- 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?
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