acceptodds
Under review as a conference paper at ICLR 2027

PKTime:Lightweight Time Series Forecasting via Kronecker-Structured Modeling with Adaptive Period Selection

Abstract

Lightweight time-series forecasting requires compact models that capture temporal dependencies effectively. We propose PKTime, a lightweight forecasting model combining Kronecker-structured modeling with adaptive period selection. PKTime selects periods from training data and organizes historical observations into intra-period and inter-period dimensions. Each branch uses a diagonal-plus-low-rank mapping for intra-period interactions and a dense mapping for inter-period dependencies, yielding a compact Kronecker-separable predictor. Multiple branches and their fusion weights are jointly learned to capture different periodic structures, with parameters shared across variables. Period selection occurs only at initialization, and forecasting operates entirely in the time domain. Experiments on eight benchmark datasets demonstrate state-of-the-art forecasting performance. Source code is available at https://anonymous.4open.science/r/PKTime-CE08.

Primary area:
learning on time series and dynamical systems
Keywords:
lightweight time series forecasting, kronecker decomposition, adaptive period selection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_SnacOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PKTime is a linear forecaster. It selects a period set P\mathcal{P} from the training data (FFT plus a permutation test), centers each variable's look-back window (L=720L = 720), folds it into Zp∈Rp×NpZ_p \in \mathbb{R}^{p \times N_p} (positions within a period ×\times past cycles), and predicts

Qp=DpZpWp⊤,Dp=diag(dp)+UpVp⊤,Up,Vp∈Rp×r, r=2,Q_p = D_p Z_p W_p^\top, \qquad D_p = \mathrm{diag}(d_p) + U_p V_p^\top, \quad U_p, V_p \in \mathbb{R}^{p \times r},\ r = 2,

with a full cross-cycle matrix Wp∈RMp×NpW_p \in \mathbb{R}^{M_p \times N_p}. Branches are combined with global weights α=softmax(a)\alpha = \mathrm{softmax}(a), 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
  1. The method is simple and fast at inference, and anonymized code is available.
  2. Period selection uses training data only and is fully specified in the appendix
Critical weaknesses
  1. 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 vec(DZW⊤)=(W⊗D) vec(Z)\mathrm{vec}(DZW^\top) = (W \otimes D)\,\mathrm{vec}(Z) (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).

  2. 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.

  3. 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
AuthorsOpenReview ↗

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 DD . 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×NZ\in\mathbb R^{p\times N} and W∈RM×NW\in\mathbb R^{M\times N}, where pp is the period length, and NN and MM are the numbers of historical and future cycles, respectively. The rows of ZZ 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⊤.Y_{\mathrm{shared}}=ZW^\top, \qquad Y_{\mathrm{PKTime}}=DZW^\top.

The shared head applies the same WW to each phase independently. With DD, the prediction can be expanded as

Yi,m=∑j=1p∑n=1NDijWmnZj,n.Y_{i,m}=\sum_{j=1}^{p}\sum_{n=1}^{N}D_{ij}W_{mn}Z_{j,n}.

Thus, when predicting a particular future phase, the model can also use historical information from other phases. DD learns relationships between phase pairs, while WW learns relationships between historical and future cycles. DD is shared across cycle pairs, and WW 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⊤,D=\operatorname{Diag}(d)+UV^\top,

where U,V∈Rp×rU,V\in\mathbb R^{p\times r}. Its action can be written as

Dz=d⊙z+U(V⊤z).Dz=d\odot z+U(V^\top z).

The diagonal term preserves each phase’s own information; VV extracts rr shared linear combinations of phases, and UU distributes these combinations to the output phases. Thus, r=2r=2 corresponds to two shared phase combinations. The rank constraint applies to the interaction correction term, while DD as a whole can still be full rank. This parameterization uses p+2prp+2pr parameters. For example, with p=24p=24 and r=2r=2, DD contains 120 parameters—96 more than a diagonal mapping and 456 fewer than a full 24×2424\times24 matrix.

This modification also has a clear implication for expressivity. For a shared linear prediction head with a fixed nonzero WW, if another shared matrix W~\widetilde W satisfies

DZW⊤=ZW~⊤,∀Z,DZW^\top=Z\widetilde W^\top, \qquad \forall Z,

then necessarily

D=cIp,W~=cW.D=cI_p, \qquad \widetilde W=cW.

