acceptodds
Under review as a conference paper at ICLR 2027

FIRE: Fourier-Informed Regime Adaptive Estimation for Non-Stationary Time Series Forecasting

Abstract

Non-stationarity remains a critical challenge in multivariate time series forecasting because data distributions and temporal dependencies continuously evolve over time. Recent frequency-domain models have demonstrated the value of spectral representations, yet most of them jointly model amplitude and phase within unified complex embeddings, overlooking their distinct physical roles and evolutionary behaviors. From a signal processing perspective, amplitude characterizes the strength of frequency components, whereas phase encodes their temporal alignment. These distinct physical roles and transformation properties motivate dedicated modeling of amplitude and phase under non-stationarity. We propose FIRE (Fourier Informed Regime Adaptive Estimation), a forecasting framework that explicitly decouples amplitude and phase into dedicated learning branches. To adapt these branches to evolving dynamics, FIRE introduces a signal-informed modulation mechanism driven by two historical window statistics: the Amplitude Non-stationarity Index (ANI) and the Phase Irregularity Index (PII). Across 12 diverse real-world benchmarks, FIRE achieves state-of-the-art forecasting performance against competitive baselines, securing the lowest average MSE and MAE on 9 and 10 datasets, respectively. Controlled ablations demonstrate the benefits of amplitude-phase decoupling and ANI/PII guided adaptive modulation, while channel-level analyses show larger forecasting gains for the evaluated channels with higher ANI and PII.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Frequency-domain, Signal-processing Informed Modulation

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

The paper proposes FIRE, a patch-based frequency-domain forecaster. Each look-back window is RevIN-normalized, patched and transformed by FFT; amplitude and phase are embedded and processed in two disjoint Branchformer-style branches with causal attention over patches. Two per-channel scalar statistics of the look-back window, the Amplitude Non-stationarity Index (ANI, the coefficient of variation of patch spectral energy) and the Phase Irregularity Index (PII, the circular dispersion of inter-patch phase increments), are mapped by small MLPs to multiplicative gates on the MLP and attention outputs of each branch. After cross-variable attention, the branches are recombined into a complex representation, a projection head predicts the future spectrum, and an inverse FFT gives the forecast. Training uses a Huber loss plus the FreDF frequency loss. Experiments cover seven standard long-term benchmarks (look-back 96) against nine baselines, five additional datasets against two baselines, a two-variant ablation, a loss ablation, gate visualizations, a channel-level analysis on eight selected channels, and efficiency numbers.

Critical strengths
  1. A clearly stated and physically motivated design question. Asking whether amplitude and phase should be given separate representations, and whether window-level spectral statistics can condition the network, is a reasonable hypothesis, and the architecture is a clean instantiation of it (cos/sin phase coordinates, masking of ill-conditioned phases, disjoint branches). This matters to my assessment because the idea is testable; my concerns are about whether the paper has tested it.

  2. Breadth of the experimental surface and candour in places. Twelve datasets, a loss ablation, four sensitivity studies and an efficiency table are provided, and the paper acknowledges that WPMixer is cheaper and that gains vary by dataset. The appendix tables are what allowed me to check the main claims, which I credit.

Critical weaknesses
  1. The evidence does not attribute the reported gains to the two claimed contributions (decoupling and ANI/PII modulation).

