acceptodds
Under review as a conference paper at ICLR 2027

PhiMixer: Forecast-Origin Phase Conditioning for Long-Term Time-Series Forecasting

Abstract

Multivariate time-series forecasting involves temporal and cross-channel dependencies whose dynamics can vary across recurring cycles. We investigate whether forecast-origin phase can serve as a conditioning signal for adapting forecasting transformations to the underlying rhythmic state. We introduce PhiMixer, a lightweight phase-conditioned architecture that separates learnable multi-period cyclic components from the input and uses forecast-origin phase to modulate temporal and cross-channel processing. Its rhythm-aware relative attention (RRF) captures periodic temporal structure, while the Phase-Informed Low-Rank Channel Mixer (PILCM) generates phase-dependent low-rank channel transformations without dense O(C²) interactions. PhiMixer achieves the best performance on 6/8 datasets on MSE/MAE, respectively, among the compared methods. Ablation studies show that active phase conditioning provides consistent gains, particularly on periodic traffic data, with conditioning consistently outperforming phase-as-feature concatenation. With approximately 126K parameters and linear channel scaling, PhiMixer offers an efficient approach to phase-aware long-term multivariate forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, phase conditioning, low-rank channel mixing, long-term forecasting, efficient architectures, multivariate 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_cXFPOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhiMixer uses forecast-origin phase to condition temporal and cross-channel processing for multivariate time-series forecasting. It combines cyclic decomposition, FiLM modulation, rhythm-aware relative attention, STAR-style aggregation, and a Phase-Informed Low-Rank Channel Mixer, followed by a dual-path prediction head. Experiments cover twelve datasets, with component, phase, loss, sensitivity, and efficiency analyses.

Critical strengths
  1. Forecast-origin phase provides information about recurring temporal patterns that may not be clear from the observed window alone. Using this information to condition forecasting operations is relevant to periodic data such as electricity consumption and traffic.
  2. The model uses shared processing and global aggregation to keep inference cost linear in the number of channels. Its parameter count is relatively small under the standard ETTh1 configuration.
  3. The evaluation covers twelve datasets and multiple prediction horizons. The component, phase, and loss ablations provide evidence that several parts of the model contribute to its performance.
Critical weaknesses
  1. The released code does not implement the experiment described in the paper. The training script uses plain MSE instead of the stated composite objective, and the data loaders return only inputs and targets while the training loop requires an additional forecast-origin index. The provided entry point therefore cannot consume its own batches, and the central phase-conditioning pipeline is incomplete for real data. This release is insufficient to reproduce or verify the reported results.
  2. The released data pipeline contradicts the stated evaluation protocol. Validation and test boundaries subtract the lookback length twice, causing prediction targets to overlap the preceding split. At L=H=96, the first ETTh validation sample exactly duplicates the final training-set sample, including its target. This is an evaluation defect. The authors need to identify the code version used for the tables and establish that those results were obtained with correctly separated targets.
  3. The methodological contribution remains insufficiently established. PhiMixer combines existing cyclic decomposition, FiLM, and aggregate-redistribute processing without a direct accuracy comparison with CycleNet. The central PILCM claim is also overstated: shared phase-conditioned projections and uniform pooling do not substantiate the claimed recovery of asymmetric channel interactions. The paper does not demonstrate a distinct modeling advance beyond this integration of established components.
  4. Wrapping a complete cycle can be motivated by periodicity, but RRF wraps the input window without ensuring that it spans a complete cycle. On PEMS and Weather, the windows cover only part of the stated daily cycle, yet their first and last patches are treated as neighbors. Appendix A.2 argues that attention can override this soft prior, but does not justify the assumed relationship between these window boundaries. The ablation replaces RRF with an MLP rather than removing only the circular bias, leaving the claimed rhythmic benefit of this boundary assumption unverified.
  5. The architecture diagram and implementation disagree on the composition of the channel modules. Figure 1 shows STARPool and PILCM in parallel, whereas the code applies STARPool first and then PILCM to the updated representation. This changes the computation being described. The figure's dense layout, small labels, and unexplained final output projection further obscure the forward path. The diagram, equations, and released implementation do not provide a consistent account of the evaluated model.
  6. The channel interaction provided by PILCM may be restricted. Its phase-dependent transformations are shared across channels, while information is exchanged through uniform global aggregation. This does not establish the claimed ability to recover asymmetric channel interactions. Further analysis is needed to show whether the gains come from modeling channel relationships or from additional phase-conditioned feature processing.
Questions for authors
  1. Which exact code version generated the reported results? Provide an executable training and evaluation pipeline that supplies the forecast-origin index and implements the stated composite objective. Explain why the current release instead uses plain MSE and incompatible dataset outputs.
  2. Were the reported results generated with the released split-boundary logic? Identify the actual validation and test target ranges, and provide corrected results if overlapping targets were used.
  3. What substantive modeling contribution is established beyond the existing components? Which operation supports PILCM's claimed asymmetric channel interactions, and where is the comparison with a simpler CycleNet-based alternative under the same protocol?
  4. What evidence supports circular adjacency across partial-cycle window boundaries? Report the comparison with the same attention module after removing only its circular bias.
  5. Are STARPool and PILCM applied in parallel or sequentially in the model used for the experiments? Reconcile Figure 1 with the released forward computation, including the additional output projection shown after inverse normalization.
3 comments