This follows because, for any Wmn≠0W_{mn}\ne0, the corresponding operator blocks must satisfy

WmnD=W~mnIp.W_{mn}D=\widetilde W_{mn}I_p.

Therefore, a phase-dependent mapping cannot, in general, be reproduced by adjusting the shared WW. Even a rank-1 correction of the form

D=Ip+βeiej⊤,D=I_p+\beta e_i e_j^\top,

with i≠ji\ne j and β≠0\beta\ne0, already establishes a predictive connection from input phase jj to output phase ii.

The full SparseTSF model also includes convolutional aggregation. Convolution shares weights according to temporal lag, whereas DD 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 DD 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.

AuthorsOpenReview ↗

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.

DatasetH=96H=96H=192H=192H=336H=336H=720H=720
ETTh10.358 ± 0.002 / 0.388 ± 0.0010.399 ± 0.002 / 0.411 ± 0.0010.427 ± 0.002 / 0.426 ± 0.0010.424 ± 0.002 / 0.446 ± 0.002
ETTh20.280 ± 0.003 / 0.339 ± 0.0020.341 ± 0.006 / 0.377 ± 0.0020.364 ± 0.002 / 0.399 ± 0.0020.385 ± 0.003 / 0.424 ± 0.001
ETTm10.317 ± 0.002 / 0.357 ± 0.0020.342 ± 0.003 / 0.372 ± 0.0030.371 ± 0.002 / 0.388 ± 0.0020.420 ± 0.003 / 0.416 ± 0.001
ETTm20.179 ± 0.002 / 0.273 ± 0.0020.234 ± 0.001 / 0.308 ± 0.0010.287 ± 0.001 / 0.342 ± 0.0010.371 ± 0.003 / 0.394 ± 0.001
Electricity0.139 ± 0.000 / 0.234 ± 0.0010.152 ± 0.000 / 0.245 ± 0.0010.167 ± 0.000 / 0.261 ± 0.0000.206 ± 0.000 / 0.295 ± 0.001
Traffic0.393 ± 0.000 / 0.268 ± 0.0010.402 ± 0.000 / 0.272 ± 0.0000.415 ± 0.001 / 0.280 ± 0.0010.452 ± 0.001 / 0.299 ± 0.001
Solar0.197 ± 0.002 / 0.237 ± 0.0030.243 ± 0.065 / 0.267 ± 0.0540.230 ± 0.005 / 0.250 ± 0.0040.235 ± 0.005 / 0.250 ± 0.002
Exchange0.084 ± 0.001 / 0.202 ± 0.0010.171 ± 0.001 / 0.294 ± 0.0020.318 ± 0.001 / 0.407 ± 0.0020.866 ± 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.

AuthorsOpenReview ↗

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 WW 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 configurationMSEMAE
Single period 48 selected by validation0.3788690.395806
Two period-48 branches with independent parameters0.3645990.385095
Two branches with different periods [48, 96]0.362860.383284
Single period 48 with an approximately matched parameter budget0.3788940.395548
Duplicate period-48 branches with an approximately matched parameter budget0.3639650.384708

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.

Reviewer_CyqdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

To capture temporal dependencies effectively, this paper proposed a lightweight forecasting model, PKTime, combining Kronecker-structured modeling with adaptive period selection. The proposed Kronecker parameterization of the historical-to-future mapping captured both within-period interactions and cross-period dependencies. Meanwhile, this component unified parameter compression and temporal modeling within a single forecasting operator. Furthermore, experiments on eight benchmarks were conducted to validate the proposed method.

Critical strengths
  1. PKTime models temporal dependencies using periodic structures via period-specific representation on intra-period and inter-period relationships. The presented Kronecker-structured modeling represents temporal dependencies from intra-period positions and inter-period sequences.

  2. In sub-section 4.2, the author provided the computation cost analysis of PKTime to show the batch-level computational complexity. Moreover, sub-section 5.4 makes a runtime efficiency analysis. Both demonstrate the author paid attention to the efficiency of the proposed PKTime, which is essential for practical applications.

  3. As different periods provide distinct structured forecasting views of the same historical sequence, this paper adopted multi-period fusion using weighted combination and the fusion policy were jointly optimized through the forecasting objective.