    • Loss confound. FIRE is trained with Huber + FreDF loss; most baselines are reported under their own objectives. Table A5 shows that FIRE with plain MSE obtains average MSE 0.442 / 0.383 / 0.247 on ETTh1 / ETTh2 / Weather, which is worse than WPMixer (0.435 / 0.371 / 0.244), TimeMixer (ETTh2 0.364, Weather 0.240) and DistDF in Table 1. With Huber only it is 0.441 / 0.373 / 0.245. So on the three datasets where this was checked, the architecture alone does not beat the baselines; the margin comes from the borrowed training objective. No baseline is retrained with the same loss.
    • Ablation. Removing modulation changes average MSE by 0.001–0.005 on six of seven datasets (Table 4), with no seeds or standard deviations. On Electricity, the dataset the paper highlights as its largest win (4.7% over APDNet), the full model, "w/o Mod." and "Joint Branch" score 0.162 / 0.163 / 0.163: neither claimed contribution matters there, so the gain must come from something else (cross-variable attention, loss, or tuning). The only large Joint-Branch gap (ETTh1) is driven by two horizons (0.547, 0.551) that look like an unstable run rather than a systematic effect. There is no ANI-only / PII-only ablation and no control that feeds the gates a constant or a random scalar.
    • The gates barely act. Figure 3 shows learned gate values within roughly 1.00–1.05, applied to sub-layer outputs that are then normalized. A ≤5% rescaling is hard to reconcile with "regime-adaptive" modeling.
    • Tuning variance exceeds the margins. The sensitivity tables, which share the main protocol (T = 96), give ETTh1 averages of 0.428–0.441, Weather 0.244–0.248 and Traffic 0.440–0.467, never reaching the main-table values (0.424, 0.241, 0.433). The headline numbers therefore come from a per-dataset/per-horizon search over patch length, batch size and learning rate whose spread (0.005–0.03) is larger than most gaps to the second-best baseline (often 0.001–0.006). Only Fredformer and WPMixer were re-run; it is not stated that baselines received a comparable search.
  2. The central analytical claim, that FIRE's advantage "scales progressively with the severity of non-stationarity", rests on eight hand-selected channels, and the indices themselves are weakly justified.

    • Table 5 reports 4 of 21 Weather channels and 4 of 7 ETTm1 channels, described as "selected". Weather channels 13 and 12 have essentially identical indices (ANI 0.3446 vs 0.3456, PII 1.2977 vs 1.2996) yet gains of 8.9% and 16.8%, so the indices do not explain the gain even within the selected set. No correlation over all channels and datasets is given, and the comparison is against a single baseline.
    • PII is nearly saturated: seven of the eight reported values lie in 1.14–1.31 rad, so it hardly discriminates between channels. It weights all above-threshold bins equally, so it is likely dominated by low-energy noisy bins.
    • ANI is computed after RevIN and, by Parseval, the patch spectral norm equals the time-domain patch energy; ANI is therefore the coefficient of variation of patch RMS, a time-domain statistic, not a spectral diagnostic.
    • The motivating Figure 1(a,b) is close to tautological (gain perturbations change amplitude, circular shifts change phase).
    • The closest prior work on amplitude/phase modeling that the paper itself cites (APRNet, TimeAPN) is not compared against, so novelty relative to it is not established empirically.
Questions for authors

Clarifications that could change my assessment:

  1. Please retrain at least WPMixer, Fredformer and iTransformer with the same Huber + FreDF objective and the same hyperparameter search budget, or report FIRE with each baseline's native loss on all seven datasets. Does FIRE still lead?
  2. Report mean ± std over ≥3 seeds for Table 1 and Table 4. Which differences survive?
  3. On Electricity, all three ablation variants tie. What produces the gain over the baselines there?
  4. Provide the ablations described in A.5.2 (amplitude-only, phase-only, ANI-only, PII-only) and a control where the gate input is replaced by a constant or shuffled across samples.
  5. Replace Table 5 with the gain-versus-ANI/PII relationship over all channels of all seven datasets (scatter plus rank correlation), against more than one baseline.
  6. What are the loss weights, and the selected patch length / batch size / learning rate for each dataset and horizon? Was the search validated on the validation set for every reported cell?
  7. How does FIRE compare with APRNet and TimeAPN, and with baselines at longer look-backs (Table A9 shows FIRE benefits from T up to 512)?

Suggestion for future work (does not affect my rating): per-frequency or per-patch conditioning rather than one scalar per channel.

3 comments

We appreciate the detailed examination of attribution and evaluation settings. We clarify these issues with additional comparisons and ablations.

Q1. Training objective.

