acceptodds
Under review as a conference paper at ICLR 2027

RouteTS: Budgeted Spectral Component Routing across Time and Frequency Domains for Time Series Forecasting

Abstract

Time-domain and frequency-domain predictors offer complementary inductive biases for recurring patterns and localized variations. However, existing architectures generally determine how the two domains are used through fixed design choices, rather than adapting spectral-component allocation to the input spectrum. We propose RouteTS, a lightweight framework that makes this allocation explicit through amplitude-based component routing. For each input window and variable, RouteTS ranks spectral components by amplitude, sends the Top- components to a shared complex-valued linear predictor, and transforms the complementary spectrum back to the time domain for modeling by a shared MLP. The dataset-specific routing budget is selected using only validation data, and varying yields time-only, mixed-domain, and frequency-only configurations within the same architecture. Across diverse benchmarks, RouteTS achieves the best average MSE on the majority of datasets while remaining among the most parameter- and memory-efficient models evaluated. Controlled analyses further show that the preferred allocation varies across datasets. In the mixed-domain cases examined, assigning dominant spectral components to the frequency predictor and the remainder to the temporal predictor improves accuracy over alternative component selections and reversed assignments.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, frequency-domain modeling, time-frequency modeling, spectral component routing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_okZaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RouteTS, a lightweight channel-independent forecaster. For each input window and variable, RouteTS ranks spectral components by amplitude, sends the Top-K components to a shared complex-valued linear predictor, and transforms the complementary spectrum back to the time domain for modeling by a shared MLP.

Critical strengths

This paper provides many details for others to reproduce results. For example, this paper reports the validation landscape over K for every dataset (Figure 3), compares against endpoint-only selection (Table 2), includes Broadcast, Swap, random and low-frequency controls, keeps the validation-selected K = 4 on Exchange although it is 23% worse on test, and states that results are single runs.

This paper provides an interesting observation of the benchmark. A RevIN + two-layer MLP, or a single complex linear layer, chosen per dataset on validation data, matches or beats ten recent models on most datasets at a fraction of the parameters and memory.

Critical weaknesses

This paper claims that the proposed spectral-component routing improves forecasting. However, the evidence presented in this paper cannot support that claim. First, none of the seven wins depends on mixed routing. Validation selects an endpoint on nine of twelve datasets. Six of the seven best-MSE results in Table 1 are endpoints: ILI, COVID-19 and ETTm1 use K = 0 (a plain MLP), and ETTh1, ETTh2 and ETTm2 use K = full (a complex linear layer, within 0.001 of FITS on ETTh2 and ETTm2). The seventh, Solar (K = 1, 0.201), would also be the best with the K = 0 endpoint (0.204 in Table 2; next best baseline 0.210). Second, where routing is active, the gain is within noise or negative. As Table 2 lists, Traffic improves by 0.001 MSE (0.2%), Solar by 0.003 MSE (1.5%) with unchanged MAE, and Exchange degrades by 23.1%. All trained models are single runs; seeds are varied only for the random routing mask, so there is no estimate of training variance against which 0.001–0.003 can be judged. Third, this paper's own controls point away from routing. Broadcast, which removes the partition and feeds the full input to both branches, is within 1.1% on Solar and slightly better on Traffic (Figure 4). On Solar, routing a random or the lowest-frequency component gives 0.218 / 0.216, which is worse than using no frequency branch at all (0.204). What the data support is "a linear spectral path beside the MLP sometimes helps a little", not complementary component allocation. Fourth, K behaves as a noisy hyperparameter. Appendix B.1 finds no relation between the selected K and spectral or autocorrelation properties (ETTm1 and ETTm2 choose opposite endpoints), validation error changes by up to 422% across K on COVID-19 where the validation split has on the order of a hundred time steps, and on Exchange the selection is driven by one horizon and fails on test. The abstract's statement that the preferred allocation "varies across datasets" is thus hard to separate from selection noise.

The comparisons against baselines are unfair. Large improvements come from settings where baselines appear to be untuned. RouteTS is the best of eight configurations per dataset on validation, with dataset-specific learning rates (1e-4 to 1e-2) and batch sizes whose selection procedure is not described. Baselines use "official implementations and recommended hyperparameters", but at a look-back of 336 (48 for ILI and COVID-19) and on datasets (COVID-19, PM2.5) for which most of them have no recommended configuration. Besides, TFPS scores 1.122 on Traffic and FreqMoE 3.927 on ILI; FreqMoE gives 0.715 / 0.742 on ETTh1 at horizons 336 / 720; iTransformer has MSE 0.389 with MAE 0.341 on ETTm2-336. On COVID-19 (raw cumulative counts, 1,143 steps, 8 series) the second-best model is DLinear, and seven of the ten baselines are 2.5 to 14 times worse than RouteTS. The two largest reported gains (ILI −16%, COVID-19 −31%) are on these two small datasets and are single runs.