Critical weaknesses
  1. Adaptive period selection depend hard on the training data, especially the periods and branch count. It means for each dataset, the related periods and branch count will be estimated once.

  2. Lack of description about Zp,cZ_{p,c}, which was used in Equation 6 and 7. The μc\mu_{c} occurred only once in Equation 5. What is the meaning of dpd_p ? In this case, there are many notations without clear description.

  3. The performance differences between the proposed PKTime and baselines are on the 3rd effective position after the dot. The author should provide significant testing on such results.

Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer Cyqd

W1: Thank you for your suggestion. We agree that, for a given training set and model configuration, the period set and number of branches are determined once before model fitting and remain fixed throughout training and prediction. Period detection uses only the training segment used for parameter optimization. If the model is refitted using a different training interval, selection is repeated using the new training data.

Here, “adaptive” refers to selecting the model structure based on the training data. We first extract candidate periods from the training series, then determine which periods to retain through spectral tests, merging nearby periods, filtering by relative energy, and enforcing a parameter budget. Each retained period corresponds to a prediction branch. Thus, the period lengths and number of branches are jointly determined by the same set of selection rules. Once the structure is fixed, each branch’s intra-period and cross-cycle mappings, together with the fusion weights, are learned through the forecasting loss. This concentrates the computation for period discovery before model fitting, while prediction directly uses the established structure.

To address your concern about dependence on the training data, we additionally examined sensitivity to the training interval. Specifically, we performed period selection using the first 50%, middle 25%–75%, last 50%, and entirety of the training segment used for parameter optimization. The table below reports all results, with the number of branches in parentheses. Period lengths are measured in sampling steps.

DatasetFirst 50%Middle 25%–75%Last 50%Full training segment
ETTh1[24] (1)[24, 12] (2)[24, 12] (2)[24] (1)
ETTh2[24] (1)[24] (1)[24] (1)[24] (1)
ETTm1[96] (1)[96, 48] (2)[96, 48] (2)[96, 48] (2)
ETTm2[96] (1)[96] (1)[96] (1)[96] (1)
Electricity[24, 12] (2)[24, 12] (2)[24, 12] (2)[24] (1)
Traffic[24] (1)[24] (1)[24] (1)[24] (1)
Solar[143, 72] (2)[143, 72] (2)[145, 72] (2)[144] (1)
Exchange[332, 221, 190] (3)[266, 295] (2)[332, 241, 266] (3)[312, 280, 253] (3)

For each training interval, the selections above were identical across forecast horizons of 96, 192, 336, and 720. ETTh2, ETTm2, and Traffic retain the same structure across all four training intervals. For ETTh1, ETTm1, and Electricity, the primary period remains unchanged, while additional branches vary with the training interval. Solar’s primary period ranges from 143 to 145, and the period-72 branch is retained in all three subintervals. Exchange shows more pronounced changes in both the period set and the number of branches, indicating that its selection results are more sensitive to the training interval. These results illustrate how period selection depends on the training data: different training intervals can affect both the candidate period values and whether additional branches are retained. This design allows the model to configure prediction branches according to the time scales identified in the training data while controlling the associated parameter overhead. We will explicitly state in the method section that “the periods and number of branches are selected once before each model fit, after which the structure remains fixed.” We will also include the analysis above to clarify the meaning and implementation of “adaptive.”

W2: Thank you for pointing out the unclear definitions in the manuscript. We will clarify these definitions and incorporate them into the manuscript. Specifically, μc\mu_c is the mean of the current input window for variable cc, Zp,cZ_{p,c} is the input after centering and folding by period, and dpd_p contains the learnable diagonal coefficients of the intra-period mapping.

