors should clarify the total training budget and compare models under similar budgets or after sufficient training.

  5. Table 4 mainly tests additional modules while keeping the ring backbone unchanged. Comparisons with fixed-length segments and low-rank models of similar size are needed to show whether ring decomposition itself helps.

  6. The lossless reconstruction result keeps the full shape information, so it does not guarantee good forecasts from the compressed model. The rank-2 choice also needs more support, given the 48.98% shape energy reported on Electricity.

Questions for authors
  1. What does N-ring add beyond ordinary monotonic segmentation, and can you show that the gains come from this design rather than the other model components?

  2. Do the reported gains remain when all models receive similar total training budgets, including both the backbone and calibration stages of N-ring?

1 comment

Thank you for your criticism of our manuscript. We acknowledge that the initial version was not adequately prepared, and your sharp feedback will help us improve the quality of our manuscript. Based on your feedback, we have redrawn the architecture diagram, added corresponding explanations, and conducted relevant experiments. The specific changes include the following four main parts.Firstly, in order to reduce MSE, we proposed horizon_calib to modify the Nring backbone, which slightly increases model parameters and computation time. Across the 32 configurations, 29 showed lower MAE and 24 showed lower MSE. Improvements in multivariate datasets such as electricity and transportation are substantial. Secondly, we added an appendix-only accuracy comparison with the official WPMixer implementation on eight 96-to-96 tasks using three seeds. N-ring achieves lower mean MAE on six datasets and lower mean MSE on four. We could not verify an official implementation of WaveletMixer and therefore do not report results for it. Thirdly, we reported the use of MASE, sMAPE, ND, and NRMSE as metric extensions in GIFT Eval. Fourthly, we added 72 complete test set evaluations and checkpoint level component interventions under clean and damaged inputs for Electricity 192 to 192 and Weather 192 to 96. New evidence suggests that the model is sensitive to recent continuous interruptions. We also clarified statistical uncertainty, sensitivity control, implementation details, and conclusion scope.The following are responses to specific questions:

RQ1: Thank you for your valuable suggestion. The model combines point-aligned monotone-run descriptors with raw values, shared forecasting experts, and horizon-dependent calibration. We disabled the 86-parameter descriptor pathway at inference while holding fitted weights fixed. MAE increases by 0.24%–4.26% and MSE by 0.44%–5.06% in the two three-seed settings. This result shows that the forecaster uses the observed-run structure. We have revised the corresponding passage in the revision draft as follows: On Electricity 192-to-192 and Weather 192-to-96, removing descriptors, periodic correction, or future routing from fitted checkpoints increases relative MSE by 2.75%, 9.68%, and 4.64%, respectively. Descriptor removal raises MAE by 0.24%–4.26%. These fixed-checkpoint interventions show that all three components contribute to the fitted model's accuracy in the tested settings.

RQ2: Thank you for your valuable suggestion. We report actual stage epochs and total measured training cost. The first stage trains the backbone; the second freezes it while each refinement method is trained separately. Each stage has a 10-epoch maximum, with observed averages of 9.89 and 7.86 across 96 runs. Baselines use a 10-epoch maximum, and the cost table includes both N-ring stages. In the summary of results, our model outperforms the lighter model MixLinear and SparseTSF, and maintains comparable or even slightly better performance compared to other mainstream models, with the lowest MAE and MSE average rank. However, due to the parameter differences of several orders of magnitude between these comparative models, it is difficult to fully utilize the performance of the corresponding models by forcibly matching training/inference budgets of the same level with the same parameters. Matched-total-compute comparisons are identified as further evaluation in the Discussion. We have revised the corresponding passage in the revision draft as follows: We report test MAE/MSE on chronological splits, with L,H∈{96,192} and seeds 2025–2027 for the short matrix. Baselines are local reruns of official configurations limited to ten epochs. To compare refinement methods, N-ring uses two stages: train the backbone, then freeze it and train calibration separately. Each stage has at most ten epochs, patience 3, and validation-MAE checkpoint selection; the effective batch size is 64. Across 96 runs, the stages average 9.89 and 7.86 epochs.

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

This paper proposes N-ring, a compact shallow architecture for multivariate long-horizon time-series forecasting. The main idea is to decompose temporal trajectories into directed monotonic “rings” with learned length, direction, phase, and local waveform components, rather than relying on a deep monolithic forecasting backbone or temporal downsampling. The model combines shared ring experts, a lightweight structural router, periodic calibration, and an optional short-window state auxiliary module.

