- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
PeakTS augments a decomposition-based forecaster with an ordered categorical distribution and general/peak prototype memories queried through time, difference, and frequency representations. A learned gate blends the point forecast with a mixture of the distributional mean and 90th percentile. Training combines L1 forecasting, frequency-weighted cumulative cross-entropy, and consistency losses. Experiments cover nine datasets and four overall forecasting horizons, with high-value evaluation at horizon 96.
Critical strengths
- Consistent MAE improvements across diverse datasets. Table 1 reports the best Top-25% MAE on all nine datasets. For example, ETTm1 improves from the best competing 0.395 to 0.378, and CzeLan from 0.310 to 0.294. This is meaningful evidence for usefulness on the evaluated high-value subsets.
- Component-level evidence supports the design. Removing correction or memory worsens Top-25% performance on all three ablation datasets (Table 2). Appendix F/Table 5 additionally tests the temporal views; the full combination improves over each reduced-view configuration. These experiments are more informative than aggregate benchmark wins alone.
Critical weaknesses
- Comparisons do not sufficiently isolate the contribution. PeakTS explicitly optimizes L1 (Eq. 17), but baseline training losses and matched tuning budgets are not documented in Section 4.1/Appendix C. Thus, the especially consistent MAE gains cannot be cleanly attributed to the proposed mechanisms. The closest ordinal baseline, OCE-TS, has Weather overall MSE 36.91 versus PeakTS's 0.155 (Table 1). The authors acknowledge modified preprocessing to remove leakage (p. 7), but do not specify or validate that modification. This is not evidence of an implementation error, but it leaves a crucial comparison unauditable. No seed variability is reported, including for small gains such as Electricity Top-25% MAE 0.261 to 0.259.
- The distributional correction's tail behavior is insufficiently specified. Figure 1 standardizes using historical mean/std, while Section 3.1 initializes bin support near [-1,1], without explaining boundary adaptation or out-of-support targets. For normalized values, if y_true >= y_raw > the upper bin edge, the bounded mean/quantile target pulls the forecast downward whenever the gate is positive. Proposition 3.1 assumes appropriately directed, non-overshooting correction; it does not establish how frequently that condition holds. Boundary-case evidence is therefore important. More generally, Top-25% evaluation at horizon 96 supports a narrower claim than reliable rare-peak forecasting. Its MSE gains are also not universal: Weather is 0.311 versus the best baseline's 0.283, and ZafNoo 0.646 versus 0.614.
Questions for authors
- Do the gains persist against L1-trained, equally tuned DLinear and a verified input-only-normalized OCE-TS? Please specify the OCE-TS preprocessing changes and report variability across seeds.
- How are bin boundaries updated and out-of-support targets handled? Please give the actual normalization, smoothing/stability parameters, out-of-support frequency, and correction performance on those cases.
I have read the authors’ rebuttal. My questions have been addressed. The additional Top-5% experiments address my concern about forecasting performance in more extreme high-value regions. I would like to raise my rating.