- 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
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 and . Table 6 lists the final settings across datasets, with batch sizes ranging from 8 to 64 and learning rates from to . 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 (, , , and ) 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.
| Dataset | RouteTS (MSE / MAE) | DLinear (MSE / MAE) | FITS (MSE / MAE) |
|---|---|---|---|
| ILI | 1.557 / 0.796 | 2.456 / 1.050 | 2.292 / 0.993 |
| COVID-19 | 0.146 / 0.158 | 0.146 / 0.250 | 0.451 / 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- Fourier coefficients directly to a complex-valued frequency predictor and reconstructs their complement for the temporal MLP. Varying 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.
| Dataset | RouteTS | DLinear + FAN | Koopa |
|---|---|---|---|
| ILI | 1.640 / 0.807 | 3.124 / 1.225 | 2.085 / 0.909 |
| COVID-19 | 0.125 / 0.135 | 1.101 / 0.593 | 0.871 / 0.368 |
| PM2.5 | 0.418 / 0.435 | 0.421 / 0.463 | 0.524 / 0.472 |
| Weather | 0.240 / 0.276 | 0.234 / 0.291 | 0.249 / 0.278 |
| Traffic | 0.402 / 0.274 | 0.437 / 0.305 | 0.636 / 0.397 |
| Solar | 0.201 / 0.249 | 0.206 / 0.255 | 0.299 / 0.307 |
| ECL | 0.164 / 0.256 | 0.171 / 0.270 | 0.622 / 0.393 |
| ETTh1 | 0.407 / 0.419 | 0.469 / 0.476 | 0.459 / 0.456 |
| ETTh2 | 0.338 / 0.383 | 0.427 / 0.441 | 0.377 / 0.412 |
| ETTm1 | 0.352 / 0.380 | 0.363 / 0.392 | 0.374 / 0.397 |
| ETTm2 | 0.257 / 0.315 | 0.281 / 0.341 | 0.316 / 0.345 |
| Exchange | 0.398 / 0.423 | 0.581 / 0.515 | 0.458 / 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.
Response to Reviewer okZa (Part 3)
Q6. Difference from FITS at .
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 into a sequence of length and takes the final steps as the forecast. This formulation allows FITS to supervise both the reconstructed look-back window and the future horizon. At , RouteTS maps the full input spectrum directly to the spectrum of the -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 , 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 ?
At , 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.
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 K. 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=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=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 K and covers multiple daily cycles on both datasets.
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 K 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 K and which budgets perform close to the test optimum.
Algorithm 1 selects one K per dataset. Training-set probes shortlist a single-domain endpoint, which we compare with K=1 using validation MSE at one reference horizon. With just two candidates evaluated at one validation horizon, the rule matches the test-best K 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).