Important related work is missing. For example, FAN (presented at NeurIPS 2024 by Ye et al.) selects the top-K amplitude components of each input instance, predicts them with a separate small network, and gives the remainder to the backbone. Koopa (presented at NeurIPS 2023 by Liu et al.) splits the spectrum by amplitude and sends the two parts to different predictors.

Questions for authors

Please report mean ± std over at least 3–5 training seeds for K = 0, the selected K and K = full on Solar and Traffic, and for the Broadcast variant. Is the mixed-domain gain larger than seed variance?

Please report test MSE for every K on every dataset (as done for Exchange in Figure 10). These results allow readers to understand how often validation selection finds the test-best budget and how flat the test landscape is.

How were the per-dataset learning rate and batch size of RouteTS chosen? Were baselines given a comparable search (including learning rate) at look-back 336 / 48 on each dataset?

If each strong baseline (e.g., FITS, CFPT, Amplifier) were also given eight validation-selected configurations, or if the comparison were "best of {DLinear, FITS} on validation", how many of the seven wins remain?

How does RouteTS differ from, and compare with, FAN and Koopa's amplitude-based spectral split?

With K = full the model is close to FITS without the low-pass cut-off. What explains the remaining difference on ETTh1 (0.407 vs 0.413)?

For K = 1 on Solar and Traffic, which frequency bin is selected, and how often does it change across windows? If it is almost always the daily bin, the "input-adaptive" mask is effectively a fixed one.

5 comments
AuthorsOpenReview ↗

Response to Reviewer okZa (Part 1)

Thank you for your detailed review and constructive comments. In the revised manuscript, we introduce a rule for selecting KK. It changes the selected configurations for PM2.5, Weather, ECL, and Exchange, and we update Table 1 and the corresponding appendix tables accordingly. After these updates, RouteTS still achieves the lowest average test MSE on seven of the twelve datasets and the lowest average test MAE on five.

Weakness. Practical value of routing when most datasets select a single domain.

RouteTS allows the choice of forecasting domain to vary by dataset within a single framework. Algorithm 1 uses training-set probes to shortlist a single-domain endpoint and compares it with mixed-domain routing on validation data. The revised rule selects a single-domain configuration on eight datasets and mixed-domain routing on four.

On all four datasets where the rule selects K=1K=1, the mixed configuration improves mean test MSE over the shortlisted endpoint (Table 3). On Solar and Traffic, it improves over both endpoints. The reductions are 25.56% and 4.96% relative to frequency-only forecasting, and 1.47% and 0.25% relative to the stronger time-only endpoint, respectively. This is the practical value of supporting all three configurations in RouteTS: the same framework can use a single domain on datasets where validation favors it and combine domains on datasets where the two branches can improve forecasting.

Q1. Training-seed variability and mixed-domain gains.

At the 336-step horizon, mixed-domain routing (K=1K=1) achieves the lowest mean test MSE among the three routing configurations on both Solar and Traffic. We use three training seeds (2021, 2024, and 2027) and report the 336-step setting because it is the reference horizon for selecting KK and covers multiple daily cycles on both datasets.

  • Frequency-only endpoint. Table 3 shows that mixed-domain routing reduces MSE averaged across the four forecast horizons by 25.56% on Solar and 4.96% on Traffic. At the 336-step horizon, the absolute MSE reductions are 0.082 on Solar and 0.030 on Traffic for each of the three seeds. Both gains far exceed the largest observed seed-to-seed MSE standard deviation of approximately 0.0006.
  • Time-only endpoint. Mixed-domain routing also lowers mean test MSE by 0.0020 on Solar and 0.0007 on Traffic. The gain holds across all three Solar seeds. On Traffic, it appears in two seeds, with a tie in the third at the reported precision, and is comparable to seed-to-seed variation.
DatasetConfigurationTest MSE (mean ± SD)Test MAE (mean ± SD)
SolarTime-only (K=0K=0)0.2093 ± 0.00060.2550 ± 0.0010
SolarMixed (K=1K=1)0.2073 ± 0.00060.2540 ± 0.0000
SolarFrequency-only0.2893 ± 0.00060.3053 ± 0.0006
TrafficTime-only (K=0K=0)0.4067 ± 0.00060.2770 ± 0.0000
TrafficMixed (K=1K=1)0.4060 ± 0.00000.2743 ± 0.0006
TrafficFrequency-only0.4360 ± 0.00000.2900 ± 0.0000

The consistent gains over frequency-only routing on both datasets, together with the gains over the stronger time-only endpoint on Solar, support retaining mixed-domain routing as an explicit option in RouteTS.

Q2. Routing-budget selection and test sensitivity.