Let X∈RL×CX\in\mathbb R^{L\times C} denote the input window standardized using training set statistics. For variable cc,

\widetilde x_{t,c}=X_{t,c}-\mu_c.$$ The standardizer remains fixed after being fitted on the training set, while $\mu_c$ is updated based on the currently observed input window. Let $N_p=\lceil L/p\rceil$ and $M_p=\lceil H/p\rceil$. We pad the centered input with zeros on the right and define $$(Z_{p,c})(i,n) = \tilde{x}_{np+i,c} \quad \text{if}\quad np+i < L\quad\text{else}\quad 0$$ for $0\le i<p$ and $0\le n<N_p$. **W3**: Thank you for your question about siginificancy test. 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 in the revised manuscript.
Reviewer_rSoGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PKTime, a lightweight time-series forecasting model. It selects periods from training data and folds the history into within-period and cross-period dimensions. Each branch combines a diagonal-plus-low-rank within-period map with a dense cross-period map, and forecasts are combined through globally learned weights. Experiments on eight datasets, structural ablations, rank sensitivity analysis, and runtime measurements evaluate forecasting performance and efficiency.

Critical strengths
  1. The motivation is clear and the design follows it coherently. The paper explains both the cost of dense forecasting maps and the sharing restrictions of a single separable map, then develops a compact formulation with explicit roles for within-period and cross-period modeling.
  2. PKTime offers a promising compact forecasting model. It uses only 241–1,021 parameters on Electricity while remaining competitive in Table 1, and Table 3 provides direct evidence that the low-rank within-period interactions are useful.
Critical weaknesses
  1. The configurations and full-model results across experiments need to be reconciled. Averaging the four ETTm2 horizons in Table 1 gives MSE/MAE of 0.25275/0.31525, whereas both Full in Table 3 and the default r=2 setting in Table 5 give 0.266/0.328. Clarifying the configurations, seeds, and aggregation procedures is necessary to establish whether the main results and ablations describe the same setting.
  2. The manuscript states, “At the longest horizon, PKTime reduces the parameter count by 99.902%, 99.503%, and 96.849% compared with DLinear, FITS, and SparseTSF, respectively.” However, the abstract of SparseTSF states that “the SparseTSF model uses fewer than 1k parameters.” The authors should explain in more detail how the parameter counts were calculated and compared.
Questions for authors
  1. Which ranks, period sets, batch sizes, and seed aggregation procedures were used in Tables 1, 3, and 5? Please reconcile the ETTm2 full-model averages and clarify whether the training batch size was 128 or 32.
2 comments
AuthorsOpenReview ↗

Response to reviewer rSoG

W1: Thank you for pointing out this discrepancy. We have rechecked the data reported in the manuscript. The ETTm2 results in Table 1 use rank 8, whereas the Full configuration in Table 3 and the default configuration in Table 5 use rank 2. We apologize for this inconsistency and will correct it in the revised manuscript.

W2: Thank you for your question about how model parameters are counted. Specifically, we use the following code:

parmas = sum(p.numel() for p in model.parameters())

For PKTime, the parameter count is P=K+∑p∈P[p+2pr+MpNp].P = K + \sum_{p\in\mathcal{P}}\left[p+2pr+M_pN_p\right].

Here, KK is the number of fusion logits, Np=⌈L/p⌉N_p=\lceil L/p\rceil, and Mp=⌈H/p⌉M_p=\lceil H/p\rceil. The term p+2prp+2pr counts the parameters of Dp=diag⁡(dp)+UpVp⊤D_p=\operatorname{diag}(d_p)+U_pV_p^\top, while MpNpM_pN_p counts the parameters of WpW_p.

For SparseTSF, the parameter count is P=1+2⌊p2⌋+MpNp,P = 1+2\left\lfloor\frac{p}{2}\right\rfloor+M_pN_p,

where pp is the period length, Np=L/pN_p=L/p, and Mp=H/pM_p=H/p.

Using these formulas and the TFB/OpenTS implementations, we obtain the following parameter counts for PKTime and SparseTSF. Under these settings, SparseTSF has substantially more than 1K parameters.

Q1: Thank you foir your question about training details. We use rank 8 for ETTm2 and rank 2 for all other datasets. The period sets match those reported in Table 7 of the manuscript. We use three random seeds (2021, 2022, and 2023) and report the arithmetic mean of the results across these seeds. The training batch size is 128. We will update the revised manuscript with the corrected results.

Thank you for the clarification. The different rank settings explain the discrepancy between the ETTm2 results in the main and ablation tables, and the training batch size is now clear. Please state these settings explicitly in the revised manuscript. Regarding SparseTSF, the counting formula is helpful, but the actual parameter counts mentioned in the response appear to be missing. I maintain my weak acceptance rating.