We thank the reviewer for the careful audit of the released code and the paper. We address each weakness and question below; all corresponding changes are reflected in the revised manuscript and the revised release.

Weaknesses 1–2 / Q1–Q2 — Code corrections: Last time, we uploaded an older version of the code by mistake. The released code matches the version used for the paper's results: composite loss in run_main.py, t0 returned by the data loader, and split boundaries with a single - seq_len (validation/test targets are strictly disjoint from the preceding split).

Weakness 3 / Q3 — Methodological contribution and CycleNet comparison: We have added a direct comparison with CycleNet under the identical protocol — L = 96, the same splits, horizons, and seed as PhiMixer. A CycleNet column now appears in the main results Tables 1 and 2 (Page 08) and in the full per-horizon Tables 13–14 (Pages 22–23). The reproduced averages on the main benchmarks are:

DatasetPhiMixer MSECycleNet MSEPhiMixer MAECycleNet MAE
ETTh10.4330.4500.4270.438
ETTh20.3560.3770.3840.402
ETTm10.3590.3800.3770.398
ETTm20.2540.2670.3020.315
ECL0.1520.1890.2430.278
Weather0.2310.2430.2550.272
Exchange0.3690.3740.3990.409
Traffic0.4670.5120.2930.324

On the PEMS benchmarks:

DatasetPhiMixer MSECycleNet MSEPhiMixer MAECycleNet MAE
PEMS030.1170.1460.2270.252
PEMS040.1270.1660.2420.273
PEMS070.1330.1410.2330.245
PEMS080.1800.1930.2650.274

PhiMixer achieves lower average MSE and MAE than CycleNet on every dataset in both tables. Taken together with the phase-only ablations in Table 3, these results identify the distinct advance as the use of forecast-origin phase as a shared conditioning signal for both the temporal mixer (via FiLM) and the low-rank cross-channel operator (via PILCM): the CycleNet comparison isolates the gain over static cyclic tables, and Table 3 isolates the contribution of phase conditioning within each module.

Weaknesses 3 and 6 / Q3 — PILCM asymmetry and source of gain: We agree the original text overstated the asymmetric-interaction claim. Appendix A.2 now states that PILCM uses uniform channel-mean aggregation and does not model pairwise asymmetric interactions; learned asymmetric aggregation is left to future work (Appendix A.2, Page 16). To address whether PILCM's gain comes from channel mixing or from phase-conditioned feature processing, we decompose Table 3. On PEMS04 (average over horizons), removing PILCM entirely changes MSE from 0.127 to 0.136 (+7.1%), while removing only its phase conditioning changes it to 0.132 (+3.9%); the channel-mixing component and the phase-conditioned component therefore both contribute, and their contributions are separable (approximately 3.2% and 3.9% respectively). On ETTh1 the corresponding values are 0.433 (full), 0.449 (w/o PILCM, +3.7%), and 0.445 (w/o phase in PILCM, +2.8%). Thus the gain is not attributable to one component alone, and the phase-conditioned portion is separately measurable.

