acceptodds
Under review as a conference paper at ICLR 2027

TSComp: Cross-scale Compositional Pre-training for Time Series Foundation Models

Abstract

Time series foundation models (TSFMs) aim to learn transferable temporal representations across diverse downstream tasks. However, time series are inherently multi-scale: the observed signal is jointly composed of distinct temporal structures at different scales, each capturing different temporal dynamics. This poses two challenges for transferable pre-training: scale-specific structures should be explicitly distinguished and their cross-scale correspondence must also be preserved because their roles depend on how they interact across scales. Conventional time-domain masked reconstruction does not explicitly enforce either property, leaving both scale-specific structures and cross-scale correspondence implicitly supervised. To address these limitations, we propose TSComp, a cross-scale compositional pre-training framework for TSFMs. TSComp uses scale-balanced tokenization to construct aligned scale-specific representations, scale coefficient attention to model interactions among corresponding structures across scales, and multi-scale directional masking to infer missing components from complementary cross-scale context. Its pre-training objectives jointly supervise scale energy distribution calibration, signal recomposition, and coefficient reconstruction, making scale-specific recovery and cross-scale composition explicit learning targets. The resulting scale-specific representations are adapted to downstream tasks through a scale-aware adaptation method. Experiments across forecasting, imputation, classification, and anomaly detection demonstrate effective transfer across diverse downstream settings and competitive performance against both multi-task foundation models and task-specific architectures. Our code is available at https://anonymous.4open.science/r/TSComp-66B0.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time-series Foundation Models, Multi-scale Representation Learning, Cross-scale Composition

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 3 reviews, above 27% of the others.
Reviewer_cTe8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TSComp, a time-series foundation model that explicitly decomposes signals into wavelet scales, models temporally aligned cross-scale interactions, and pre-trains with directional masking plus energy, time-domain, and coefficient reconstruction objectives. The learned representations are transferred through task-specific scale-aware pipelines for forecasting, imputation, classification, and anomaly detection. Results across a large pre-training corpus and multiple downstream benchmarks are competitive or state of the art on several metrics.

Critical strengths
  1. The paper identifies a concrete limitation of time-domain reconstruction and connects each architectural component to scale separation, temporal alignment, or cross-scale composition.

  2. TSComp is tested on four distinct task families, including 27 classification datasets and 245 anomaly datasets, rather than being optimized for forecasting alone.

  3. The scale-wise reconstruction study, cross-scale correspondence probe, component ablations, and scale-aware adaptation comparison provide evidence beyond aggregate task scores.

Critical weaknesses
  1. The method uses Symlet-4 and a fixed scale hierarchy throughout. The paper acknowledges this limitation, but does not provide enough sensitivity analysis across wavelet families, decomposition depths, nonstationary settings, or datasets with irregular sampling.

  2. TSComp changes the pre-training representation, backbone, objectives, and downstream heads at the same time. Since the proposed scale-aware task pipelines themselves improve performance, the comparison does not establish how much of the gain comes from cross-scale pre-training rather than task-specific architectural adaptation. Compute, parameter, and pre-training-data matching against all foundation-model baselines is also not fully clear.

  3. Removing coefficient reconstruction or scale-energy calibration improves forecasting in the reported ablation, TSComp-base is worse than TSComp-small for classification, and GTM remains better on two anomaly metrics. The near-perfect reconstruction and correspondence-probe scores also deserve scrutiny because they may partly reflect an easy diagnostic aligned with TSComp's training objective.

Questions for authors
  1. Can the authors provide a controlled comparison using the same backbone and downstream head, varying only the pre-training objective, and report matched parameter counts and pre-training compute for the main baselines?

  2. How sensitive are the results to wavelet family, decomposition depth, patch alignment, and the magnitude or type of cross-scale misalignment used by the probe?