Figure 5 of the revised manuscript shows test MSE for every evaluated KK on all 12 datasets. Each cell reports the percentage increase over the best result for that dataset, using test MSE averaged across the four forecast horizons. The heatmap makes it easy to see how sensitive each dataset is to changes in KK and which budgets perform close to the test optimum.

Algorithm 1 selects one KK per dataset. Training-set probes shortlist a single-domain endpoint, which we compare with K=1K=1 using validation MSE at one reference horizon. With just two candidates evaluated at one validation horizon, the rule matches the test-best KK on 8 of 12 datasets and achieves mean test MSE within 1.3% of the best evaluated budget on 11 of 12 datasets (Table 2).

AuthorsOpenReview ↗

Response to Reviewer okZa (Part 2)

Q3–Q4. Hyperparameter selection and fairness of baseline comparisons.

For RouteTS, we used defaults based on DLinear’s settings on most datasets. On a few datasets, including ILI and COVID-19, we tried limited adjustments, such as batch sizes of 16 and 32 and learning rates of 10−310^{-3} and 10−210^{-2}. Table 6 lists the final settings across datasets, with batch sizes ranging from 8 to 64 and learning rates from 10−410^{-4} to 10−210^{-2}. This procedure required little dataset-specific tuning across the benchmarks. For the baselines, we used the authors’ recommended settings where available and made targeted adjustments for datasets without published settings.

We ran an additional search on ILI and COVID-19 to test whether the comparison holds after tuning two lightweight baselines. These smaller datasets are particularly relevant to the reviewer’s concern because published settings are unavailable for some baselines and simple models can be competitive on them. We used a look-back window and forecast horizon of 48 and tested four learning rates (10−410^{-4}, 10−310^{-3}, 5×10−35\times10^{-3}, and 10−210^{-2}) with batch sizes of 16 and 32. For FITS on ILI, batch size 32 produced NaN validation loss at all four learning rates, so we replaced it with 8. The supplementary material includes the full training logs.

For each baseline, the table reports the lowest observed test MSE across its grid and the MAE from the same run.

DatasetRouteTS (MSE / MAE)DLinear (MSE / MAE)FITS (MSE / MAE)
ILI1.557 / 0.7962.456 / 1.0502.292 / 0.993
COVID-190.146 / 0.1580.146 / 0.2500.451 / 0.294

At this horizon, RouteTS achieves 32.1% lower MSE on ILI than the best result in either baseline grid. On COVID-19, it matches DLinear’s lowest MSE to three decimal places and reduces MAE from 0.250 to 0.158. Thus, any validation-based choice among the tested baseline configurations would leave RouteTS ahead on ILI and at least tied on COVID-19 in MSE at the reported precision. This targeted comparison shows that RouteTS remains competitive on both datasets after tuning DLinear and FITS. The seven wins in the main table are based on mean MSE across four forecast horizons.

Q5. Comparison with FAN and Koopa.

RouteTS makes component-to-domain assignment explicit. For each input window and variable, it sends the top-KK Fourier coefficients directly to a complex-valued frequency predictor and reconstructs their complement for the temporal MLP. Varying KK allows time-only, frequency-only, or mixed-domain forecasting within the same framework.

  • FAN also selects dominant frequencies by amplitude for each input window. It reconstructs the selected components in the time domain and forecasts them with an MLP that also sees the full input history. Its backbone forecasts the residual.
  • Koopa identifies dominant frequencies from amplitudes aggregated over the training set and reuses that split across input windows. It models the resulting components with time-invariant and time-varying Koopman predictors.

We evaluated all three methods using the same data splits, lookback lengths, and forecast horizons. The table reports test MSE / MAE averaged over four horizons. RouteTS achieves the lowest mean MSE on 11 of 12 datasets and the lowest mean MAE on all 12. DLinear + FAN achieves the lowest mean MSE on Weather. Table 11 of the revised manuscript reports the per-horizon results.

DatasetRouteTSDLinear + FANKoopa
ILI1.640 / 0.8073.124 / 1.2252.085 / 0.909
COVID-190.125 / 0.1351.101 / 0.5930.871 / 0.368
PM2.50.418 / 0.4350.421 / 0.4630.524 / 0.472
Weather0.240 / 0.2760.234 / 0.2910.249 / 0.278
Traffic0.402 / 0.2740.437 / 0.3050.636 / 0.397
Solar0.201 / 0.2490.206 / 0.2550.299 / 0.307
ECL0.164 / 0.2560.171 / 0.2700.622 / 0.393
ETTh10.407 / 0.4190.469 / 0.4760.459 / 0.456
ETTh20.338 / 0.3830.427 / 0.4410.377 / 0.412
ETTm10.352 / 0.3800.363 / 0.3920.374 / 0.397
ETTm20.257 / 0.3150.281 / 0.3410.316 / 0.345
Exchange0.398 / 0.4230.581 / 0.5150.458 / 0.448

The Swap ablation in Section 4.4 tests the role of domain assignment directly. It keeps the spectral split fixed and reverses which branch receives each component, increasing mean test MSE by 27.8% on Solar and 8.0% on Traffic. This supports routing the dominant coefficients to the frequency predictor and their complement to the temporal predictor in these mixed-domain settings.