Critical strengths

Interesting structured modeling idea. The directed-ring formulation is a novel and intuitive way to represent local monotonic temporal segments while preserving native sampling resolution. This gives the method a clear inductive bias and distinguishes it from both deep Transformer-style forecasters and simple linear/downsampled models.

Strong parameter efficiency with competitive empirical performance. The reported parameter counts are very small relative to many mainstream forecasting models, yet the method achieves strong MAE results across many dataset/horizon configurations. This is consequential because efficient forecasting models are practically important for deployment in resource-constrained settings.

Broad empirical comparison against relevant baselines. The paper compares against a range of standard and recent forecasting baselines, including PatchTST, DLinear, SparseTSF, MixLinear, TimeMixer, and iTransformer. The matched ten-epoch evaluation and long-history comparisons help position the method within the current forecasting literature.

Critical weaknesses

The empirical advantage is less consistent under MSE. While the MAE results are strong, the MSE results are more mixed, with other models often doing better on datasets such as Electricity, Traffic, and Weather. This matters because squared-error performance is important in many forecasting applications and suggests that N-ring may struggle with large deviations, sharp extrema, or phase-displaced events. The paper acknowledges this to some extent, but the limitation deserves more analysis.

Some methodological details are hard to fully verify from the current presentation. The paper would benefit from clearer pseudocode and implementation-level details for ring extraction, long-ring subdivision, routing, calibration, and the state auxiliary branch. Since the method is somewhat nonstandard, reproducibility depends heavily on these details. The current description is mathematically motivated but not always easy to translate into an exact implementation.

More ablation and sensitivity analysis would strengthen the claims. The paper includes ablations, but I would like to see more systematic sensitivity to the maximum ring length, rank choice, number of experts, calibration components, and the constant-run/missing-value filtering rules. This is important because the method’s success may depend on several design choices that are currently presented as relatively fixed.

Questions for authors

How sensitive is N-ring to the choice of maximum compact ring length N≤18N \leq 18? Would performance degrade significantly if this threshold were smaller or larger?

Can the authors provide clearer pseudocode for the complete training and inference pipeline, including ring extraction, routing target construction if any, long-ring subdivision, and final forecast assembly?

The method appears stronger on MAE than MSE. Do the authors have diagnostic evidence about when large squared errors occur? For example, are failures concentrated around sharp turning points, regime shifts, or high-amplitude channels?

Are the reported improvements statistically significant across seeds for the main comparisons, especially against MixLinear and SparseTSF? Confidence intervals or paired tests would make the empirical claims more convincing.

How much of the performance gain comes from the ring representation itself versus the periodic calibration and state auxiliary modules? A more detailed component-level ablation across datasets would be useful.

2 comments

Thank you for reviewing our manuscript. We have made four major changes based on the feedback from all experts. Firstly, in order to reduce MSE, we proposed horizon_calib to modify the Nring backbone, which slightly increases model parameters and computation time. Across the 32 configurations, 29 showed lower MAE and 24 showed lower MSE. Improvements in multivariate datasets such as electricity and transportation are substantial. Secondly, we added an appendix-only accuracy comparison with the official WPMixer implementation on eight 96-to-96 tasks using three seeds. N-ring achieves lower mean MAE on six datasets and lower mean MSE on four. We could not verify an official implementation of WaveletMixer and therefore do not report results for it. Thirdly, we reported the use of MASE, sMAPE, ND, and NRMSE as metric extensions in GIFT Eval. Fourthly, we added 72 complete test set evaluations and checkpoint level component interventions under clean and damaged inputs for Electricity 192 to 192 and Weather 192 to 96. New evidence suggests that the model is sensitive to recent continuous interruptions. We also clarified statistical uncertainty, sensitivity control, implementation details, and conclusion scope. The following are responses to specific questions:

RQ1: Thank you for your valuable suggestion. We evaluated encoding caps of 18, 16, 12, and 8 and expert banks of 34, 32, 24, and 16 on ETTh1 and Exchange with three seeds. In the feature-input control, reducing the cap from 18 to 8 lowers mean relative MAE/MSE by 0.93%/0.23%. Reducing 34 experts to 16 cuts parameters by 44.35%, with dataset-dependent accuracy changes. Original Figure 2 used smoothing and extrema selection; the revised figure uses the implemented unsmoothed, non-strict detector and excludes constant runs and runs containing at least eight equal consecutive values. The manuscript summary follows, with individual results in the appendix.