  • We retrained WPMixer, TimeMixer, and DistDF, highlighted in the review, with FIRE’s Huber + FFT objective and their public tuning ranges.

Table 1. Shared-objective results (average MSE/MAE).

DatasetFIREWPMixerTimeMixerDistDF
ETTh1.424/.427.425/.427.432/.431.438/.431
ETTh2.362/.392.365/.394.378/.400.366/.392
ETTm1.371/.387.378/.392.379/.391.385/.394
ETTm2.276/.319.271/.316.271/.318.277/.321
Weather.241/.265.243/.269.243/.272.250/.273
Electricity.162/.255.176/.265.182/.273.175/.265
Traffic.433/.266.485/.302.499/.313.589/.367
  • FIRE leads on both average metrics on six of seven datasets, with 18/19 MSE/MAE best counts across 28 settings. This supports an advantage beyond objective choice alone.

Q2–Q3. Variability and component contributions.

  • We added 10-seed ablations.

Table 2. Average MSE over 10 seeds.

DatasetFIREw/o Mod.Joint
ETTh1.4292±.0032.4313±.0018.4601±.0153
ETTh2.3674±.0044.3717±.0068.3872±.0065
ETTm1.3726±.0009.3727±.0005.3745±.0007
ETTm2.2786±.0013.2794±.0007.2820±.0005
Weather.2442±.0020.2449±.0007.2464±.0006
Electricity.1630±.0009.1634±.0008.1626±.0006
Traffic.4364±.0032.4515±.0021.4591±.0013
  • Without modulation, separate branches achieve lower mean MSE than joint modeling on six of seven datasets, while the difference on Electricity is small. This supports the proposed inductive bias, with benefits varying across datasets.

Table 3. Additional modulation ablations (MSE, 10 seeds).

DatasetFIREw/o Mod.Reduction
PV.4552±.0202.4635±.00881.79%
L2C.8424±.0105.8834±.00764.64%
  • Results average four horizons, or three for PV/L2C. These new seed statistics are separate from the original point estimates. We will revise “substantial complementary gains” to reflect dataset-dependent benefits.

Q4. Individual modulation and gates.

  • Both branches remain intact. ANI-only removes PII conditioning, and vice versa.

Table 4. Individual modulation (average MSE, 10 seeds).

DatasetFIREANI-onlyPII-onlyw/o Mod.
ETTh2.3674±.0044.3706±.0022.3704±.0030.3717±.0068
Traffic.4364±.0032.4493±.0015.4480±.0014.4515±.0021
  • Joint conditioning yields the lowest mean MSE here. It combines learned mappings with signal-processing-informed cues about evolving amplitude and phase patterns. The gains are input-dependent and feature-wise, applied before residual addition. Layer normalization does not generally cancel them; Figure 3’s averages near 1 do not imply identical gains across features or inputs.

Q5. Analysis across channels.

  • We cover all channels of four datasets, grouping window–channel pairs by ANI/PII quartiles. Each cell reports relative MSE reduction from pooled errors, averaged equally across the four horizons.

Table 5(a). ETTm1 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1-.35-.34-.29.27
Q2-.52.28.51.67
Q3.15.54.43.82
Q4.29.67.53.96

Table 5(b). ETTm2 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.30.43.49.61
Q2.83.501.00.46
Q3.45.731.69.77
Q41.211.781.381.31

Table 5(c). Weather gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.41.51.76.93
Q2-.17.63.79.87
Q3.86.82.86.95
Q4.81.951.021.13

Table 5(d). Traffic gains (%).

ANI / PIIQ1Q2Q3Q4
Q11.902.933.903.82
Q23.023.893.955.12
Q33.182.984.264.97
Q43.523.564.995.52
  • Higher-index regions generally benefit more, most clearly on ETTm1/Traffic. We will revise “substantial complementary gains” to reflect these general trends.

  • ANI relates to patch energy variation, which whole-window RevIN need not eliminate. It does not uniquely identify spectral-shape changes. Similar channel-average PII values alone do not establish instance-level saturation. Figure 1 illustrates the Fourier motivation, not independent evidence of forecasting gains.