AuthorsOpenReview ↗

Response to Reviewer okZa (Part 3)

Q6. Difference from FITS at K=fullK=\mathrm{full}.

RouteTS and FITS both use a complex-valued linear layer, but they learn different output mappings even with the same input frequencies. FITS interpolates the look-back window of length LL into a sequence of length L+HL+H and takes the final HH steps as the forecast. This formulation allows FITS to supervise both the reconstructed look-back window and the future horizon. At K=fullK=\mathrm{full}, RouteTS maps the full input spectrum directly to the spectrum of the HH-step forecast and trains on forecast error alone. These differences can contribute to the 0.006 mean test MSE gap on ETTh1 (0.407 vs. 0.413). Isolating the contribution of either design choice would require a matched ablation.

Some frequency-domain forecasters discard part of the input spectrum to reduce model size. FITS, for example, uses a low-pass cutoff and reports only marginal accuracy gains from raising it in its ETTh2 case study (FITS, Section 4.3). That experiment measures the effect of passing more coefficients to the frequency predictor. RouteTS asks a complementary question: can the components left out of the frequency branch still improve the forecast through a temporal predictor? At intermediate KK, RouteTS reconstructs those components and routes them to the temporal MLP. On Solar and Traffic, this mixed-domain configuration achieves lower mean test MSE than either single-domain endpoint (Table 3), demonstrating the value of retaining and modeling the complementary spectrum on these datasets.

Q7. Which frequency bins does RouteTS select at K=1K=1?

At K=1K=1, RouteTS selects the FFT bin with the largest Fourier magnitude independently for each input window and variable. We use the 24-hour period as a reference for interpreting the selected bins. With both the look-back window and forecast horizon set to 336 steps, we obtain the following results:

  • Solar. Bin 2 is selected in 95.38% of window–variable pairs. The 24-hour frequency falls between bins 2 and 3, which together account for 96.12% of selections. For the same variable, the selected bin changes between consecutive windows in 0.37% of comparisons.
  • Traffic. Bin 14 is selected in 98.83% of window–variable pairs and coincides with the 24-hour frequency at this look-back length. The pooled switch rate between consecutive windows for the same variable is 0.022%.

These results show that the amplitude-based rule produces nearly fixed masks on these datasets at the 336-step setting, especially on Traffic. Table 3 shows that the resulting mixed-domain allocation is beneficial on both Solar and Traffic. On both datasets, this split yields lower mean test MSE across the four forecast horizons than either single-domain configuration, supporting our central claim that routing the dominant component to the frequency branch and its complement to the temporal branch can improve forecasting.

Reviewer_dQM5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a lightweight time-series forecasting framework called RouteTS, which combines time-domain and frequency-domain modeling. Specifically, it selects the Top-K spectral components according to their amplitudes and sends them to a frequency-domain predictor. The remaining complementary spectrum is transformed back into the time domain and processed by a temporal predictor. The authors argue that this selective routing allows the model to switch among time-only, frequency-only, and mixed-domain configurations while remaining lightweight. The method achieves competitive results on multiple datasets.

Critical strengths

S1. The paper is clearly written, well organized, and easy to read.

S2. RouteTS achieves competitive forecasting performance with relatively few parameters and offers an interesting way to combine time- and frequency-domain modeling.

S3. The experiments are comprehensive, covering a variety of benchmark datasets, recent baselines, and several ablation studies and sensitivity analyses.

Critical weaknesses

W1. The authors provide intuitive explanations but do not fully justify why high-amplitude components are better suited to frequency-domain prediction and the remaining components to temporal modeling. More direct evidence would make this design more convincing.

W2. The selection of routing budget K mainly relies on a restricted hyperparameter search. Although the paper extensively analyzes K sensitivity and the validation–test discrepancy on Exchange, a more principled or adaptive selection strategy is still lacking.

W3. Most datasets select single-domain configurations, while mixed-domain routing offers only marginal improvements or even degrades performance. These results provide limited evidence for its practical benefits.

W4. Although the input spectra are complementary, both predictors can generate overlapping output frequencies. The actual outputs may therefore differ from the intended division between periodic patterns and local variations. Further analysis of the branch output spectra would be helpful.

W5. The temporal MLP and complex-valued frequency predictor differ substantially in architecture and parameter count. The observed routing preferences may therefore reflect predictor capacity rather than domain suitability. Capacity-controlled comparisons would help clarify this issue.