Reviewer_RvisOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PKTime, a lightweight time-series forecasting model that combines training-data-driven period selection with Kronecker-structured temporal mappings. Each branch folds the input sequence according to a selected period and models intra-period interactions using a diagonal-plus-low-rank matrix and inter-period dependencies using a dense matrix. Branch predictions are combined through learned global weights. Experiments on eight datasets report competitive forecasting accuracy with compact parameter counts, supported by structural ablations and batched inference measurements.

Critical strengths
  1. The model provides a useful accuracy–efficiency tradeoff through a clear structural design. Assigning separate mappings to intra-period and inter-period relationships gives an understandable inductive bias and substantially reduces parameter counts. On Electricity, the reported model uses approximately 241–1,021 parameters across forecasting horizons while maintaining competitive accuracy. The runtime measurements on Electricity and Traffic provide additional evidence that the compact parameterization can translate into faster batched inference.

  2. The experiments provide meaningful support for the main architectural choices. PKTime achieves the best reported MSE on 15 of 32 tasks and MAE on 22, including ties. The ablations show consistent benefits from intra-period low-rank interactions. On ETTm1, using multiple distinct periods also outperforms both a single-period model and multiple branches with the same period, helping distinguish the benefit of period diversity from simply adding branches.

Critical weaknesses
  1. The incremental value of the period-selection procedure is not sufficiently isolated. Six of the eight datasets retain only one period, and the detailed period-diversity ablation is limited to ETTm1. While this supports multiple periodic views in that setting, it does not establish how much the proposed selection procedure improves over a fixed natural period or a simpler spectral selection rule. Similarly, uniform fusion performs almost identically to learned fusion on ETTm1. These results leave some uncertainty about which components are necessary for the overall gains.

  2. The empirical advantage needs better uncertainty reporting and a clearer evaluation protocol. Several improvements over the strongest baselines are around 1%, but Table 1 provides no variability estimates. Although the paper states that baselines are reproduced using TFB, the baseline configuration and tuning details are insufficient to fully assess comparability. The efficiency evidence is also limited to batched inference on two datasets, which supports that setting but leaves the broader resource-constrained deployment motivation less thoroughly validated.

Questions for authors
  1. How does the proposed period selection compare with a fixed natural period and a simple top-k spectral selection rule under comparable parameter budgets? Results on ETTm1 and Exchange would be particularly informative because these are the two datasets that retain multiple periods.

  2. Could you provide mean and standard deviation across training seeds for the main comparisons, particularly where the reported gains are small? Please also clarify the input lengths and hyperparameter-selection procedures used for the baselines.

  3. What batch size was used for training? Appendix A.1 specifies 128, whereas Appendix A.2 specifies 32. For the inference measurements, please clarify the timing protocol, including warm-up, device synchronization, and which operations are included.

3 comments
AuthorsOpenReview ↗

Response to reviewer Rvis

W1: Thank you for pointing out the need for more direct experimental comparisons of period selection and fusion strategies. We have added experiments with fixed periods, simple spectral selection, and alternative fusion strategies, and conducted comparisons on Exchange under similar parameter budgets.

The following experiments use an input length of 720, a batch size of 128, a learning rate of 0.03, and three training seeds (2021, 2022, and 2023). The reported values are first averaged across the three seeds for each forecast horizon, then averaged across the four horizons of 96, 192, 336, and 720.

On ETTm1, our method selects periods 96 and 48. This combination achieves lower MSE and MAE than using either period 48 or period 96 alone:

Selection strategyPeriod setMSEMAE
Fixed period 48[48]0.3788690.395806
Fixed period 96 / spectral top-1[96]0.4482290.4412
Spectral top-2 / our method[96, 48]0.3624970.38327

Spectral top-2 and our method identify the same period combination on ETTm1. Compared with fixed period 48, this combination reduces the average MSE by approximately 4.32%.