6 comments

Dear Reviewer cTe8,

We thank the reviewer for their careful evaluation and constructive feedback. In response to their comments, we have (1) clarified our wavelet selection rationale and examined sensitivity to decomposition choices and patch alignment (Table 1; Appendix E), (2) conducted additional robustness evaluations under non-stationary perturbations and simulated irregular sampling (Tables 2–3), (3) introduced a capacity-matched MOMENT baseline and additional pre-training objective ablations to examine the contribution of cross-scale compositional learning (Tables 4–6), (4) strengthened the diagnostic evaluation with additional probing controls and misalignment sensitivity analyses (Tables 7–8), and (5) clarified the interpretation of task-dependent ablation results and model scaling behavior. We will incorporate these additional experiments and clarifications into the revised manuscript. Detailed responses are provided below.


Sensitivity on Wavelet Decomposition and Analysis on Challenging Settings

(W1-1) The method uses Symlet-4 and a fixed scale hierarchy throughout

We appreciate the reviewer for highlighting the importance of sensitivity analysis. We emphasize that the primary contribution of TSComp lies in its pre-training framework for explicitly learning scale-specific temporal structures and their cross-scale compositional relationships, not in selecting a particular wavelet basis.

We adopt Symlet-4 with a three-level decomposition as the default configuration based on its structural compatibility with our pre-training design. Specifically, Symlet-4 combines orthonormality, compact support, near symmetry, and four vanishing moments. Orthonormality ensures energy preservation and supports our theoretical formulation of scale-wise reconstruction and scale-energy distribution calibration, while compact support provides temporal localization for modeling scale-specific structures and their cross-scale correspondence. Thus, our choice of Symlet-4 is motivated by its suitability for the proposed learning objectives, without assuming that it achieves optimal performance on every downstream task.

(W1-2) does not provide enough sensitivity analysis across wavelet families, decomposition depths, nonstationary settings, or datasets with irregular sampling.
(Q2-1) How sensitive are the results to wavelet family, decomposition depth, patch alignment ...?

We investigate the sensitivity of TSComp to wavelet bases (Appendix E.3, Table 23), decomposition depths (Appendix E.2, Figure 6), and alternative decomposition strategies (Appendix E.1, Table 22).

Wavelet basis (Appendix E.3): Table 23 compares Symlet-4, Bior 4.4, Db4, and Haar across four downstream tasks. Symlet-4 achieves the best forecasting performance on ETTh1, whereas Bior 4.4, Db4, and Haar perform best in imputation, classification, and anomaly detection, respectively. These results indicate that no single wavelet basis is optimal across all tasks. As discussed above, our choice of Symlet-4 is motivated by its structural compatibility with cross-scale compositional pre-training, not by task-specific performance optimization.

Decomposition depth (Appendix E.2): Figure 6 evaluates decomposition levels from 1 to 3. The results show that the three-level decomposition achieves the lowest forecasting and imputation MSE on ETTh1, supporting the benefit of representing temporal structures at multiple resolutions.

Decomposition strategy (Appendix E.1): Table 22 compares orthonormal DWT with Fourier decomposition, learnable wavelets, and residual pyramids. Orthonormal DWT achieves the best forecasting and imputation performance on ETTh1, while learnable wavelets yield slightly higher classification accuracy. Although orthonormal DWT is not optimal for every task, our choice prioritizes the four structural requirements for cross-scale compositional pre-training—temporal localization, an explicit scale hierarchy, perfect recomposition, and energy preservation—while maintaining competitive performance across downstream tasks.

In response to the reviewer's comment, we have extended our sensitivity analysis to include patch alignment, nonstationary settings, and irregularly sampled datasets.

Patch alignment: Time-aligned patching preserves temporal correspondence across wavelet scales, enabling SCA to model interactions among scale-specific structures describing the same temporal region and MDM to encourage reconstruction from complementary cross-scale context. We compare TSComp with TSComp-NoAlign, which flattens coefficient patches into a single token sequence and applies conventional self-attention without explicitly enforcing temporal alignment in cross-scale interactions. As reported in Table 1, TSComp achieves lower MSE across all four evaluated settings, supporting the effectiveness of explicitly modeling cross-scale relationships among aligned coefficient representations

Table 1. Comparison of TSComp with TSComp-NoAlign (MSE).