W6. Figure 1 is not sufficiently intuitive. The claimed peak timing and amplitude deviations of DLinear are difficult to assess in Figure 1(b). Including the FITS prediction curve would make the comparison clearer.

Questions for authors

1.What evidence supports assigning high-amplitude components to the frequency-domain predictor and the remaining components to the temporal predictor? Please also analyze the output spectra of both branches.

2.Could the authors explore a more principled or adaptive K-selection strategy, especially given the different preferences across forecasting horizons?

3.Would the observed routing preferences and benefits of mixed-domain routing still hold if the two predictors had comparable model capacities?

4 comments
AuthorsOpenReview ↗

Response to Reviewer dQM5 (Part 1)

Thank you for your detailed review and insightful comments. In the revised manuscript, we introduce a rule for selecting KK. It changes the selected configurations for PM2.5, Weather, ECL, and Exchange, and we update Table 1 and the corresponding appendix tables accordingly. After these updates, RouteTS still achieves the lowest average test MSE on seven of the twelve datasets and the lowest average test MAE on five.

Q1 (W1, W4). Limited evidence for branch assignment and spectral specialization.

Two controlled experiments in Section 4.4 support routing high-amplitude components to the frequency predictor and the remaining components to the temporal predictor.

  • Which branch should receive each component? The Swap ablation (Section 4.4) holds the spectral split fixed and reverses the branch assignments. Test MSE increases at every evaluated horizon, by 27.8% on Solar and 8.0% on Traffic on average.
  • Which component should go to the frequency branch? The component-selection ablation (Table 4, Section 4.4) fixes K=1K=1 and varies only the selection criterion. Amplitude-based selection reduces average test MSE on Solar by 6.6% relative to low-frequency selection and 7.5% relative to random selection. The MSE differences are smaller on Traffic.

The branch-output spectrum analysis (Table 8) shows that the two branches emphasize different frequencies on Solar. Across test window–variable pairs, the frequency branch places an average of 66.98% of its non-DC output power at the dominant frequency selected from the input, compared with 19.55% for the temporal branch, whose mean output spectrum peaks at a different bin. This concentration supports the frequency branch’s role in modeling the dominant periodic component. The representative forecast in Appendix B.3 complements the spectral evidence by showing how the branches contribute in the time domain. The temporal branch captures much of the sharp peak structure, while the frequency branch provides a smoother complementary correction.

Q2 (W2). Lack of a principled strategy for selecting KK.

In the revision, we introduce a probe-guided rule for selecting KK (Algorithm 1). The rule is motivated by the different inductive biases of the linear frequency predictor and the nonlinear temporal predictor, and by the possibility that the two are complementary. We first test whether adding nonlinear features improves forecasts on held-out blocks of the training split to shortlist a single-domain endpoint. We then use validation MSE at one reference horizon to choose between that endpoint and the mixed configuration (K=1K=1).

We use a reference horizon of 336 for the longer-horizon benchmarks to give the validation comparison enough time steps to capture recurring patterns, and 48 for ILI and COVID-19, which use shorter horizons. We share the selected KK across all four forecast horizons. As shown in Table 2, the rule matches the test-best KK on 8 of 12 datasets and comes within 1.3% of the best horizon-averaged test MSE among the evaluated budgets on 11 of 12 datasets. on Exchange, where the gap is somewhat larger than on the other datasets, the selected configuration’s mean test MSE is within 6.99% of the lowest among the evaluated budgets (Appendix B.4). Finding the KK with the lowest mean validation MSE under the original procedure required evaluating each of the eight candidates at all four forecast horizons. A key contribution of RouteTS is that Algorithm 1 reduces this search to two candidates at a single reference horizon and still achieves near-best mean test MSE on most datasets.

AuthorsOpenReview ↗

Response to Reviewer dQM5 (Part 2)

Q3 (W5). Unequal capacity of the two predictors.

After approximately matching the parameter counts of the two branches, mixed-domain routing achieves the lowest test MSE on both Solar and Traffic at the 336-step forecast horizon. We reduced the temporal MLP width from 512 to 85 to match the size of the frequency predictor. We reran K=0K=0 and K=1K=1 with this width and used the same frequency-only result in both comparisons because it uses only the frequency predictor.

DatasetMLP widthTime-only (K=0K=0)Mixed (K=1K=1)Frequency-only (K=K=full)
Solar5120.210 / 0.2560.207 / 0.2540.289 / 0.306
Solar850.228 / 0.2670.223 / 0.2660.289 / 0.306
Traffic5120.406 / 0.2770.406 / 0.2740.436 / 0.290
Traffic850.430 / 0.2980.425 / 0.2930.436 / 0.290

Each entry reports test MSE / MAE. Reducing the MLP width increases both errors for the time-only and mixed configurations on both datasets, showing that absolute forecasting accuracy is sensitive to temporal predictor capacity. The mixed configuration continues to improve on both endpoints in MSE in the matched-width comparison. On Traffic, the frequency-only endpoint gives the lowest MAE in that comparison. This ablation controls parameter counts per branch. The mixed configuration uses both predictors and therefore has more total parameters than either endpoint.