Q6. Tuning protocol.

  • Both FIRE and reproduced baselines were tuned using model-specific search spaces. We followed baselines’ official ranges and reference configurations, selecting on validation data and reserving test data for evaluation (Appendix A.4).

  • Appendix A.5.4 uses one-factor sweeps with other settings fixed, distinct from configuration selection and seed variability. We will clarify reference settings and report loss weights and selected patch length/batch size/learning rate for each dataset/horizon.

Q7. APRNet, TimeAPN, and longer contexts.

Table 6. Comparison at look-back 96 (average MSE/MAE).

DatasetFIREAPRNetF+TP+TS+T
ETTh1.424/.427.443/.432.441/.448.438/.430.447/.437
ETTh2.362/.392.375/.397.384/.411.367/.395.374/.402
ETTm1.371/.387.385/.395.373/.406.377/.395.391/.404
ETTm2.276/.319.281/.324.277/.328.279/.324.278/.326
Weather.241/.265.256/.280.261/.317.245/.287.237/.287
Electricity.162/.255.171/.263.170/.274.161/.259.164/.270
Best count4/60/00/01/01/0
  • F/P/S denote FEDformer/PatchTST/S_Mamba. T denotes TimeAPN. All entries use look-back 96. Results average four horizons, best counts cover 24 dataset–horizon settings. FIRE achieves the best counts in average MSE(4)/MAE(6). For all cases, Fire also wins 14/24 for MSE and 22/24 for MAE, we'll add full results in the revised PDF.

Table 7. Traffic comparison at look-back 96 (average MSE/MAE).

DatasetFIREAPRNetF+T
Traffic.433/.266.464/.293.479/.304
Best count4/40/00/0
  • On Traffic, FIRE achieves the lowest MSE and MAE at all four horizons among the eligible models. Against APRNet, FIRE achieves lower average MSE and MAE on all seven datasets.

  • Appendix Table A.9 evaluates contexts up to 512: Weather’s average MSE decreases from .244 (96) to .227 (512). This supports using longer histories, without establishing superiority under matched longer-context comparisons.

The supplementary PDF will document full results and settings, with claims revised to match the evidence.

Thank the authors for the extensive rebuttal. While these additions substantially improve the evaluation, the new results still reinforce my core concern regarding component attribution: outside of Traffic, ANI/PII modulation yields virtually identical MSE to w/o Mod. across six of the seven datasets. Moreover, without the requested constant- or shuffled-scalar gating control, it remains unclear whether the Traffic gain stems from genuine spectral conditioning or simply added MLP capacity. Therefore, I maintain my current score.

Reviewer_jD9sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that amplitude and phase have distinct physical roles and transformation properties, therefore motivating dedicated modeling under non-stationarity. Based on this idea, the authors propose FIRE to explicitly decouple amplitude and phase into separate learning branches. Two statistics derived from historical observations guide branch-specific modulation. The resulting representations are combined to produce forecasts. Experiments across twelve datasets report competitive forecasting performance.

Critical strengths

The paper addresses an important forecasting problem and connects its frequency-domain representation to a concrete architectural design. The motivation of separating amplitude and phase modeling branches due to distinct physical roles/properties is sound. To capture these different charactertics, trigonomertic phase representation is used to respect periodicity. Extensive experiments are conducted to validate the proposed method and the benefit of separating these spectral components.

Critical weaknesses

I have the following concerns:

  • unclear how much novelty is presented over earlier methods. APRNet [1] already uses amplitude and phase information. The paper discusses both methods but does not compare FIRE against it.
  • The separate branch comparison seems unfair as the two branches version could have more parameters than the single branch (the authors did not report the number of parameters), two branches may perform better simply because the model is larger.
  • The ANI/PII test compares scaling that changes for each input with no scaling. The authors should also test a scale learned during training and kept the same for every input. This would show whether changing the scale for each input provides a benefit.
  • The authors claim the gains from ANI/PII as “substantial complementary gains.” However, some differences are small. In Table 4, adding ANI/PII reduces average MSE from 0.372 to 0.371 on ETTm1 and from 0.163 to 0.162 on Electricity.
  • no standard deviations are reported. How much these results change across runs? It is hard to judge whether these small gains are reliable without std.
  • The claim about larger gains under stronger non-stationarity also relies on eight selected channels from Weather and ETTm1. Tests across all channels and controlled changes in amplitude and phase would provide stronger support.