ModelForecasting ETTm1Forecasting WeatherImputation ETTh1Imputation ETTm1
TSComp0.36090.23400.18130.0849
TSComp-NoAlign0.36510.23410.19200.0854

Robustness to Non-stationary Conditions: We investigate TSComp's forecasting performance under non-stationary perturbations, including a comparison with TSComp-Fourier, a variant that replaces wavelet decomposition with Fourier decomposition. We conduct 96-step forecasting on the Exchange dataset, selected because its exchange-rate series exhibit inherently non-stationary temporal dynamics. We introduce three representative perturbations: Trend, which introduces gradual temporal drift; Level Shift, which introduces abrupt changes in the signal mean; and Frequency Regime, which introduces changes in oscillatory frequencies across temporal segments. Each perturbation is applied at three severity levels (L1–L3) to the raw time series before standardization, with normalization statistics computed exclusively from the perturbed training split.

Across all nine evaluated conditions in Table 2, TSComp-small achieves the lowest forecasting MSE, consistently outperforming TSComp-Fourier and the other evaluated baselines. Wavelet decomposition does not require a stationarity assumption and provides temporally localized, scale-specific representations, enabling TSComp to model cross-scale relationships through SCA as temporal dynamics change. In particular, the improvements over TSComp-Fourier suggest that temporally localized wavelet representations are beneficial for cross-scale compositional learning under non-stationary conditions. These findings support TSComp's robustness to the evaluated changes in temporal dynamics.

Table 2. Forecasting performance (MSE) under three types of non-stationary perturbations at three severity levels (L1–L3) on the Exchange dataset.

ConditionLevelTSCompTSComp-FourierMOMENTTSPulse
TrendL10.1020.1350.1550.144
L20.0680.0910.1080.097
L30.0250.0330.0400.036
Level ShiftL10.1360.1690.2060.172
L20.1580.1980.2510.200
L30.1710.2180.2700.221
Frequency RegimeL10.1530.1800.1940.216
L20.2290.2600.2700.328
L30.3290.3720.3780.470

Robustness to Irregular Sampling: We investigate TSComp's robustness to irregular sampling using six forecasting benchmarks (ETTh1, ETTh2, ETTm1, ETTm2, Weather, and Exchange), selected for a computationally efficient sensitivity analysis. We use a 512-step lookback window and four prediction horizons (96, 192, 336, and 720). To simulate irregular sampling, we randomly remove observations at missing ratios of 10%, 30%, and 50%, with all channels sharing the same missing timestamps. Missing values are replaced with zeros, while forecasting targets remain fully observed. We report the average MSE across the six datasets and four prediction horizons.

As shown in the Table 3, TSComp-small achieves the lowest average MSE under fully observed conditions and at missing ratios of 10% and 30%, outperforming all compared baselines. At 50% missingness, FreTS achieves better performance, while TSComp-small continues to outperform SimpleTM and TimeMixer. These results suggest that TSComp's multi-scale representations remain effective under moderate observation loss, although performance deteriorates when observations become substantially sparse. Since wavelet decomposition operates on a regular time grid, extensive missingness may distort scale-specific temporal structures and their cross-scale relationships, limiting the effectiveness of the proposed framework under severe observation sparsity.

Table 3. Forecasting performance (MSE) under simulated irregular sampling.

Model0%10%30%50%
TSComp-small0.3470.3440.4320.644
FreTS0.4340.4420.4610.474
SimpleTM0.4080.4280.5670.812
TimeMixer0.3560.4120.5350.776

Effect of Cross-scale Pre-training

(W2-1) the comparison does not establish how much of the gain comes from cross-scale pre-training rather than task-specific architectural adaptation. 

To examine the contribution of cross-scale pre-training independently of the proposed scale-aware downstream adaptation, we compare TSComp without scale-aware adaptation (w/o Scale) against a capacity-matched MOMENT model pre-trained on the same corpus using conventional time-domain masked reconstruction. By excluding TSComp's task-specific scale-aware adaptation and controlling for pre-training data and model size, this comparison provides a more direct assessment of the benefits of cross-scale compositional pre-training over conventional time-domain reconstruction-based pre-training. As shown in the table 4, TSComp w/o Scale consistently outperforms the capacity-matched MOMENT across all three evaluated tasks. These improvements are achieved without the proposed scale-aware downstream adaptation, supporting the contribution of TSComp's cross-scale compositional pre-training framework to transferable representation learning.