W3. Limited practical gains from mixed-domain routing.

The revised Table 3 compares mixed-domain routing with both single-domain endpoints. On all four datasets where the rule selects K=1K=1, the mixed configuration improves mean test MSE over the shortlisted endpoint. On Solar and Traffic, it also outperforms both endpoints, reducing MSE by 25.56% and 4.96%, respectively, relative to the frequency-only configuration.

These results illustrate the practical value of supporting all three configurations in one framework. Algorithm 1 gives each dataset a way to choose between a single-domain endpoint and mixed-domain routing using its own training and validation data. Please see our response to Reviewer vpVw, W4(a), for the full endpoint comparison.

W6. Readability of Figure 1(b).

We added the FITS prediction curve to Figure 1(b) and shaded the intervals between successive peaks where DLinear deviates from the ground truth. In this Traffic example, FITS follows the recurring pattern more closely, especially in the highlighted intervals. This observation supports our motivation for modeling dominant periodic components in the frequency domain.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts, but I still think the proposed routing design is largely heuristic, and the improvements over the stronger single domain model are limited. I will therefore keep my current score.

AuthorsOpenReview ↗

Response to Reviewer dQM5 (Part 3)

Thank you for your follow-up comment. Our routing design is motivated by the hypothesis that dominant spectral components and their complement benefit from predictors with different inductive biases. Keeping the spectral split fixed and reversing the branch assignments increases average test MSE by 27.8% on Solar and 8.0% on Traffic (Section 4.4). The distinct spectral emphases of the branch outputs on Solar provide further support for this assignment (Table 8).

We agree that gains over the stronger single-domain endpoint are modest. RouteTS also provides a unified framework for choosing among time-only, frequency-only, and mixed-domain configurations. Algorithm 1 guides this choice using dataset-specific forecasting evidence and achieves mean test MSE within 1.3% of the best evaluated budget on 11 of 12 datasets. We believe that the experimentally supported component assignment and systematic configuration selection, together with RouteTS’s lightweight architecture, offer practical value alongside the incremental accuracy gains.

We hope this clarifies our contribution and would welcome any further feedback.

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

This paper proposes RouteTS, a lightweight long-term time-series forecasting framework that assigns different spectral components to time- and frequency-domain predictors. For each normalised input window and variable, RouteTS applies the Fourier transform, ranks frequency components by amplitude, and routes the Top-K components to a shared complex-valued linear predictor. RouteTS is evaluated on 12 benchmark datasets and four prediction lengths against ten recent forecasting baselines. The reported results demonstrate competitive forecasting accuracy with relatively low parameter count, training time, and memory consumption. Additional experiments investigate different component selection strategies, reversed branch assignments, alternative predictors, parameter sharing, robustness to noise, and input lengths.

Critical strengths

RouteTS is simple, lightweight, and easy to interpret, with low parameter and memory requirements. The paper evaluates the method on 12 datasets and includes useful analyses of component selection and branch assignment. These strengths make the method practical and relatively easy to reproduce.

Critical weaknesses
  1. The proposed decomposition is closely related to FAN [2], which also performs instance-wise amplitude-based Top-K frequency selection, separates the dominant components from the residual, predicts the two parts through different branches, and combines their forecasts.
  2. The paper does not discuss or compare with Dualformer [1], a closely related method that also combines time- and frequency-domain branches and adaptively assigns their contributions. The paper should clearly explain this distinction and provide a direct comparison to establish its novelty and empirical advantage.
  3. Please conduct two comparative experiments on [1] and [2].
  4. Only three of the 12 datasets use both the time- and frequency-domain branches, while the others select a single branch. Moreover, K is chosen separately for each dataset through validation rather than being learned by the model, and it is unclear how the proposed routing mechanism generalises across datasets.

Reference

[1] J. Bai and Y. Kawahara, “Dualformer: Time-Frequency Dual Domain Learning for Long-term Time Series Forecasting,” in Proceedings of the 29th International Conference on Artificial Intelligence and Statistics, vol. 300, pp. 271–279, 2026.

[2] W. Ye, S. Deng, Q. Zou, and N. Gui, “Frequency Adaptive Normalization for Non-stationary Time Series Forecasting,” in Advances in Neural Information Processing Systems, vol. 37, 2024.

Questions for authors

Regarding the weakness in the content, please consider providing a more detailed discussion of each point. Additionally, this paper mainly shows that different datasets prefer different predictors, rather than proposing a truly adaptive routing method. Its contribution is mostly based on validation-based model selection, with limited methodological novelty.