References:
[1] Amplitude–Phase Reconstruction for Non-Stationary Time-Series Forecasting, TKDE 2026

Questions for authors

please refer to weaknesses

1 comment

We appreciate the reviewer’s questions about component contributions, model capacity, and statistical reliability. We address these points with additional experiments and design clarifications.

W1. Relation to APRNet.

  • APRNet models amplitude–phase interactions across temporal and channel dimensions. FIRE combines distinct amplitude/phase embeddings and branch parameters with historical ANI/PII conditioning. Its contribution is this structured combination, rather than using amplitude and phase alone.

  • We reproduced APRNet with a 96-step look-back. FIRE leads on both average metrics across all seven datasets, winning 28/28 MSE and 27/28 MAE comparisons.

Table 1. APRNet comparison (average MSE/MAE).

DatasetFIREAPRNet
ETTh1.424/.427.443/.432
ETTh2.362/.392.375/.397
ETTm1.371/.387.385/.395
ETTm2.276/.319.281/.324
Weather.241/.265.256/.280
Electricity.162/.255.171/.263
Traffic.433/.266.464/.293

W2. Separate branches and capacity.

  • Amplitude and phase have different representations: spectral magnitude versus periodic sine/cosine coordinates. Separate parameters provide an inductive bias for learning their respective transformations.

  • Parameter count alone does not determine accuracy. Table 6 reports 1.231M parameters for FIRE versus approximately 48.79M for Fredformer. FIRE achieves better reported forecasting results with 2.52% of Fredformer’s parameters.

  • Within FIRE, separate branches without modulation yield lower mean MSE than joint modeling on six of seven datasets over 10 seeds. This supports the architecture, although neither comparison isolates representation effects from capacity effects.

W3. Why input-dependent modulation?

  • Other mechanisms, including learned constant gains, may also help. FIRE combines data learning with a signal-processing-informed inductive bias: amplitude variation and phase irregularity provide distinct cues about evolving temporal patterns and condition the corresponding branch updates. The mappings are learned, while their conditioning signals have explicit interpretations.

  • A learned input-independent gain applies the same scaling across windows. FIRE instead derives gains from each window’s ANI/PII statistics, allowing branch updates to respond to the observed signal characteristics.

W4–W5. Gain magnitude and run variability.

  • We repeated the ablations with 10 random seeds and added PV/L2C. “w/o Mod.” removes ANI/PII conditioning while retaining both branches.

Table 2. Modulation ablation (MSE).

DatasetFIREw/o Mod.
ETTh1.4292±.0032.4313±.0018
ETTh2.3674±.0044.3717±.0068
ETTm1.3726±.0009.3727±.0005
ETTm2.2786±.0013.2794±.0007
Weather.2442±.0020.2449±.0007
Electricity.1630±.0009.1634±.0008
Traffic.4364±.0032.4515±.0021
PV.4552±.0202.4635±.0088
L2C.8424±.0105.8834±.0076
  • Results average four horizons, or three for PV/L2C. Mean MSE is lower on all nine datasets, with reductions ranging from negligible on ETTm1 to 3.34% on Traffic, 1.79% on PV, and 4.64% on L2C. We do not claim significance for every difference and will revise “substantial complementary gains” to reflect dataset-dependent benefits.

W6. Analysis across all channels.

  • We cover all channels of ETTm1, ETTm2, Weather, and Traffic. We group window–channel pairs by ANI/PII quartiles. Each cell reports the relative MSE reduction computed from pooled errors, averaged equally across the four horizons.