RQ2: Thank you for your valuable suggestion. We provide the complete training and inference procedure, covering observed-run extraction, eight descriptors, 34 forecasting experts, structural supervision, and calibration. Both stages use at most 10 epochs, with the backbone frozen during calibration. Relative to the matched backbone, calibration lowers MAE in 29/32 configurations and MSE in 24/32, with mean relative reductions of 5.81% and 4.95%. The main text summarizes these results, and the implementation appendix gives the procedure.

RQ3: Thank you for your valuable suggestion. We added three-seed diagnostics on eight 96-to-96 tasks. High-amplitude positions contribute 21.92%–71.37% of squared error, with enrichment factors of 1.02–10.18; turning-neighborhood enrichment ranges from 0.88 to 1.40. These results identify substantial squared-error concentration in high-amplitude regions. The Discussion relates this finding to the model design and possible optimization directions.

RQ4: Thank you for your valuable suggestion. We added three-seed means, standard deviations, paired confidence intervals, and conventional two-sided paired t-tests. Against MixLinear, lower errors with p<0.05 occur in 27/32 MAE and 24/32 MSE settings; against SparseTSF, the counts are 22/32 and 20/32. N-ring has the lowest average MAE/MSE ranks, 1.406/2.781, among the eight evaluated models. The appendix reports every comparison and supplementary Holm-adjusted p values.

RQ5: Thank you for your valuable suggestion. We added 32 matched-backbone comparisons and fixed-checkpoint component interventions in two settings with three seeds. Calibration lowers relative MAE/MSE by 5.81%/4.95% on average. With fitted weights held fixed, disabling descriptors, periodic correction, or future routing increases mean relative MSE by 2.75%, 9.68%, and 4.64%, respectively. These results quantify the fitted model’s dependence on each component.

Thank you again for your valuable suggestion. We are always committed to improving model performance from a structural perspective rather than parameter optimization. If you have any new ideas or suggestions for our model, you can also provide them, which may become our future research direction.

Thank you for reviewing our manuscript. Although these discussions about weaknesses have been fully addressed in the previous RQs, we still need to summarize the point by point responses for each weakness here. We have expanded the squared-error analysis, clarified the complete computational procedure, and added component and sensitivity experiments. Below, we explain the original presentation and identify the corresponding revisions.

RW1: Original Section 4 and Appendix A.7 acknowledged the weaker MSE performance and provided an Electricity-specific diagnostic of extremum and boundary errors. Revised Section 4 and Appendix A.6, Table A3, extend the diagnostics to eight datasets, showing that high-amplitude regions account for 21.92%–71.37% of squared error. Section 2.2 specifies horizon-dependent routing and periodic/amplitude calibration, while Section 3.4 and Table 4 report its contribution relative to the same backbone: MSE decreases in 24/32 configurations, with average relative reductions of 13.79% on Electricity, 12.67% on Traffic, and 8.18% on Weather. Section 4 discusses the remaining dataset-dependent errors and the possible effects of low-rank prediction, expert averaging, and the training objective. 

RW2: Original Sections 2.1–2.2 and Appendices A.3–A.4 provided the mathematical representation, routing formulation, and handling rules, but the complete computational sequence needed clearer specification. Revised Figure 1 and Sections 2.1–2.2 explicitly describe observed-ring encoding, the retained raw-value pathway, full-horizon expert predictions, weighted fusion, and horizon_calib. Appendix B.1 provides a stepwise training and inference procedure, including the eight descriptor fields, shared (8-2) mapping, expert mixing, routing supervision, backbone freezing, and periodic-reference fallback. Appendix A.3 specifies plateau assignment and training-label smoothing, while Appendix A.4 explains the subdivision rule used in encoding-cap controls; the complete model retains natural observed-run lengths. 