On Exchange, we further compare different selection strategies. By adjusting the ranks of the comparison models, we keep the parameter count difference at each forecast horizon within 10% of our method’s parameter count:

Selection strategyPeriod setParameter count rangeMSEMAE
Fixed period 96, adjusted rank[96]4145–43290.3929390.413357
Spectral top-1, adjusted rank[312]4060–40660.388140.410734
Spectral top-3[312, 280, 332]4632–46500.3650450.401147
Our method[312, 280, 253]4237–42550.3597820.399773

With similar parameter counts, our method reduces the average MSE by approximately 8.44% and 7.31% relative to fixed period 96 and spectral top-1, respectively. Compared with spectral top-3, which also uses three branches, our method selects a different period combination, reducing the average MSE by approximately 1.44% while also reducing the parameter count by approximately 8.5%. This comparison directly demonstrates the role of period selection in forecasting accuracy and model size.

Our method uses learned fusion: one trainable scalar per branch is converted into a global weight through softmax, and these scalars are optimized jointly with the branch parameters:

αk=exp⁡(ak)∑j=1Kexp⁡(aj).\alpha_k=\frac{\exp(a_k)}{\sum_{j=1}^{K} \exp(a_j)}.

This design allows the forecasting loss to directly determine how predictions from different periods are combined. Fusion adds only two and three trainable scalars for ETTm1 and Exchange, respectively. The comparisons with the period sets held fixed are as follows:

DatasetOur learned fusion MSE / MAEUniform fusion MSE / MAESpectral energy weighting MSE / MAE
ETTm10.362497 / 0.3832700.362812 / 0.3836530.362586 / 0.383400
Exchange0.359782 / 0.3997730.361172 / 0.4000170.360657 / 0.399811

Learned fusion achieves the lowest MSE and MAE averaged across the four forecast horizons on both datasets. Relative to uniform fusion, it reduces the average MSE by approximately 0.09% and 0.38%, respectively. Using just one scalar per branch, the model can further optimize the combination of predictions from the selected periods.

Six datasets retain a single period, while ETTm1 and Exchange retain two and three periods, respectively. This reflects our design of determining period lengths and the number of branches from the training data. These results support constructing prediction branches from a compact set of periods and combining their predictions through a small number of learnable weights: ETTm1 benefits from an effective two-period representation; on Exchange, period selection further improves accuracy and reduces the parameter count; and learned fusion reduces the average forecasting error with very little parameter overhead.

AuthorsOpenReview ↗

Response to reviewer Rvis

W2&Q2&Q3: Thank you for your suggestion. 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.

Dataset96192336720
ETTh10.358 ± 0.002 / 0.388 ± 0.0010.399 ± 0.002 / 0.411 ± 0.0010.427 ± 0.002 / 0.426 ± 0.0010.424 ± 0.002 / 0.446 ± 0.002
ETTh20.280 ± 0.003 / 0.339 ± 0.0020.341 ± 0.006 / 0.377 ± 0.0020.364 ± 0.002 / 0.399 ± 0.0020.385 ± 0.003 / 0.424 ± 0.001
ETTm10.317 ± 0.002 / 0.357 ± 0.0020.342 ± 0.003 / 0.372 ± 0.0030.371 ± 0.002 / 0.388 ± 0.0020.420 ± 0.003 / 0.416 ± 0.001
ETTm20.179 ± 0.002 / 0.273 ± 0.0020.234 ± 0.001 / 0.308 ± 0.0010.287 ± 0.001 / 0.342 ± 0.0010.371 ± 0.003 / 0.394 ± 0.001
Electricity0.139 ± 0.000 / 0.234 ± 0.0010.152 ± 0.000 / 0.245 ± 0.0010.167 ± 0.000 / 0.261 ± 0.0000.206 ± 0.000 / 0.295 ± 0.001
Traffic0.393 ± 0.000 / 0.268 ± 0.0010.402 ± 0.000 / 0.272 ± 0.0000.415 ± 0.001 / 0.280 ± 0.0010.452 ± 0.001 / 0.299 ± 0.001
Solar0.197 ± 0.002 / 0.237 ± 0.0030.243 ± 0.065 / 0.267 ± 0.0540.230 ± 0.005 / 0.250 ± 0.0040.235 ± 0.005 / 0.250 ± 0.002
Exchange0.084 ± 0.001 / 0.202 ± 0.0010.171 ± 0.001 / 0.294 ± 0.0020.318 ± 0.001 / 0.407 ± 0.0020.866 ± 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.

The training batch size used in the main experiments was 128. The description in Appendix A.1 is correct; the value of 32 in Appendix A.2 is a typo. We will correct this typo and make the related configuration descriptions consistent.