4 comments
AuthorsOpenReview ↗

Response to Reviewer vpVw (Part 1)

Thank you for your careful review and constructive comments. In the revised manuscript, we introduce a rule for selecting KK. It changes the selected configurations for PM2.5, Weather, ECL, and Exchange, and we update Table 1 and the corresponding appendix tables accordingly. After these updates, RouteTS still achieves the lowest average test MSE on seven of the twelve datasets and the lowest average test MAE on five.

W1–W3. Relationship to FAN and Dualformer and direct comparisons.

RouteTS makes component-to-domain assignment explicit. The top-KK Fourier coefficients go directly to a complex-valued frequency predictor, and only their reconstructed complement goes to the temporal MLP. This gives KK a concrete role: it controls the inputs to the two predictors and allows time-only, frequency-only, or mixed-domain forecasting within one framework.

FAN shares the window-wise amplitude-based split. It reconstructs the dominant components in the time domain and predicts them with an MLP that also sees the full input history. Its backbone forecasts the residual.

Dualformer samples frequency bands across layers and uses harmonic energy to weight the time- and frequency-domain features. The revised Related Work clarifies these design differences.

We evaluated Dualformer and DLinear + FAN using the same data splits, lookback lengths, and forecast horizons as RouteTS. The table reports mean test MSE / MAE across four horizons. RouteTS achieves the lowest mean MSE on 10 of 12 datasets and the lowest mean MAE on 11 of 12. Dualformer leads on both metrics on PM2.5, and DLinear + FAN has the lowest MSE on Weather. Table 11 of the revised manuscript provides the per-horizon results.

DatasetRouteTSDualformerDLinear + FAN
ILI1.640 / 0.8072.278 / 0.9543.124 / 1.225
COVID-190.125 / 0.1350.678 / 0.3511.101 / 0.593
PM2.50.418 / 0.4350.401 / 0.4220.421 / 0.463
Weather0.240 / 0.2760.256 / 0.2950.234 / 0.291
Traffic0.402 / 0.2740.730 / 0.4200.437 / 0.305
Solar0.201 / 0.2490.359 / 0.3930.206 / 0.255
ECL0.164 / 0.2560.221 / 0.3270.171 / 0.270
ETTh10.407 / 0.4190.540 / 0.5100.469 / 0.476
ETTh20.338 / 0.3830.387 / 0.4260.427 / 0.441
ETTm10.352 / 0.3800.494 / 0.4720.363 / 0.392
ETTm20.257 / 0.3150.313 / 0.3590.281 / 0.341
Exchange0.398 / 0.4230.818 / 0.5730.581 / 0.515

The Swap ablation in Section 4.4 directly tests the importance of the branch assignment. Keeping the spectral split fixed and reversing the assignments increases average test MSE by 27.8% on Solar and 8.0% on Traffic. This supports assigning the dominant coefficients to the frequency predictor and their complement to the temporal predictor in the evaluated mixed-domain settings.

AuthorsOpenReview ↗

Response to Reviewer vpVw (Part 2)

W4(a). Why did only three of the twelve datasets select mixed-domain routing?

The benefit of using both branches depends on the dataset. RouteTS supports time-only, frequency-only, and mixed-domain prediction within a single framework and selects a routing configuration for each dataset. Both single-domain endpoints are valid routing outcomes in which all components are assigned to one prediction domain.

The revised rule selects mixed-domain routing on four datasets. The table below reports test MSE averaged across the four forecast horizons, as in Table 3.

DatasetTime-only (K=0K=0)Mixed (K=1K=1)Frequency-only (K=K=full)
Weather0.2370.2400.249
Traffic0.4030.4020.423
Solar0.2040.2010.270
ECL0.1630.1640.170

Solar illustrates the benefit of keeping the mixed-domain option: K=1K=1 reduces MSE by 25.56% relative to the frequency-only model and by 1.47% relative to the stronger time-only model. On Traffic, the mixed configuration also outperforms both endpoints, with a 0.25% gain over the stronger endpoint.

W4(b). Reliance on validation and reliability of KK selection across datasets.

In the revision, we introduce a probe-guided rule for selecting KK and apply it to all datasets (Algorithm 1). The rule is motivated by the different inductive biases of the nonlinear temporal MLP and the linear frequency predictor, and by the possibility that they offer complementary forecasts. We compare a linear forecasting probe with one augmented with nonlinear features on three held-out blocks within the training split. We shortlist the frequency-only endpoint if the linear probe yields lower MSE on at least two blocks, and the time-only endpoint otherwise. We then use validation MSE at a single reference horizon to choose between the shortlisted endpoint and the mixed configuration (K=1K=1), and use the selected KK across all four forecast horizons.