Weakness 4 / Q4 — RRF boundary justification and bias-only ablation: Section 3.3 adds a paragraph acknowledging that the circular distance is approximate when the look-back window spans only part of a full cycle (e.g., L = 96 against P = 288 on PEMS). It explains that the learned decay parameters adapt the strength of the prior per dataset, so the bias acts as a soft preference that the query–key product can override when non-cyclic structure is present (Page 05). In addition, we ran the specific bias-only ablation you suggested (attention module kept intact, circular bias set to S = 1). Removing only the bias changes average MSE by less than 1.3% across the five evaluated datasets (a small degradation on ETTh2, a small improvement on ETTh1). This shows that the primary benefit of RRF comes from the attention mechanism itself rather than the specific circular-distance prior. We report this finding in Appendix D (Pages 23–24) and reference it from Section 3.3, noting that the learned decay parameters can neutralize the bias when non-cyclic structure is present.

Weakness 5 / Q5 — Figure 1: We corrected the arrow flow in Figure 1. STARPool and PILCM are shown sequentially, matching the code, and the post-denormalization projection box has been removed. Fonts have been enlarged.

Thank you for your detailed response and revisions. You have partially addressed my concerns, although I still find the methodological novelty limited. Taking these improvements into account, I will raise my score from 1 to 2

We thank the reviewer for the score revision. On the remaining point -- methodological novelty -- we would like to state the contribution precisely, since our earlier response focused on correcting the PILCM description and may not have made the advance itself clear.

The contribution is not any individual component: cyclic decomposition, FiLM and STAR-style aggregation are prior work, as the reviewer notes. The advance is that forecast-origin phase is used as a shared conditioning signal for two operators that are otherwise independent — the temporal mixer (through FiLM, Section 3.3 (Page 05) ) and the cross-channel mixer (through PILCM, Section 3.4 (Page 06)) -- so that one parameter set produces different effective transformations at different positions of a cycle, while channel interaction remains linear in the number of channels (Section 4.5 (Page 10)). Conditioning the cross-channel operator, not only the temporal operator, on forecast-origin phase is the specific choice that prior periodic-modeling methods do not make (Section 4); the phase-weighted low-rank parameterization is what makes it affordable, and it is what allows phase to modulate how channel information is shared rather than only how temporal features are mixed.

Three results in the paper support this reading:

(i) CycleNet comparison (Tables 1–2 (Page 08) and 13–14 (Pages 22–23)): identical protocol, same look-back, splits, horizons, and seed. PhiMixer achieves lower MSE and MAE on all 12 datasets (e.g., ETTh1 0.433 vs. 0.450; ECL 0.152 vs. 0.189), which isolates the gain over the closest periodic-modeling baseline — a baseline that uses learnable cyclic tables but does not condition its transformations on phase.

(ii) Phase-only ablations (Table 3 (Page 08)): removing phase from both FiLM and PILCM produces the largest degradation among phase variants on every evaluated dataset (ETTh1 0.433 → 0.446; PEMS04 0.127 → 0.196), while removing phase from one module degrades less. This separates the phase-conditioning contribution from the architecture itself.

(iii) PILCM decomposition (Table 3 (Page 08)): removing PILCM entirely costs 7.1% MSE on PEMS04, while removing only its phase conditioning costs 3.9% (ETTh1: 3.7% and 2.8%). The phase-conditioned component of the channel operator is therefore separately measurable, and it contributes even when the channel-mixing path is intact.

Together these show that the phase-conditioning choice is doing work, that the gain is not attributable to cyclic decomposition or channel mixing alone and that the effect holds across datasets. These three results establish the contribution: phase conditioning of the channel operator is a specific architectural choice, it is measurable independently of the other components, and it produces consistent gains across datasets.

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

The paper addresses the problem of multivariate long-term time-series forecasting. The authors introduce PhiMixer, a lightweight phase-conditioned architecture, which includes two core components: rhythm-aware relative attention (RRF) and the Phase-Informed Low-Rank Channel Mixer (PILCM). The authors conduct experiments on 12 standard long-term forecasting benchmarks to demonstrate the effectiveness of the proposed solution.

Overall, the paper's motivation and problem definition are clear. The related work section is generally well explained. Extensive experiments on real datasets demonstrate that the proposed method performs better than state-of-the-art baseline methods. However, the technical contribution of this paper is somewhat unclear. It seems the proposed method mostly combines existing work into a single framework. Also, the accuracy improvement seems modest, indicating that the effectiveness of the proposed approach may be limited.

Critical strengths

(S) The motivation and problem definition are clear

(S) The authors conduct experiments on 12 standard long-term forecasting benchmarks

Critical weaknesses

(W) The technical contribution of this paper is somewhat unclear

(W) the accuracy improvement seems modest

Questions for authors

I think the font size of Figure 1 is too small. Is it possible to enlarge it?