Table 3(a). ETTm1 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1-.35-.34-.29.27
Q2-.52.28.51.67
Q3.15.54.43.82
Q4.29.67.53.96

Table 3(b). ETTm2 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.30.43.49.61
Q2.83.501.00.46
Q3.45.731.69.77
Q41.211.781.381.31

Table 3(c). Weather gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.41.51.76.93
Q2-.17.63.79.87
Q3.86.82.86.95
Q4.81.951.021.13

Table 3(d). Traffic gains (%).

ANI / PIIQ1Q2Q3Q4
Q11.902.933.903.82
Q23.023.893.955.12
Q33.182.984.264.97
Q43.523.564.995.52
  • Higher-index regions generally benefit more, most clearly on ETTm1/Traffic. We will revise “substantial complementary gains” to reflect these general trends.

We will incorporate these results into the supplementary PDF and revise the discussion to distinguish the design motivation from the conclusions supported by the experiments.

Reviewer_2EJAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FIRE, a supervised forecasting model for multivariate time series. The authors argue that existing frequency-domain forecasters put amplitude and phase in a single complex embedding, which ignores their different roles and their different behaviour under non-stationarity. FIRE instead handles amplitude and phase in two separate branches, and scales each branch by a gain computed from a window-level statistic: the Amplitude Non-stationarity Index (ANI) for amplitude and the Phase Irregularity Index (PII) for phase. The claimed contributions are the decoupled architecture, the two indices and the modulation they drive, and an evaluation on twelve datasets where FIRE gets the lowest average MSE and MAE on most of them, with bigger gains on more non-stationary channels.

Critical strengths
  1. The method is clearly written and fully specified, and ANI and PII are cheap to compute and easy to interpret.
  2. The evaluation covers twelve datasets from several domains, with extensive sensitivity analyses in the appendix.
Critical weaknesses
  1. The modulation seems to do very little. In Table A4, removing it changes MSE by at most 0.003 on 19 of 28 datasets.
  2. ANI and PII are never ablated separately, so their individual contributions are unknown.
  3. The motivation of this paper is unclear. Published zero-shot results already beat FIRE’s supervised results on two standard benchmarks. For example, on ETTh1, Moirai-Small (0.400/0.424) outperforms FIRE (0.424/0.427) on both metrics. Moirai has no explicit amplitude-phase decoupling, which suggests a model with enough data may learn this behaviour on its own.
  4. The components are only tested inside FIRE’s own Branchformer-style backbone. It would help to see whether ANI/PII modulation and amplitude-phase decoupling also help other architectures, such as a Transformer (PatchTST), a CNN (TimesNet or ModernTCN), an MLP (DLinear or TimeMixer), or a residual network. That would show whether the idea is generally useful or specific to this design.
Questions for authors
  1. The connection to prior work is unclear. APRNet also splits the spectrum into amplitude and phase and learns separate adjustments for each, and TimeAPN also uses amplitude and phase to handle non-stationarity. Could the authors explain how FIRE differs in design and motivation, and why these methods are not in the comparison?
1 comment

Thank you for the constructive feedback. We respond to each comment in detail below.

W1. Modulation gains.

  • We repeated the ablations with 10 random seeds and added PV/L2C.

  • Table 1. Modulation ablation (MSE).

DatasetFIREw/o Mod.
ETTh1.4292±.0032.4313±.0018
ETTh2.3674±.0044.3717±.0068
ETTm1.3726±.0009.3727±.0005
ETTm2.2786±.0013.2794±.0007
Weather.2442±.0020.2449±.0007
Electricity.1630±.0009.1634±.0008
Traffic.4364±.0032.4515±.0021
PV.4552±.0202.4635±.0088
L2C.8424±.0105.8834±.0076
  • Results average four horizons, or three for PV/L2C. Mean MSE reductions vary in different datasets which are negligible on ETTm1, but 3.34% on Traffic, 1.79% on PV, and 4.64% on L2C.

