acceptodds
Under review as a conference paper at ICLR 2027

LUMA: Ultralight Low-Rank Forecasting for Periodic Time Series

Abstract

Long-term Time Series Forecasting (LTSF) faces a growing dichotomy between parameter-heavy foundation models and lightweight linear models. While recent trends favor massive scale, in this paper, we investigate whether the spectral compactness of periodic signals can motivate efficient low-rank forecasting. By combining period-folding with a rank-constrained operator, we reduce parameter growth through a compact cross-period representation. We introduce **LUMA** (**L**ow-rank **U**ltralight **M**atrix **A**pproximation), a minimalist architecture designed to exploit this efficiency. LUMA couples period-folding with a rank-constrained operator to 1) preserve within-period positions in a cross-period representation and 2) forecast their evolution through shared low-rank factors. We further prove that for a -harmonic signal, the period-folded representation has rank at most , providing a structural motivation for rank-constrained forecasting. Extensive experiments on seven benchmarks demonstrate that LUMA achieves the best MSE on four datasets (ETTh1, ETTh2, ETTm2, Weather) while remaining competitive on the rest. In the Electricity profiling configuration, LUMA uses only **635 parameters**, approximately fewer than PatchTST and fewer than TimeKAN. Relative to SparseTSF, this configuration reduces parameters by approximately **36%** and MACs by approximately **18%**, with a reported CPU inference latency of **1.3 ms**. These results support low-rank cross-period modeling as an effective approach to accurate forecasting with a small parameter budget.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DXCCOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes LUMA, an ultralight, channel-independent linear model for long-term forecasting of periodic time series. The input window is mean-centered, smoothed by a residual convolution, and folded by its dominant period so that values at the same within-period position are lined up across successive cycles. A single low-rank linear operator, shared across positions and channels, then maps past cycles to future cycles, and the result is unfolded back into a forecast. The low-rank choice is motivated by a theoretical argument that folded periodic signals are compact, that is, they can be described by only a few components. Experiments on seven standard benchmarks compare LUMA with Transformer, linear and efficient baselines, and add efficiency profiling, rank and period sensitivity, noise robustness, ultra-long horizons, a fixed-lookback comparison and weakly periodic datasets.

Critical strengths

The model is simple and transparent, and the low-rank constraint has a sensible intuition. Folding the dominant period removes that period, so the cross-cycle forecasting map only has to capture a few slow patterns, and restricting its rank is a reasonable inductive bias in the spirit of reduced-rank regression. The whole model is a structured linear map with explicit parameter and compute formulas and public code, which makes it easy to inspect and to deploy in resource-constrained settings such as CPU-only devices.

Beyond the standard benchmarks and horizons, the paper includes a fixed-lookback comparison, rank and period sensitivity, noise robustness and ultra-long horizons, and it reports multi-seed variability for LUMA. It also shows that the method falls behind on weakly periodic or aperiodic data and discusses this as a limitation, which helps readers judge where the inductive bias applies.

Critical weaknesses

The contribution over SparseTSF is incremental, and the low-rank mechanism is not cleanly demonstrated. Mean-centering, the residual convolution and period folding are the same as in SparseTSF, and the only change is factorizing the cross-cycle linear layer; framing the two as spectral versus spatial sparsity is misleading, since SparseTSF's cross-cycle map is itself dense. The theory bounds the rank of the folded data rather than of the forecasting operator, and it does not match the high effective ranks the paper itself measures. The rank ablation's U-shape comes mainly from a single short-horizon setting, where widths beyond the folded output dimension cannot raise the attainable rank, while other datasets show a plateau. The authors should compare dense and low-rank operators in the same pipeline across datasets and horizons, and tie the theory to the operator itself.

LUMA uses a long lookback with rank and period tuned per dataset and horizon, whereas baselines largely keep their reported settings; under a common lookback, most of the headline wins disappear, and baselines carry no uncertainty estimates. The parameter savings over SparseTSF amount to a few hundred parameters and nearly vanish at short horizons, and the reported latency reduction is much larger than the compute reduction, which points to implementation differences. The authors should rerun the comparison with matched lookback, tuning and seeds, and time both models in the same codebase.