2 comments

We thank the reviewer for the constructive comments.

W1 Technical Contribution. We have sharpened the contribution statement in Section 1. The specific contribution is not the combination of FiLM, patching, and aggregate–redistribute processing; it is the use of forecast-origin phase as a shared conditioning signal for both the temporal mixer (via FiLM) and the low-rank cross-channel operator (via PILCM), at O(Cdr) cost. This distinguishes PhiMixer from prior uses of periodicity: as decomposition, as frequency content, and as static cyclic tables (CycleNet). Furthermore, we have added a direct comparison with CycleNet under the identical protocol to Tables 1–2 and 13–14. This confirms that our conditioning-based approach yields empirical benefits over static cyclic tables (e.g., ETTh1 0.433 vs. 0.450; ECL 0.152 vs. 0.189).

W2 Modest Accuracy Improvement. We appreciate the concern and have made the trade-off explicit. PhiMixer's gains are obtained at ~126K parameters, linear channel scaling, and 1.37 ms CPU / 1.91 ms GPU inference — the smallest footprint among the highly competitive models. The paper's aim is efficient long-term forecasting with competitive accuracy: PhiMixer achieves the best average MSE on 6 of 12 benchmarks and the best average MAE on 8 of 12; Section 4.5 discusses this efficiency–accuracy balance explicitly. Where PhiMixer does not achieve the best absolute MSE (ETTh1, ETTh2, Traffic), Appendix I (Limitations, Failure Cases, and Future Work) documents the gap and its reasons.

We have enlarged the fonts in Figure 1.

Thank you for the detailed explanations. I do not have any further questions. I will maintain my original score.

Reviewer_ipnMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PhiMixer, a method for long-term time-series forecasting that incorporates forecast-origin phase information and combines it with modules including FiLM, Rhythm-Aware Relative Attention (RRF), and Phase-Informed Low-Rank Channel Mixer (PILCM) to model temporal features and channel interactions. The authors conduct comparative experiments against existing methods to investigate the effectiveness of the proposed approach and the roles of its individual components.

Critical strengths

S1. A potentially useful modeling direction. Explicitly incorporating forecast-origin phase information offers a reasonable approach to exploiting periodic structure in long-term time-series forecasting.

S2. A concrete architectural formulation. PhiMixer integrates phase information into temporal feature processing and channel interactions, providing a specific framework for investigating phase-conditioned forecasting.

Critical weaknesses

W1. Insufficient evidence to assess baseline comparison fairness. The paper does not adequately report the hyperparameter search spaces, tuning budgets, search strategies, or configuration-selection procedures for the baselines. As a result, it is difficult to establish whether the competing methods received reasonable and comparable tuning opportunities. This uncertainty directly undermines the interpretability and reliability of the main performance comparisons.

W2. The incremental novelty is not sufficiently established. PhiMixer combines forecast-origin phase information with several existing modeling mechanisms, including feature modulation, attention, and channel mixing. However, its substantive distinction from prior periodic modeling and conditioning approaches remains insufficiently clear.

W3. The independent contribution of phase conditioning is not fully established. The performance of the complete architecture alone cannot demonstrate that phase conditioning is responsible for the observed improvements. The experimental evidence should more clearly distinguish its contribution from those of the other components.

W4. The implications of period sensitivity remain unclear. The effectiveness of phase conditioning may depend on period specification and the stability of periodic patterns. The reported sensitivity analysis raises questions about how period variations affect robustness and applicability, which deserve a clearer discussion.

Questions for authors

Q1. Baseline tuning and comparison fairness: What hyperparameter search spaces, tuning strategies, computational budgets, and model-selection procedures were used for each baseline? Please distinguish results taken from original papers, official default configurations, and independently reproduced or tuned implementations.

Q2. Consistency of experimental protocols: Were all models evaluated using consistent data splits, input lengths, preprocessing procedures, and validation-based selection criteria? What evidence demonstrates that the baselines received reasonable and comparable tuning opportunities?

Q3. Novelty and component attribution: What is the central technical novelty of PhiMixer relative to existing periodic modeling and conditioning approaches? Which experiments directly establish the independent contribution of forecast-origin phase conditioning?

Q4. Period sensitivity and robustness: How are the period parameters determined, and how sensitive is the method to period misspecification or changes in periodic patterns? What do the reported sensitivity results imply for the method's robustness and applicability?