The inference efficiency comparison measures model forward-pass latency with the input already resident on the target device. The default measurement procedure is as follows:

  1. Model state and warm-up. We set the model to eval() mode and execute it under torch.inference_mode(). Each configuration undergoes 50 warm-up iterations, followed by CUDA synchronization before timing begins.

  2. Repeated measurements and device synchronization. We perform five rounds of 200 measurements each. For each measurement, we first call torch.cuda.synchronize(), start timing with time.perf_counter_ns(), execute one forward pass, synchronize again, and record the end time. This measures the wall-clock time required to complete the forward pass, including the host-side invocation and final synchronization overhead.

  3. Aggregation. We compute the median of the 200 latency measurements within each round, then take the median of the five round medians as latency_median_ms. Figure 3 uses an inference batch size of 128, so this metric is expressed in milliseconds per batch. Results across forecast horizons are aggregated using an equally weighted average over 96, 192, 336, and 720.

The timed region includes centering, branch mappings, fusion, and mean restoration within the forward pass, as well as any FFTs, reconstruction, memory allocation, and copying performed by each model within that pass. It also includes the host-side overhead of invoking the forward pass and waiting for CUDA operations to complete.

We use randomly initialized models and random FP32 tensors generated with the corresponding input shapes to measure the execution cost of a fixed model structure. Model construction, period selection, and input preparation are completed before timing begins.

Throughput is measured using a separate set of loops. Each block executes 20 consecutive forward passes, with synchronization only at block boundaries, and throughput is calculated from the block’s elapsed time. The latency reported in Figure 3 is the latency_median_ms obtained from the measurements with synchronization before and after each forward pass, as described above.

We will expand the efficiency experiment description to specify the number of warm-up iterations, synchronization points, aggregation procedure, and operations included in the timed region. We will also explicitly distinguish the training batch size from the inference batch size.

AuthorsOpenReview ↗

Response to reviewer Rvis

Thank you for your suggestion. We have added comparisons with fixed periods and simple spectral top-k selection on ETTm1 and Exchange, adjusting the ranks of the comparison models to bring their parameter counts close to that of our model. All methods use the same forecasting architecture, learned fusion, and training settings: an input length of 720, a batch size of 128, a learning rate of 0.03, and three training seeds.

On ETTm1, we use the natural daily period of 96 and additionally consider the half-day period of 48. The spectral baselines include top-1 and top-2, with top-2 using the same number of branches as our method:

Selection strategyPeriod setParameter count rangeMaximum parameter deviationMSEMAE
Fixed half-day period, adjusted rank[48]751–10423.84%0.3788940.395548
Fixed daily period / spectral top-1, adjusted rank†[96]681–92910.40%0.4057210.419591
Spectral top-2[96, 48]760–10110%0.3624970.38327
Our method[96, 48]760–1011—0.3624970.38327

The parameter deviation for fixed period 96 is approximately 10.4% at forecast horizons of 96 and 192, slightly exceeding the predefined 10% tolerance. At horizons of 336 and 720, the deviations are 4.42% and 8.11%, respectively. The table reports the actual parameter counts after rank adjustment and the results averaged across all four forecast horizons.

Our method and spectral top-2 both select [96, 48] and have identical parameter counts and forecasting results. Compared with the single-period 48 baseline, which meets the budget tolerance at all four forecast horizons, our method reduces the average MSE by 4.33% and MAE by 3.10%. This shows that, with similar parameter counts, allocating parameters to two period branches achieves lower average errors than using a single half-day period branch.

On Exchange, we use a predefined fixed period of 96 and compare against spectral top-1 and top-3, with top-3 using the same number of branches as our method. All the following baselines have parameter deviations below 10% at each of the four forecast horizons:

Selection strategyPeriod setParameter count rangeMaximum parameter deviationMSEMAE
Fixed period 96, adjusted rank[96]4145–43292.17%0.3929390.413357
Spectral top-1, adjusted rank[312]4060–40664.44%0.388140.410734
Spectral top-3[312, 280, 332]4632–46509.32%0.3650450.401147
Our method[312, 280, 253]4237–4255—0.3597820.399773

Our method reduces the average MSE by 8.44%, 7.31%, and 1.44% relative to fixed period 96, spectral top-1, and spectral top-3, respectively. The comparison with top-3 keeps the number of branches unchanged: our method replaces the third period, 332, with 253, reducing the average error while also decreasing the parameter count by approximately 8.5%.