Table 4. Comparison of cross-scale compositional pre-training with a capacity-matched time-domain pre-training baseline

VariantETTh1 Imp. MSEClassification Avg. Acc.Anomaly VUSPR
Capacity-matched MOMENT0.302447.670.0266
TSComp w/o Scale0.188567.000.0343

Effect of Pre-training Objectives

(Q1-1) Can the authors provide a controlled comparison using the same backbone and downstream head, varying only the pre-training objective,

Table 3(c) compares variants trained with different combinations of pre-training objectives while keeping the backbone architecture and downstream adaptation settings unchanged. The results show that the effects of individual objectives vary across downstream tasks, with some ablated variants outperforming the complete configuration in specific settings.

TSComp's pre-training objectives are designed to jointly capture scale-specific temporal structures and their cross-scale compositional relationships within a unified framework, rather than to maximize performance on individual downstream tasks. To assess whether the proposed pre-training objectives achieve these intended representational properties, we conduct additional diagnostic ablations using the same backbone and probing protocol (Table 5). The complete configuration achieves the strongest combination of scale-specific reconstruction fidelity and cross-scale correspondence discriminability. While using only LTIMEL_{\mathrm{TIME}} preserves high scale-wise reconstruction fidelity, it yields lower correspondence probe performance. Conversely, using only LCOEFL_{\mathrm{COEF}} leads to weaker fine-scale reconstruction (D1 correlation drops from 0.999 to 0.665). These findings suggest that the objectives provide complementary learning signals for capturing scale-specific temporal structures and their cross-scale relationships, although their contributions to downstream performance remain task-dependent.

Table 5. Diagnostic analysis of pre-training objective ablations. Reconstruction fidelity is measured by Pearson correlation in the time domain and at individual wavelet scales.

VariantTimeA3D3D2D1B.AccAUROCF1
TSComp0.9991.0001.0000.9990.9990.9991.0000.999
w/o LTIMEL_{\mathrm{TIME}}0.9950.9940.9980.9980.9980.9901.0000.990
w/o LCOEFL_{\mathrm{COEF}}0.9990.9990.9980.9960.9960.9910.9990.991
w/o LSEDL_{\mathrm{SED}}0.9991.0001.0000.9990.9990.9810.9980.981
Only LTIMEL_{\mathrm{TIME}}0.9970.9980.9950.9930.9960.9190.9620.920
Only LCOEFL_{\mathrm{COEF}}0.9270.9500.8730.8130.6650.9860.9990.985
Reviewer_EBa5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TSComp, a pre-training framework for multi-task time series foundation models. A RevIN- normalized univariate window is decomposed by a 3-level orthonormal Symlet-4 DWT into four coefficient streams. Each stream is normalized (SCNorm) and time-aligned patched; the patch length is set to p = 8 in the original time domain and scaled per stream as p_s = p/2^s (with p_0 = p/2^L for the approximation), so all streams have the same number of patches. Each stream is given a stream summary token. Scale Coefficient Attention alternates attention across the four aligned tokens with attention within each stream. Multi-scale directional masking hides a reference coefficient patch together with its aligned coarser or finer patches. The objective is L_SED + L_TIME + L_COEF: a KL term on the energy distribution over scales predicted from the summary tokens, an iDWT time-domain reconstruction loss, and a coefficient reconstruction loss weighted by the predicted energies. Two sizes (12.17M and 47.93M parameters, Table 15) are pre-trained on 4.12B observations and adapted with scale-aware heads for forecasting, imputation, classification and anomaly scoring. The paper reports gains over GTM, TSPulse, SymTime, MOMENT and task-specific models on 8 forecasting/imputation datasets, 27 UEA datasets and the UCR anomaly archive. It adds an analysis of reconstruction quality and cross-scale correspondence(Table 1), an industrial regression task, ablations and theoretical analysis.

