acceptodds
Withdrawn from ICLR 2027

GatedLinear: Adaptive Routing of Complementary Linear Bases for Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series forecasting requires models to capture diverse temporal dynamics, from smooth trend continuation to nonstationary drift and strict phase-aligned recurrence. While recent deep learning models have improved accuracy, many process these diverse patterns through a shared computational backbone with fixed architectural inductive biases. Representative mixture-of-experts models introduce multiple experts but retain homogeneous expert architectures, leaving their forecasting roles to emerge through training. To address temporal heterogeneity directly at the level of prediction mechanisms, we propose GatedLinear: a lightweight framework that frames forecasting as the adaptive routing of complementary linear bases. GatedLinear leverages three heterogeneous forecasting experts: a global trend-seasonal basis for smooth projection, a difference-based incremental basis for nonstationary drift, and a phase-aligned recurrence basis for explicit cyclic reuse. These experts build distinct extrapolation priors directly into their forecasting operations. To combine their predictions, we introduce a Tri-Factorized Fusion Gate that decomposes routing logits into channel-specific preferences, horizon-aware offsets, and phase-indexed biases derived from known future time marks. This design enables point-wise soft routing across variables, forecast steps, and periodic phases without stacking computationally heavy neural modules. Experiments on eight standard benchmarks show that our method achieves the lowest average MSE and MAE among the evaluated baselines, while offering interpretable routing patterns and a favorable accuracy–cost trade-off.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Adaptive Routing of Complementary Linear Bases for Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BBDCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GatedLinear, a lightweight framework that frames forecasting as the adaptive routing of complementary linear bases. It leverages three heterogeneous forecasting experts: a global trend-seasonal basis for smooth projection, a difference-based incremental basis for nonstationary drift, and a phase-aligned recurrence basis for explicit cyclic reuse, and introduces a Tri-Factorized Fusion Gate that decomposes routing logits into channel-specific preferences, horizon-aware offsets, and phase-indexed biases derived from known future time marks. Experiments on eight standard benchmarks report the lowest average MSE and MAE among the evaluated baselines, while offering interpretable routing patterns.

Critical strengths
  • GatedLinear has a simple and computationally lightweight architecture. It produces three full-horizon candidate forecasts using four channel-shared temporal linear layers in total, with the branches capturing moving-average trend-seasonal projection, differenced incremental evolution, and phase-aligned cyclic recurrence. A tri-factorized softmax gate then combines these candidates through convex weights.
  • The fusion gate provides an interpretable soft weighting mechanism. Its channel, horizon, and phase factorization reveals the learned mixture at each variable and forecast step, although these input-independent weights reflect learned preferences rather than dynamic mechanism selection.
Critical weaknesses
  • The claimed novelty is narrow relative to existing work. RTSF [1] already shows that linear layers learn periodic patterns effectively and that reversible normalization improves trend forecasting. TSMixer [2] further provides a theoretical justification that time-step-dependent linear projections can capture periodic and smooth trend structure, and explicitly warns that standard LTSF benchmarks under-reward cross-variate information, with its M5 results showing that conclusions change under richer interactions. MoLE [3] already combines multiple linear experts with a timestamp-conditioned channel-specific router, and MoHETS [4] motivates heterogeneous experts for trends, periodicities, and nonstationary regimes. ETSformer [5] also established decomposition into interpretable level, growth, and seasonality components. These works narrow the contribution of this paper. However, RTSF, MoHETS, and ETSformer are not discussed at all, and the paper does not compare against TSMixer or MoLE in its main experiments.
  • The fixed lookback window size compromises the fairness of the empirical comparison. The main evaluation fixes the lookback length at 336 for all datasets and models, although the optimal lookback depends on both the model and the dataset. Prior work shows that linear models benefit from longer histories [1, 2], while fixed windows can reverse rankings between channel-independent and channel-dependent models [6]. This issue is also reflected in the baseline results. For example, with a lookback length of 336 and prediction length 96, PatchTST is reported as 0.139/0.242 on ECL and 0.310/0.364 on ETTh2, whereas the original paper reports 0.130/0.222 and 0.274/0.336 [7]. Moreover, TimeMixer, iTransformer, and CycleNet use a lookback of 96 in their main comparisons, so the reported results are not directly comparable. Since GatedLinear is a channel-independent linear model while several baselines prefer shorter windows, the fixed setting may systematically favor GatedLinear.
  • The benchmark suite is too narrow and has known validity issues. Although eight datasets are used, four are ETT variants and one is Electricity, so five belong to the same broad energy domain. PatchTST also cautions that Exchange is weakly discriminative because simply repeating the last value can match or outperform the best methods [7]. Recent audits further show that these benchmarks are weakly channel-dependent and prone to ranking reversals [6, 8]. The paper neither adopts newer supervised benchmarks such as TFB and BasicTS+ [9, 10] nor includes broader collections such as GIFT-Eval or Monash [11, 12]. This limits claims about heterogeneous dynamics, especially because Weather is a single-station dataset with 21 highly regular indicators, so strong performance there is consistent with the periodic behavior of linear mappings [1] rather than broad multivariate capability.