Questions for authors

What lookback does LUMA use in Tables 1 and 4, and what lookback do the periodic baselines in Table 9 use? How were validation and early stopping handled at the longest Electricity horizon?

Were SparseTSF and LUMA timed in the same implementation, with identical batch size, thread settings and warm-up?

Reviewer_eu2jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies efficient long-term forecasting of periodic multivariate time series. It proposes LUMA, which combines residual temporal convolution, period folding, and a shared low-rank cross-period linear operator. Experiments on seven standard benchmarks, additional weakly periodic datasets, rank and period sensitivity analyses, and CPU profiling show competitive accuracy with a small parameter budget, including 635 parameters in the reported Electricity configuration.

Critical strengths
  • Treating compactness as a constraint on the cross-period forecasting operator is a sensible solution strategy for efficient forecasting.
  • The paper proposes LUMA, a transparent forecasting model with a controllable rank budget and few components to implement or tune.
  • The measured accuracy-efficiency trade-off is useful.
Critical weaknesses
  • The technical novelty over the previous cross-period linear baseline is limited. The paper does not establish why this particular factorization is preferable to directly compressing the same trained map or applying conventional matrix regularization.
  • The profiling does not validate the stated deployment setting. Table 2 and Appendix B.4 use a Xeon server CPU to simulate deployment, and the reported process memory remains 90-94 MB for LUMA. Parameter counts alone therefore do not demonstrate feasibility on the few-megabyte devices motivating the introduction.
  • For reproducibility, the rank and folding configurations used for the reported results are incompletely specified.
Questions for authors
  • How does LUMA compare with truncated-SVD compression of SparseTSF?
  • Can you provide the exact lookback, folding interval, requested and effective rank, and selection rule for every dataset-horizon pair, together with the validation search space and corresponding baseline tuning budgets?
Reviewer_8vxNOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LUMA exploits the periodic structure of time series by transforming the original sequence into a cross-period representation through period folding, followed by a low-rank mapping for forecasting.

Critical strengths

Compared with methods that rely on complex neural network architectures, this approach of designing a lightweight model based on the structural properties of time series is commendable.Novelty of the proposed method.

Critical weaknesses

none.

Questions for authors

Although LUMA employs a low-rank mapping for forecasting, the core operation appears to be essentially a conventional low-rank linear transformation. The authors should further clarify whether the main methodological novelty lies in the period-folding strategy, the low-rank mapping, or their combination, and explain how the proposed approach differs from existing low-rank forecasting methods

Comments on the paper

3 comments

Dear reviewers,

The authors have submitted their response to your review comments. Please kindly take a moment to read their response and assess whether your concerns have been adequately addressed. Further discussion with the authors to reach a deeper mutual understanding would be greatly appreciated.

Thanks, AC

Thank you for the continued discussion and for raising your score. We understand that your remaining concerns are novelty and the improvement over existing solutions. We respectfully clarify both.

1. Novelty

A new finding about the forecasting operator. LUMA's contribution is not the factorization itself but what motivates it: after period folding, the forecasting operator of periodic series is intrinsically low rank. This was not known or exploited before. SparseTSF learns a dense cross-period matrix and makes no assumption on its spectrum. We show that this matrix, trained without any constraint, already concentrates 90% of its energy in 4 of 30 (ETTh1), 9 of 60 (ETTh2) and 8 of 60 (ETTm1) directions, even though the input data has high rank (225 to 387).

A complementary prior. SparseTSF's full map W⊗IPW\otimes I_P is sparse in its support, but its rank is unconstrained, up to P⋅min⁡(L′,H′)P\cdot\min(L',H'). LUMA keeps this support and constrains the spectrum, since rank(W⊗IP)=P⋅rank(W)\mathrm{rank}(W\otimes I_P)=P\cdot\mathrm{rank}(W). To our knowledge, LUMA is the first forecaster to combine both priors.

Theory. Theorem 3.2 bounds the rank of folded harmonic signals by 2K2K, Proposition 3.1 transfers this bound to the least-squares operator, and the factorization can represent this optimum exactly whenever rr reaches its rank.