RW3: Original Table 4 primarily compared calibration variants on a fixed backbone, leaving several representation and capacity choices insufficiently tested. Revised Section 3.4 and Appendix B.3, Tables A6–A7, add three-seed controls on ETTh1 and Exchange for encoding caps of 18/16/12/8 and expert counts of 34/32/24/16. Within the stated feature-only control configuration, reducing the encoding cap from 18 to 8 produces mean relative MAE/MSE changes of −0.93%/−0.23%, with 3,558 allocated parameters unchanged; reducing the expert count from 34 to 16 saves 44.35% of parameters. Appendix B.4 separately evaluates descriptor, periodic-correction, and future-routing interventions, together with missing-input and noise robustness. Appendix A.3 documents the constant-run filtering and causal missing-value handling rules. Dedicated low-rank-width and filtering-threshold sensitivity studies remain outstanding, and Section 4 identifies complete-model sensitivity testing as further work.

Comments on the paper

1 comment

Sorry for overlooking the response to the weakness in the review comments earlier. We have made significant revisions to the entire text. We have clarified the contribution and calculation process, expanded the component analysis, and clearly reported on these two training stages. Below, we have addressed each weakness and issue, identified the relevant original content and corresponding revisions.

RW1: Original Sections 1 and 2.1–2.2 described a forecasting architecture combining structural representation, shared low-rank experts, routing, and calibration. We interpreted this comment as requesting a clearer explanation of how the representation contributes to forecasting. Revised Section 2.1 specifies the point-aligned descriptors and their integration with raw observations, while Section 3.4 and Appendix B.4 provide quantitative evidence that the fitted model uses these descriptors.

RW2: Original Section 1 explicitly listed the compact forecasting architecture among the three numbered contributions. We have sharpened this account in revised Sections 1 and 2.1–2.2 by explaining the responsibilities of structural encoding, shared prediction, and horizon-dependent calibration. Revised Section 3.4 quantifies the calibration contribution: MAE improves in 29/32 configurations and MSE in 24/32, with mean relative reductions of 5.81% and 4.95%, respectively.

RW3: Original Section 2.1, Equations (4)–(5), expressed forecast generation through weighted expert predictions, although the accompanying ring-chain illustration left the operational interpretation ambiguous. Revised Figure 1 and Sections 2.1–2.2 explicitly show the input pathways, complete HH-step expert outputs, weighted fusion, and subsequent calibration. Appendix B.1 now specifies the descriptor fields, mappings, mixing coefficients, training stages, and reference-period fallback.

RW4: Original Table 5 already disclosed ten backbone epochs followed by ten calibration epochs. Revised Section 2.4 reports separate ten-epoch limits, backbone freezing during calibration, and actual averages of 9.89 backbone epochs and 7.86 calibration epochs across 96 runs; revised Table 5 reports both stages’ training costs. Strictly matched total-compute comparisons remain an outstanding evaluation, as acknowledged in Section 4.

RW5: Original Table 4 evaluated calibration variants on the same frozen backbone, so additional representation-specific evidence was needed. Revised Section 3.4 and Appendix B.4 separately disable descriptors, periodic correction, and future routing while holding fitted weights fixed. These interventions establish component contributions in the tested models.

RW6: Original Appendix A.2 already stated that the reconstruction proposition concerns complete shape vectors and that finite expert banks and rank-RR factors are learned approximations. We have made this qualification more prominent in revised Section 2.1 and Appendix A.2, with its relationship to empirical forecasting accuracy explained in Section 4. The reported 48.98% rank-two shape energy supports the reviewer’s request for further rank sensitivity evidence; the encoding-cap and expert-count experiments address different factors.

RQ1: Original Sections 2.1–2.2 described the combined forecasting mechanism; revised Section 3.4 and Appendix B.4 now examine its components separately. Across three seeds on Electricity 192-to-192 and Weather 192-to-96, disabling the descriptor pathway increases MAE by 0.24%–4.26% and MSE by 0.44%–5.06%. Disabling periodic correction and future routing increases relative MSE by averages of 9.68% and 4.64%, respectively, across these two settings.

RQ2: The main contribution of the paper is to declare that our model performance exceeds several publicly available baselines, including those with fewer parameters and training budgets, such as SpraseTSF, as well as those with more parameters and training budgets, such as PatchTST. In our opinion, due to the difference in parameters by several orders of magnitude during the model period, it is difficult to forcefully match the parameters and training budget of these models to leverage the advantages of the corresponding models, and it is also difficult to achieve in the short term. Of course, if you recommend other models of comparable size, we would be happy to provide corresponding experiments.Revised Section 2.4 and Table 5 make the two-stage schedule and measured costs explicit, while Section 4 identifies matched-total-compute evaluation as remaining work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.