2 comments

We thank the reviewer for the substantive review. We address each weakness and question below; all corresponding changes are reflected in the revised manuscript.

W1 / Q1–Q2 — Appendix B.4 (Baseline Results and Protocol Alignment) documents the source and configuration of every baseline. All baseline results are quoted from their original publications; we did not re-tune or re-run any baseline. PhiMixer uses a single fixed configuration across all 12 datasets with no per-dataset hyperparameter search. The only baseline reproduced under our exact protocol is CycleNet, added in Tables 1–2 (Page 08) and 13–14 (Pages 22–23).

W2 / Q3 — The specific technical contribution is already stated in Section 1 (Page 02). Section 4 adds an explicit conceptual distinction: decomposition- and frequency-based methods use periodicity to construct or separate representations, cyclic-table methods use periodicity to provide a seasonal prior, and PhiMixer uses forecast-origin phase to condition the transformations applied to the representation. This is supported empirically by the phase-only ablations and by the new CycleNet reproduction in Tables 1–2 (Page 08) and 13–14 (Pages 22–23), which demonstrates that our conditioning-based approach outperforms static cyclic tables (e.g., ETTh1 0.433 vs. 0.450).

W3 / Q3 — The phase-only variants (w/o P. in FiLM, w/o P. in PILCM, w/o Phase) are already there in Table 3; Section 4.2 is the paragraph that interprets them (Pages 07–08). Removing phase from both modules produces the largest degradation among the phase variants on every evaluated dataset (ETTh1 0.433 → 0.446; PEMS04 0.127 → 0.196). The two phase contributions are complementary and, on PEMS, super-additive rather than redundant.

W4 / Q4 — Appendix D.1 (Pages 24–27) clarifies four points: (i) the period is derived from each dataset's sampling rate as ⌊1440/Δt⌋, not tuned; (ii) the sampling-derived value is optimal or near-optimal on all five evaluated datasets; (iii) the effect of phase conditioning scales with the strength of periodicity in the underlying signal — negligible on ETTh1, substantial on Weather, and largest on PEMS; (iv) for datasets without a clean dominant cycle, extending to a learned or DFT-detected period is identified as future work.

Dear Reviewer ipnM,

We are writing to respectfully follow up on our rebuttal, as we have not yet received your feedback. We deeply value your expertise (Confidence: 4), and we recognize that your concerns regarding baseline fairness (W1) and methodological novelty (W2) are the most critical to our paper's evaluation. We want to ensure that our responses have adequately addressed these specific points.

Regarding W1 (Baseline Fairness): We appreciate this concern. We want to clarify our protocol precisely: all baseline results in the main tables are quoted directly from their original publications (with their own optimized hyperparameters), and we did not re-tune them. To strengthen the fairness of the comparison, we have added a direct reproduction of CycleNet under our exact protocol (identical L=96, splits, horizons, and seed). PhiMixer outperforms CycleNet on all 12 datasets. Additionally, we have added Appendix B.4, which explicitly lists the source (original paper / official config / reproduced) and the tuning budget for every baseline. If you believe further matched-budget comparisons are necessary, we are fully prepared to include them in the revised manuscript.

Regarding W2 (Novelty): We have sharpened the contribution statement to clarify that the specific advance is not the combination of existing components (FiLM, cyclic tables, STAR pooling), but rather the use of forecast-origin phase as a shared conditioning signal for both the temporal mixer (via FiLM) and the low-rank cross-channel operator (via PILCM). The CycleNet comparison and the phase-only ablations (Table 3) directly isolate this contribution, showing that phase conditioning provides measurable gains beyond static cyclic tables.

Regarding W3 and W4: As detailed in our previous response, the phase-only ablations (Table 3) isolate the contribution of phase conditioning, and we have clarified that the period is derived directly from the sampling rate (⌊1440/Δt⌋) and is optimal or near-optimal across evaluated datasets, with phase effects scaling with the strength of periodicity.

We would be very grateful if you could share whether these clarifications and the new experimental evidence address your concerns. If there are specific additional experiments or analyses you would like to see , please let us know, and we will do our best to provide them.

Thank you for your time and for your rigorous review, which has helped us substantially improve the paper

Comments on the paper

1 comment

Dear reviewers,

The authors have now submitted their responses. Please review them and share your thoughts on whether your concerns have been addressed. I encourage you to engage with one another’s assessments, discuss any remaining disagreements, and update your reviews and scores as appropriate.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.