[1] Li, Z., Qi, S., Li, Y., & Xu, Z. (2023). Revisiting long-term time series forecasting: An investigation on linear mapping. arXiv preprint arXiv:2305.10721.

[2] Chen, S. A., Li, C. L., Yoder, N., Arik, S. O., & Pfister, T. (2023). Tsmixer: An all-mlp architecture for time series forecasting. arXiv preprint arXiv:2303.06053.

[3] Ni, R., Lin, Z., Wang, S., & Fanti, G. (2024, April). Mixture-of-linear-experts for long-term time series forecasting. In International conference on artificial intelligence and statistics (pp. 4672-4680). PMLR.

[4] Ortigossa, E. S., Lutsker, G., & Segal, E. (2026). Mohets: long-term time series forecasting with mixture-of-heterogeneous-experts. arXiv preprint arXiv:2601.21866.

[5] Woo, G., Liu, C., Sahoo, D., Kumar, A., & Hoi, S. (2022). Etsformer: Exponential smoothing transformers for time-series forecasting. arXiv preprint arXiv:2202.01381.

[6] Abdelmalak, I., Madhusudhanan, K., Choi, J., Klötergens, C., Yalavarthi, V. K., Stubbemann, M., & Schmidt-Thieme, L. (2026). Channel Dependence, Limited Lookback Windows, and the Simplicity of Datasets: How Biased is Time Series Forecasting?. In Pacific-Asia Conference on Knowledge Discovery and Data Mining (pp. 585-597). Springer, Singapore.

[7] Nie, Y., Nguyen, N. H., Sinthong, P., & Kalagnanam, J. (2022). A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730.

[8] Brigato, L., Morand, R., Strømmen, K., Panagiotou, M., Schmidt, M., & Mougiakakou, S. (2025). There are no champions in supervised long-term time series forecasting. arXiv preprint arXiv:2502.14045.

[9] Qiu, X., Hu, J., Zhou, L., Wu, X., Du, J., Zhang, B., ... & Yang, B. (2024). Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods. arXiv preprint arXiv:2403.20150.

[10] Shao, Z., Wang, F., Xu, Y., Wei, W., Yu, C., Zhang, Z., ... & Cheng, X. (2024). Exploring progress in multivariate time series forecasting: Comprehensive benchmarking and heterogeneity analysis. IEEE Transactions on Knowledge and Data Engineering, 37(1), 291-305.

[11] Aksu, T., Woo, G., Liu, J., Liu, X., Liu, C., Savarese, S., ... & Sahoo, D. (2024). Gift-eval: A benchmark for general time series forecasting model evaluation. arXiv preprint arXiv:2410.10393.

[12] Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., & Montero-Manso, P. (2021). Monash time series forecasting archive. arXiv preprint arXiv:2105.06643.

Questions for authors

Q1: Can the authors tune the lookback length separately for each dataset and model and report whether the ranking changes under model-specific lookbacks? Please also explain why the reported PatchTST and PaiFilter results differ significantly from their original results.

Q2: Why does the paper replace the standard prediction length 720 used in other works with 48?

Q3: Can the authors evaluate a subset of a broader benchmark such as Monash or GIFT-Eval to test whether the reported advantages extend beyond the current benchmark suite?

Reviewer_R297OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces GatedLinear, a lightweight time series prediction framework. This method combines three complementary linear prediction modules: trend-seasonal projection, difference-based extrapolation, and phase-aligned recurrence. It utilizes a Tri-Factorized Fusion Gate to assign weights for integrating their prediction results. Experiments conducted on eight prediction benchmarks demonstrate that this framework is competitive in terms of prediction accuracy and computational efficiency.

