- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
PKTime is a linear forecaster. It selects a period set from the training data (FFT plus a permutation test), centers each variable's look-back window (), folds it into (positions within a period past cycles), and predicts
with a full cross-cycle matrix . Branches are combined with global weights , and all parameters are shared across variables. On eight standard benchmarks against nine baselines reproduced in TFB, PKTime has the most first places (15 of 32 tasks for MSE, 22 of 32 for MAE).
Critical strengths
- The method is simple and fast at inference, and anonymized code is available.
- Period selection uses training data only and is fully specified in the appendix
Critical weaknesses
-
This is SparseTSF plus one matrix. With D = I, the model is a cross-cycle linear map shared by all positions in the period, which is the forecasting step of SparseTSF. The only addition is D, a diagonal plus rank-2 matrix. The "Kronecker structure" is the identity (the paper's Eq. 3), which holds for any two-sided linear map. The multi-period part is idle: 6 of 8 datasets select a single period (Table 7), and on ETTm1 learned fusion equals uniform fusion (MSE 0.360 vs 0.361, Table 4).
-
The gains are noise-level, and the comparison is weak. Against SparseTSF, the mean relative MSE change over the 32 tasks of Table 1 is -0.9%, and PKTime is worse on average on ETTh1 and Exchange. No standard deviations or seed counts are given. On Exchange at H = 720, PKTime ranks 7th of 10 (0.855 vs 0.578 for FITS). Five of the nine baselines are from 2022 or 2023; PatchTST, iTransformer, and CycleNet are missing, and the look-back and tuning of the baselines are not reported.
-
The ablations and efficiency study undercut the design. The single-period variant on ETTm1 (MSE 0.446) is far worse than SparseTSF (0.359), so the 23.8% multi-period gain is measured against a weak variant. A factorized W gives lower MSE on ETTh1 and ETTm1 (Table 3), and raising r above 1 changes nothing on ETTh1 and ETTh2 (Table 5). MixLinear has fewer parameters than PKTime at every horizon on Electricity (Fig. 3a), so PKTime is not even the lightest model.
Questions for authors
See weaknesses.
5 comments
Response to Reviewer Snac
W2: Thank you for your question about experiments. We have added experiments for PKTime across eight datasets and four forecast horizons, repeated with three seeds, and report the means and sample standard deviations computed from the same set of runs. The training seeds are 2021, 2022, and 2023.
| Dataset | ||||
|---|---|---|---|---|
| ETTh1 | 0.358 ± 0.002 / 0.388 ± 0.001 | 0.399 ± 0.002 / 0.411 ± 0.001 | 0.427 ± 0.002 / 0.426 ± 0.001 | 0.424 ± 0.002 / 0.446 ± 0.002 |
| ETTh2 | 0.280 ± 0.003 / 0.339 ± 0.002 | 0.341 ± 0.006 / 0.377 ± 0.002 | 0.364 ± 0.002 / 0.399 ± 0.002 | 0.385 ± 0.003 / 0.424 ± 0.001 |
| ETTm1 | 0.317 ± 0.002 / 0.357 ± 0.002 | 0.342 ± 0.003 / 0.372 ± 0.003 | 0.371 ± 0.002 / 0.388 ± 0.002 | 0.420 ± 0.003 / 0.416 ± 0.001 |
| ETTm2 | 0.179 ± 0.002 / 0.273 ± 0.002 | 0.234 ± 0.001 / 0.308 ± 0.001 | 0.287 ± 0.001 / 0.342 ± 0.001 | 0.371 ± 0.003 / 0.394 ± 0.001 |
| Electricity | 0.139 ± 0.000 / 0.234 ± 0.001 | 0.152 ± 0.000 / 0.245 ± 0.001 | 0.167 ± 0.000 / 0.261 ± 0.000 | 0.206 ± 0.000 / 0.295 ± 0.001 |
| Traffic | 0.393 ± 0.000 / 0.268 ± 0.001 | 0.402 ± 0.000 / 0.272 ± 0.000 | 0.415 ± 0.001 / 0.280 ± 0.001 | 0.452 ± 0.001 / 0.299 ± 0.001 |
| Solar | 0.197 ± 0.002 / 0.237 ± 0.003 | 0.243 ± 0.065 / 0.267 ± 0.054 | 0.230 ± 0.005 / 0.250 ± 0.004 | 0.235 ± 0.005 / 0.250 ± 0.002 |
| Exchange | 0.084 ± 0.001 / 0.202 ± 0.001 | 0.171 ± 0.001 / 0.294 ± 0.002 | 0.318 ± 0.001 / 0.407 ± 0.002 | 0.866 ± 0.010 / 0.695 ± 0.002 |
In addition, PKTime’s hyperparameter tuning experiments evaluated all six combinations of input lengths 336 and 720 and learning rates 0.001, 0.01, and 0.03. Each candidate configuration was trained with three seeds. For each dataset and forecast horizon, we selected the configuration with the lowest mean validation MSE and reported the test mean and standard deviation for that configuration.
The baseline results in the main table are from the TFB evaluation published by OpenTS and correspond to its publicly reported configurations.
We will include these complete results and the configuration selection procedure in the revised manuscript. We will also provide source citations and traceable configuration records for the OpenTS baselines to clarify the experimental settings underlying each comparison.