Precedent in this line of work. DLinear (AAAI 2023), FITS (ICLR 2024) and SparseTSF (ICML 2024, oral) each introduced one minimal structural change to a linear forecaster, together with evidence of why it works. LUMA follows the same principle. We believe its contribution should be judged by the insight and the evidence behind it rather than by architectural complexity.

2. Improvements over existing solutions

  • Accuracy. LUMA has the best average MSE on 4 of 7 datasets in Table 1, and is best or tied for best on 4 of 7 under a fixed lookback (Table 12), more than any other method, including PatchTST (2 of 7).
  • Against the closest baseline. LUMA has lower MSE than SparseTSF in 24 of 28 benchmark settings, and lower than the dense operator in 23 of 31 matched settings (sign test, p≈0.005p\approx 0.005).
  • Against periodic methods. LUMA has the lowest MSE on 6 of 7 datasets compared with CycleNet, TQNet and FreqCycle (Table 13).
  • Robustness. Under input noise (σ=0.2\sigma=0.2), LUMA's error grows by 78% against 148% for SparseTSF (Table 3).
  • Ultra-long horizons. At H=2880H=2880, LUMA outperforms PatchTST (0.254 vs. 0.265) and DLinear (0.289), with parameters growing only with H/PH/P (Table 4).
  • Efficiency. 635 parameters on Electricity: 36% fewer than SparseTSF, 36× fewer than TimeKAN and about 13,000× fewer than PatchTST, with 18% fewer MACs than SparseTSF (Table 2).

In the regime of sub-1K-parameter linear forecasters, improvements are naturally measured in a few percent of MSE. LUMA delivers them consistently, together with gains in robustness and efficiency, at no additional cost.

We respect your assessment and thank you again for a constructive and engaged discussion, which clearly strengthened the paper.

Dear Area Chair,

We sincerely thank you for your engagement during the discussion period and for encouraging the reviewers to respond to our rebuttal. We also thank the reviewers for a constructive exchange. We briefly summarize the outcome and the revisions (marked in blue in the revised PDF).

Discussion outcome. All reviewers engaged constructively. Two raised their scores after the new results, and the third already gave the maximum score.

  • Reviewer 8vxN confirmed that their concerns have been well addressed.
  • Reviewer eu2j stated that the new SVD and nuclear-norm comparisons address their main concern and updated their rating.
  • Reviewer DXCC increased their score from 1 to 2, recognizing that our rebuttal successfully addressed several key concerns, particularly the rank analysis under high-rank inputs and the expanded baseline comparisons.

Novelty: LUMA shows that, after period folding, the cross-period forecasting operator of periodic series is intrinsically low rank. We support this with theory (Theorem 3.2, Proposition 3.1) and with direct measurement: even with high-rank inputs, the operator trained without any constraint concentrates 90% of its energy in 4 of 30, 9 of 60 and 8 of 60 directions (ETTh1, ETTh2, ETTm1). Imposing this structure explicitly gives consistent gains in accuracy, robustness and parameter efficiency. In this respect, LUMA follows the line of DLinear, FITS and SparseTSF, each of which introduced one minimal structural prior on a linear forecaster together with evidence of why it works.

Main additions:

  • Controlled ablation (Appendix D). Dense vs. low-rank operator under identical configurations, with post-hoc SVD truncation and nuclear-norm regularization. LUMA improves on the dense operator in 23 of 31 matched settings (p≈0.005p\approx 0.005).
  • Operator rank (Appendix D). Effective rank of the unconstrained operator and a rank sweep at H=720H=720, where all ranks up to 14 outperform the dense map.
  • Extended baselines (Table 12). PatchTST and FEDformer added at a fixed lookback L=720L=720. LUMA remains best or tied for best on 4 of 7 datasets, more than any other method.
  • Reproducibility and clarifications. Full per-setting configurations (Appendix B.6), timing and memory protocol (Appendix B.4), clarified sparsity definitions (Definition A.1) and the parameter break-even rule (Appendix J).

Thank you again for your time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.