Critical strengths
  1. The combination of three linear models provides complementary inductive biases while maintaining a lightweight model design.

  2. GatedLinear achieved good average predictive performance on eight benchmarks. Expert module comparisons, ablation studies, and efficiency analyses provided strong empirical support for the framework.

Critical weaknesses
  1. The innovation in methodology is limited. The three modules are primarily built upon existing technologies, and the proposed gating mechanism is quite simple, making the methodological contribution appear to be merely a relative improvement.

  2. The text in Figure 1 is very small, and the spacing between the table caption and the table body is very tight. However, the spacing between the image caption and the image is quite large. The textual description also exhibits a strong AI flavor.

  3. The main experiment adopts a fixed input length, with a maximum prediction length of 336. The generalization ability of this method is questionable.

Questions for authors

See weaknesses.

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

GatedLinear combines three lightweight forecasting branches: direct projection of decomposed trend and seasonal components, prediction and integration of increments, and reuse of recent periodic templates with a learned residual. A soft gate combines their outputs using channel, forecast-horizon, and future-phase indices. Experiments on eight benchmarks report the lowest average MSE and MAE among the evaluated methods, supported by expert comparisons, ablations, routing analyses, and efficiency measurements.

Critical strengths

The paper offers a practical, inexpensive way to combine forecasting priors and evaluates more than aggregate accuracy. Table 2 compares expert pools under the same gate, Table 3 tests branch and routing ablations, Appendix G.2 compares input-conditioned gates, and Appendix G.7 examines period misspecification. These controls provide useful evidence that the proposed parameterization can help, while the acknowledged weaknesses on Traffic and incorrectly specified periods clarify its limitations.

Critical weaknesses
  1. The contribution beyond a context-dependent linear ensemble needs sharper isolation. For fixed channel, horizon, and phase, Eqs. 3-10 give an affine map of the normalized history; in particular, the unrestricted residual in Eq. 8 can absorb or cancel the recurrence template. Table 2's GatedDLinear comparison is relevant, but uses six affine layers versus four here. Branch-removal ablations also change capacity. The proposed parameterizations may provide useful optimization priors, but Figure 5's gate correlations do not establish that trained branches retain distinct forecasting roles. Thus the experiments support the complete design more strongly than their proposed mechanism interpretation.

  2. The headline improvement is modest and its reliability remains unclear. Table 1 reports average MSE of 0.260 versus 0.264 for TimeAlign and 0.265 for TQNet, but Appendix F.2 gives only GatedLinear's three-seed mean, without variability. Table 2's advantage over GatedDLinear is similarly unaccompanied by uncertainty. Appendix F.4 states common training settings, but baseline result provenance and model-specific tuning are insufficiently explicit. These omissions leave both the competitive advantage and the central expert-design comparison difficult to assess.

  3. The phase-template equation disagrees with the stated indexing convention. With history X1:LX_{1:L} and r starting at zero, Eq. 7 averages indices L-KP through L-1, rather than the last KP observations. For ECL/Traffic, the stated defaults L=336, P=168, K=2 access X0X_0. Adding one appears to correct the equation. This could be a notation error, but the evaluated implementation and first forecast phase need confirmation because alignment is central to this branch.

Questions for authors
  1. Mechanism clarification/check: Do trained branches retain different extrapolation behavior, and how much do the proposed operations help at matched capacity? A control retaining the same gate and four affine layers, but replacing increment reconstruction and recurrence reuse with direct projections, on one or two existing datasets would help. Existing branch-level evidence directly testing the assigned forecasting roles would also clarify the mechanism interpretation.

  2. Protocol clarification/check: Which Table 1 results were rerun, and how were splits, normalization, early stopping, and model-specific tuning controlled? Please provide seed-level variability for GatedLinear, GatedDLinear, and the closest rerun competitors. This would establish whether the reported advantages survive training variation under comparable conditions.

  3. Implementation clarification: Does the phase template use indices L-KP+1 through L? Please correct Eq. 7, confirm how its first position aligns with the first forecast step, and state whether correcting the printed expression changes any reported results.

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

Related papers

Discussion (0)

Sign in to comment.