Critical strengths

The paper gives a precise, correct argument for why plain time-domain MSE weights scales by energy (Theorem 3, a direct Parseval consequence). It also shows how time-domain masking before a DWT contaminates coefficients outside the mask. This makes coefficient-domain masking a well-motivated idea. The paper is candid about where its components do not help (Tables 3c, 22, 23, 24, 25 and the limitations section). It covers four task families, an industrial dataset and a large appendix.

Critical weaknesses

The central claim is that explicit cross-scale compositional pre-training fixes the limits of time-domain masked reconstruction. The evidence does not test that claim directly.

No model with the same encoder size, corpus and compute is trained with conventional time-domain patch masking. The comparisons are against released MOMENT, GTM, TSPulse and SymTime checkpoints, which differ in corpus, parameters and pre-training compute (none reported for baselines; TSComp-base took about 98 hours on 4 A100-80GB GPUs). TSComp processes four streams, each with N+1 tokens (N = T/p patches plus one SST), against N tokens for a p = 8 patch model, with no FLOPs or throughput comparison. Appendices E.1 and E.5 ablate only inside the wavelet framework, so they cannot show that wavelet-domain pre-training beats time-domain pre-training.

The diagnostic study does not conclusively support the motivation. (a) The misaligned inputs are built with the same 3-level Sym4 DWT that TSComp tokenizes (Section 2), so TSComp is probed in its native domain. (b) MOMENT has the highest probe score among baselines (balanced accuracy 0.877) yet near-zero detail-scale correlations and a time-domain correlation of 0.043 (Table 1). This suggests the probe detects any input change rather than cross-scale correspondence, and no control is given (raw-input probe, untrained TSComp, TSComp without cross-scale embedding). (c) The reconstruction protocol (masked or unmasked, mask ratio) is not stated (Section 2). L_TIME supervises the whole stream, including unmasked patches(Eq. 20, Appendix G.3), so near-perfect fidelity (0.999) may follow from the objective (Table 1). A correlation of 0.043 for MOMENT is implausibly low for a working reconstruction.

The loss design is not supported by its own ablation. L_COEF is weighted by the predicted energy r̂_s, which reintroduces energy weighting. Once coupled to L_COEF, L_SED is no longer a calibration target (Eq. 23: the optimum equals the true distribution only when per-stream errors are equal). In Table 3(c) on ETTm1 forecasting, all three leave-one-out variants (0.3562, 0.3556, 0.3575) beat the full objective (0.3609). For imputation, dropping L_SED(0.0843) and using L_COEF alone (0.0824) beat the full 0.0849. The unit loss weights are never tuned. Removing SST and L_SED (Table 24) also appears to remove the energy weights used by the classification head, so its drop (68.18 to 65.57) is confounded. The w/o Scale variant(Table 3b) changes pooling and head together. Finally, under uniform sampling of reference stream and direction, 25% of selected positions mask all four scales (coarsest reference with coarse-to-fine, or finest reference with fine-to-coarse) , leaving no cross-scale context.

Appendix C.1 discloses that Monash Electricity and Traffic are in the corpus. Table 14 reports 321 series and 8,443,584 observations for Electricity Hourly, and 862 series and 15,122,928 observations for Traffic Hourly. Dividing by the number of series gives 26,304 and 17,544 timesteps per series, which match the standard ECL and Traffic benchmark lengths. The test periods of two of the eight forecasting/imputation datasets were therefore likely seen during pre-training.

Table 19 lists validation sets with exactly the test sizes for all 27 UEA datasets(300/300, 15/15, 3,524/3,524, and so on). If validation is the test set, checkpoint selection used test labels. PatchTST imputation is nearly flat across mask ratios and large (ETTh1 0.668 to 0.673; Traffic 1.327 to 1.340, above 1.0 on standardized data). GTM imputation error falls as masking rises (ETTh1 0.657 to 0.392; Traffic 1.390 to 0.550). TSPulse scores 30.0% on BasicMotions and 32.4% on Epilepsy (4 classes each). MOMENT and SymTime show seed SDs of 20 to 35 points (MOMENT ArticularyWordRecognition 61.00 ± 35.36). All baselines use input length 512, and PatchTST, iTransformer and DLinear are absent from the forecasting table.