As shown in the revised Table 2, the rule matches the test-best KK on 8 of 12 datasets. On 11 of 12 datasets, the selected configuration's mean test MSE is within 1.3% of the lowest mean test MSE across the evaluated values of KK. These results show that the rule selects near-best configurations on most benchmarks while reducing the validation search from eight candidates across four horizons to two candidates at one horizon.

Q1. Limited methodological novelty.

RouteTS makes how to divide the spectrum and whether to use both prediction domains part of the same design decision. It assigns complementary spectral components to predictors in different domains, with pure time-domain and pure frequency-domain forecasting as endpoints of the same allocation. This makes it possible to evaluate the benefit of mixed-domain routing against both single-domain alternatives within one framework.

Algorithm 1 adds a dataset-specific basis for this choice. It uses the forecasting benefit of nonlinear features to guide which single-domain endpoint to consider alongside mixed-domain routing. This reduces validation to two candidates at one horizon. The resulting configurations come within 1.3% of the best mean test MSE among the evaluated values of KK on 11 of 12 benchmarks (Table 2). RouteTS offers a starting point for future work on how the structure of time-series data can guide the combination of predictors with different inductive biases.

Thank you for the detailed response and additional experiments. I appreciate the your great efforts, but I still feel the explanations of the critical difference/novelty between the proposed method and the two related works, especially FAN, are not clear to me. I prefer to keep my score.

AuthorsOpenReview ↗

Response to Reviewer vpVw (Part 3)

Thank you for your follow-up comment. The key differences concern how spectral inputs are organized and how the resulting components are predicted.

Relationship to Dualformer. At each layer, Dualformer processes the same sampled frequency components through its time and frequency branches, then combines their representations using harmonic-energy-based weighting. RouteTS instead assigns disjoint, complementary input components to different prediction domains: the selected Fourier coefficients enter the frequency predictor, while their reconstructed complement enters the temporal MLP.

Relationship to FAN. Although FAN and RouteTS share amplitude-based Top-KK selection, its role differs in the two methods:

  • Input and decomposition rationale. FAN is a model-agnostic normalization method. It applies spectral selection without preceding RevIN and treats the selected dominant components as non-stationary patterns to be removed and subsequently restored. RouteTS applies selection after RevIN, which removes window-level mean and scale differences. Its component allocation is motivated by modeling recurring patterns and localized variations using predictors with different inductive biases.

  • Prediction design. FAN reconstructs the selected components in the time domain and forecasts them with an MLP that also receives the full input history. RouteTS directly predicts the selected complex Fourier coefficients, while a temporal MLP models only their reconstructed complement. On Traffic, FAN uses multiple selected components, whereas our revised RouteTS configuration uses K=1K=1. These configurations illustrate the different uses of spectral selection in the two designs.

Thus, amplitude-based Top-KK selection is a decomposition tool, whose role depends on the surrounding normalization and prediction design. RouteTS further provides an explicit, forecasting-guided rule for selecting KK, offering systematic configuration guidance and reducing reliance on empirical budget tuning.

We hope this clarifies the distinctions and the scope of our contribution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all the reviewers for their insightful reviews and valuable comments, which have helped us improve the paper.

The reviewers described RouteTS as “simple, lightweight, and easy to interpret” and noted that it achieves “competitive forecasting performance with relatively few parameters.” They also highlighted the combination of time- and frequency-domain modeling, the clear presentation, and the comprehensive evaluation across 12 datasets. The detailed reporting of routing-budget sensitivity and the controlled analyses of component selection and branch assignment were noted as strengths that support reproducibility.

We made every effort to address all the concerns by providing sufficient evidence and requested results. Here is the summary of the major revisions:

  1. Routing-budget selection. We introduce a dataset-informed rule for selecting KK that reduces the validation search to two candidates at one forecast horizon (Algorithm 1). It selects the test-best budget on 8 of 12 datasets and achieves mean test MSE within 1.3% of the best evaluated budget on 11 of 12 (Table 2).

  2. Effectiveness of routing. RouteTS supports time-only, frequency-only, and mixed-domain forecasting as three routing outcomes within one framework, accommodating different dataset characteristics. The revised Table 3 shows that mixed-domain routing reduces mean test MSE relative to both single-domain endpoints on Solar and Traffic.

  3. Branch behavior. We added the FITS forecast to Figure 1(b) to illustrate the motivation for frequency-domain modeling. The output-spectrum analysis in Table 8 shows that, on Solar, the frequency branch concentrates its power at the selected bin and the temporal branch peaks at a different bin, consistent with their intended roles.

  4. Related work and direct comparisons. We clarify how RouteTS relates to FAN, Koopa, and Dualformer and add direct comparisons in Table 11. RouteTS achieves the lowest mean test MSE and MAE on most datasets in this comparison.

We have uploaded the revised manuscript and supplementary material. The point-by-point responses describe these updates in the context of each reviewer’s comments.

We would be happy to clarify any of the revisions or answer further questions.

We look forward to your feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.