  • We also group window–channel pairs by ANI/PII quartiles. Each cell reports the relative MSE reduction computed from pooled errors, averaged equally across the four horizons.

Table 2(a). ETTm1 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1-.35-.34-.29.27
Q2-.52.28.51.67
Q3.15.54.43.82
Q4.29.67.53.96

Table 2(b). ETTm2 gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.30.43.49.61
Q2.83.501.00.46
Q3.45.731.69.77
Q41.211.781.381.31

Table 2(c). Weather gains (%).

ANI / PIIQ1Q2Q3Q4
Q1.41.51.76.93
Q2-.17.63.79.87
Q3.86.82.86.95
Q4.81.951.021.13

Table 2(d). Traffic gains (%).

ANI / PIIQ1Q2Q3Q4
Q11.902.933.903.82
Q23.023.893.955.12
Q33.182.984.264.97
Q43.523.564.995.52
  • Higher-index regions generally benefit more, most clearly on ETTm1/Traffic. We will revise “substantial complementary gains” to reflect these general trends.

W2. Individual ANI/PII contributions.

We added 10-seed ablations retaining both branches. ANI-only removes PII conditioning, and vice versa.

Table 3. Individual modulation ablations (MSE).

DatasetFIREANI-onlyPII-onlyw/o Mod.
ETTh2.3674±.0044.3706±.0022.3704±.0030.3717±.0068
Traffic.4364±.0032.4493±.0015.4480±.0014.4515±.0021
  • Either signal lowers mean MSE versus removing both. Joint conditioning gives the lowest mean MSE/MAE on both datasets.

W3. Moirai and implicit learning.

  • We agree that sufficient data may enable implicit learning of amplitude–phase relationships. FIRE pursues a complementary approach: an explicit inductive bias motivated by their distinct roles in signal processing. This does not imply that decoupling is necessary for every successful model. Moirai uses external pretraining, whereas FIRE is trained per dataset, so their comparison cannot isolate the effect of decoupling. Within FIRE, separate branches without modulation improve mean MSE over joint modeling on six of seven datasets across 10 seeds.

W4. Other architectures.

  • FIRE integrates separate representations, branch processing, and spectral reconstruction. Transferring these components requires architectural adaptation because FIRE is not a plug-in method. Our ablations establish their effects within FIRE. We also compare FIRE with plug-in method TimeAPN.

Q1. APRNet and TimeAPN.

  • We added APRNet and four TimeAPN-enhanced backbones. APRNet models amplitude–phase relationships. TimeAPN provides adaptive normalization/denormalization around a backbone.

Table 4. Comparison at look-back 96 (average MSE/MAE).

DatasetFIREAPRNetF+TP+TS+T
ETTh1.424/.427.443/.432.441/.448.438/.430.447/.437
ETTh2.362/.392.375/.397.384/.411.367/.395.374/.402
ETTm1.371/.387.385/.395.373/.406.377/.395.391/.404
ETTm2.276/.319.281/.324.277/.328.279/.324.278/.326
Weather.241/.265.256/.280.261/.317.245/.287.237/.287
Electricity.162/.255.171/.263.170/.274.161/.259.164/.270
Best count4/60/00/01/01/0
  • F/P/S denote FEDformer/PatchTST/S_Mamba. T denotes TimeAPN. All entries use look-back 96. Results average four horizons, best counts cover 24 dataset–horizon settings. FIRE achieves the best counts in average MSE(4)/MAE(6). For all cases, Fire also wins 14/24 for MSE and 22/24 for MAE, we'll add full results in the revised PDF.

Table 5. Traffic comparison at look-back 96 (average MSE/MAE).

DatasetFIREAPRNetF+T
Traffic.433/.266.464/.293.479/.304
Best count4/40/00/0
  • On Traffic, FIRE achieves the lowest MSE and MAE at all four horizons among the eligible models. Against APRNet, FIRE achieves lower average MSE and MAE on all seven datasets.

Full results and settings will be included in the supplementary PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.