At individual forecast horizons, our method achieves lower errors than top-3 at 192, 336, and 720, but slightly higher errors at 96. Relative to fixed period 96, the average improvement mainly comes from the 720-step horizon, where MSE decreases from 1.009070 to 0.866241; the fixed-period baseline achieves lower errors at the first three horizons.

These results demonstrate the specific effects of period selection on the two datasets. On ETTm1, our method identifies the same two-period combination as spectral top-2 and outperforms the single-period 48 baseline under a similar parameter budget. On Exchange, further refinement of the period combination achieves lower average forecasting errors while using fewer parameters than top-3 with the same number of branches.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Responses and Revisions

We thank the reviewers for their careful reading and helpful suggestions. Our detailed responses provide additional explanations, controlled comparisons, and experimental settings. We summarize the main points below.

1. Relationship to SparseTSF and the role of the intra-period map.
PKTime extends a shared cross-cycle prediction head with a diagonal-plus-low-rank intra-period map, allowing information exchange between phases. We explain this design through its parameter sharing across phase pairs and cycle pairs. We also provide an argument at the prediction-head level: for arbitrary folded inputs and a nonzero cross-cycle matrix, a phase-dependent intra-period map generally cannot be absorbed into the shared cross-cycle matrix. The revision will make this connection explicit and explain the contribution through phase interactions and compact parameterization.

2. Separate comparisons of periodic representations, selection rules, and fusion.
The added experiments compare validation-selected single periods, duplicate-period branches, distinct-period branches, fixed periods, and spectral top-k selection. On ETTm1, the multi-period configuration reduces four-horizon average MSE by approximately 4.23% relative to the validation-selected single-period reference. In the selection-rule comparison, our method and top-2 both select [96,48]. On Exchange, our method reduces average MSE by approximately 1.44% and parameter count by approximately 8.5% relative to top-3 with the same branch count. We separately report learned, uniform, and spectral-energy fusion, and examine sensitivity to the training interval. Here, “adaptive” means selecting periods and branch count from training data before each fit, with the structure fixed thereafter.

3. Cross-cycle parameterization, rank, and configuration consistency.
The added results show the tradeoffs between dense and factorized W: on ETTh1 and ETTm1, factorized W has lower MSE while dense W has lower MAE; on ETTh2 and ETTm2, dense W has lower errors on both metrics. Increasing rank from 2 to 8 on ETTm2 reduces average MSE by approximately 3.65%, adding 1,152 parameters. We clarified that the original ETTm2 main results use rank 8, whereas the default ablation uses rank 2. These settings will be stated separately in the experimental description and table notes.

4. Repeated-run results and evaluation protocol.
Our responses include the complete mean and sample standard deviation table across eight datasets and four horizons, and clarify that the original baseline scores come from the public OpenTS/TFB evaluation. The main training batch size is 128; the value of 32 in Appendix A.2 will be corrected. PKTime’s tuning experiments compare six combinations of two input lengths and three learning rates, selecting configurations by validation error. The revision will list experiment-specific seeds, configurations, and sources, add cited comparisons with PatchTST, iTransformer, and CycleNet, and provide per-comparison statistical details with test definitions and multiple-comparison corrections. We will reconcile and correct the seed labels and aggregate significance statement in the responses against the corresponding records.

5. Model size and inference efficiency.
We clarified the original timing protocol: device-resident inputs, 50 warm-up iterations, and five rounds of 200 measurements with CUDA synchronization before and after each forward pass. On Electricity, with input length 720 and inference batch size 128, average latency is 1.762 ms/batch for PKTime and 5.533 ms/batch for MixLinear. MixLinear has fewer parameters, while PKTime has lower batched forward latency under these measurement conditions. The revision will report the parameter configurations, latency, and timing scope together.

Area_Chair_Bf2uOpenReview ↗

check authors' response

Dear reviewers,

The authors have posted their responses. Please read them and update your scores if they address your concerns. If they don't change your view, a short note on what remains unresolved would help the discussion.

Thank you for your time.
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.