Besides, we conduct siginificant test on PKtime and other baselines. The results demonstrate the improvement of PKTime is significant in 10 out of 17 cases. We have updated the results and comparisons with PatchTST, iTransformer and CycleNet in the revised manuscript.
Response to Reviewer Snac
W3: Thank you for your careful analysis of the ablation results and efficiency comparison. We have added a single-period reference selected by validation, controls with duplicate period branches, and experiments on parameterization and rank using three random seeds. Based on these results, we have reassessed the benefits of each design choice.
The original Single variant retained only the first selected period, without using validation performance to select a stronger single-period configuration. Therefore, the ETTm1 result of 0.446 was indeed an inadequate single-period reference. In addition, the original value of 23.829% used the error of Full as the denominator and represented the error increase of Single relative to Full.
We added two single-period candidates, 48 and 96, and selected between them using the mean validation MSE across three seeds. Period 48 was selected at all four forecast horizons. We also added two period-48 branches with independent parameters to distinguish the effect of adding branches from that of using different periods. All values below are errors averaged across the four forecast horizons.
| ETTm1 configuration | MSE | MAE |
|---|---|---|
| Single period 48 selected by validation | 0.378869 | 0.395806 |
| Two period-48 branches with independent parameters | 0.364599 | 0.385095 |
| Two branches with different periods [48, 96] | 0.36286 | 0.383284 |
| Single period 48 with an approximately matched parameter budget | 0.378894 | 0.395548 |
| Duplicate period-48 branches with an approximately matched parameter budget | 0.363965 | 0.384708 |
The two-period configuration reduces the average MSE by 4.23% relative to the single period selected by validation and by 0.48% relative to the duplicate period-48 branches. After adjusting the rank to keep parameter count deviations within 10%, these gaps are 4.23% and 0.30%, respectively. The improvement over duplicate period-48 branches mainly occurs at forecast horizons of 192 and 336; at 96 and 720, the duplicate period-48 branches achieve lower MSE and MAE.
These comparisons show that adding independent branches accounts for a substantial part of the improvement, while using different periods provides a smaller additional average gain. We will include these results and use the 4.23% reduction relative to the single period selected by validation to describe the corresponding benefit of using multiple periods.
Response to Reviewer Snac
W1: Thank you for highlighting the connection between PKTime and SparseTSF. At the level of the single-period prediction head, PKTime indeed extends the shared cross-cycle mapping used in SparseTSF by introducing an intra-period matrix D . We aim to preserve the simplicity of this prediction scheme while enabling information exchange across phases. The points that warrant further clarification are which structural constraints this matrix changes, how its parameter overhead is controlled, and whether the experiments support this design.
After folding the input by period, let Z∈Rp×N and W∈RM×N, where p is the period length, and N and M are the numbers of historical and future cycles, respectively. The rows of Z correspond to phases within a period, and its columns correspond to historical cycles. The shared cross-cycle prediction head and the PKTime prediction head are, respectively,
Yshared=ZW⊤,YPKTime=DZW⊤.
The shared head applies the same W to each phase independently. With D, the prediction can be expanded as
Yi,m=∑j=1p∑n=1NDijWmnZj,n.
Thus, when predicting a particular future phase, the model can also use historical information from other phases. D learns relationships between phase pairs, while W learns relationships between historical and future cycles. D is shared across cycle pairs, and W is shared across phase pairs. This design separately parameterizes phase interactions and cross-cycle prediction while avoiding an independent coefficient for every pair of input and output positions.
To control the number of parameters used for phase interactions, we adopt
D=Diag(d)+UV⊤,
where U,V∈Rp×r. Its action can be written as
Dz=d⊙z+U(V⊤z).
The diagonal term preserves each phase’s own information; V extracts r shared linear combinations of phases, and U distributes these combinations to the output phases. Thus, r=2 corresponds to two shared phase combinations. The rank constraint applies to the interaction correction term, while D as a whole can still be full rank. This parameterization uses p+2pr parameters. For example, with p=24 and r=2, D contains 120 parameters—96 more than a diagonal mapping and 456 fewer than a full 24×24 matrix.
This modification also has a clear implication for expressivity. For a shared linear prediction head with a fixed nonzero W, if another shared matrix W satisfies
DZW⊤=ZW⊤,∀Z,
then necessarily
D=cIp,W=cW.
This follows because, for any Wmn=0, the corresponding operator blocks must satisfy
WmnD=WmnIp.
Therefore, a phase-dependent mapping cannot, in general, be reproduced by adjusting the shared W. Even a rank-1 correction of the form
D=Ip+βeiej⊤,
with i=j and β=0, already establishes a predictive connection from input phase j to output phase i.
The full SparseTSF model also includes convolutional aggregation. Convolution shares weights according to temporal lag, whereas D learns weights for phase pairs within a period and reuses them across cycles. For example, the same temporal lag can correspond to different coefficients in D at different phases. The two therefore impose different structural constraints.
In the revised manuscript, we will explicitly acknowledge the direct connection to SparseTSF, explain Eq. (3) in terms of its parameter-sharing structure, and organize the method description and experimental analysis around “learning intra-period phase relationships through a diagonal-plus-low-rank mapping while retaining shared cross-cycle prediction.” We will also provide separate comparisons for the multi-period representation, the specific selection rules, and the fusion strategy to clarify the role of each design choice.