The imputation gap to GTM(0.104 vs 0.257) partly reflects these failures. Trainable components also differ across tasks: encoders are frozen for classification and anomaly detection, imputation is fully fine-tuned, and forecasting is unstated.

Forecasting: TSComp-base averages 0.330 vs 0.342 for TSPulse, and Exchange alone supplies 0.0069 of the 0.0122 gap. Exchange is a series where 720-horizon MSE is above 1.0 for every model (Table 26). Classification: 68.18 vs 67.76 for iTransformer, with TSComp-small winning 13 and losing 14 of 27 datasets. The average gain comes mainly from Epilepsy (+25.9), StandWalkJump (+17.8) and AtrialFibrillation (+13.3), the latter two with 15 test series each. TSComp-base (65.91) is below TSComp-small. Anomaly detection: by my rough tally, the five largest TSComp-small rAUCPR entries in Table 28 (0.407, 0.262, 0.148, 0.120, 0.096) sum to about 1.03, roughly 0.012 of a mean over the ~86 listed rows, which is about the whole 0.042 vs 0.030 gap. GTM is better on rAUCROC and VUSROC (0.679 vs 0.654 and 0.673 vs 0.644). There are no significance tests, and each model size is pre-trained once, so the reported SDs cover only fine-tuning seeds.

Questions for authors

Please report forecasting and imputation results excluding Electricity and Traffic, or with those series held out of pre-training. Do the Monash series cover the benchmark test period?

Which split is used for early stopping and checkpoint selection in each task, for TSComp and for each baseline? This matters because Table 19 lists validation and test sets of identical size.

Please state the Table 1 protocol(masked or unmasked reconstruction, mask ratio, probe architecture) and add controls: a raw-input probe, an untrained TSComp, and TSComp without cross-scale embedding. Why is MOMENT's time-domain correlation 0.043?

Please add a controlled baseline with the same encoder, corpus and budget trained with time-domain patch masking, and one with wavelet inputs but no cross-scale embedding.

Please audit the baselines: PatchTST and GTM imputation, TSPulse and MOMENT classification, and the large seed SDs. Compare with the original papers reported numbers, add PatchTST, iTransformer and DLinear(with tuned look-back) and a decomposition-based masked pre-training method such as ST-MTM to forecasting, and report any zero-shot or few-shot results.

For each task and method, which parameters are trainable? Please give trainable parameter counts, FLOPs and throughput, and the data, parameters and compute of each baseline checkpoint.

Please report results with tuned loss weights and a detached r̂. For classification, please ablate the energy weighting separately (uniform or learned weights). For anomaly scoring, please apply the same wavelet residual scoring to baseline reconstructions.

For downstream imputation, how are time-domain masks mapped to coefficient masks, given the contamination described in Appendix G.1? How are UEA series of length 8 to 17,984 fitted to T = 512?

How was Sym4 chosen? Table 23 shows Db4 at 70.50 classification accuracy and Haar at 0.0562 VUSPR, against 68.18 and 0.0421 for Sym4. Please report per-dataset tests, medians for anomaly detection, and variance across pre-training seeds. Why does the larger model underperform the small one on classification?

20 comments

Dear Reviewer EBa5

We thank the reviewer for their careful evaluation and constructive feedback. Their detailed analysis identified important gaps in our original experimental evaluation. In response, we have (1) added a capacity-matched MOMENT baseline and controlled cross-scale interaction ablations (Tables 1–2), (2) strengthened the diagnostic evaluations with additional probing controls and pre-training objective ablations (Tables 4–7), (3) conducted further analyses of gradient coupling and scale-aware downstream adaptation (Tables 8–11), (4) audited baseline evaluation protocols and potential pre-training data overlap (Tables 12 and 14), and (5) expanded forecasting comparisons, computational analysis, and statistical significance testing (Tables 3 and 15–16). These additional experiments and clarifications will be incorporated into the revised manuscript. Detailed responses are provided below.


Cross-scale Compositional Pre-training over Time-domain Masked Reconstruction

(W1) The central claim is that explicit cross-scale compositional pre-training fixes the limits of time-domain masked reconstruction. The evidence does not test that claim directly.

We appreciate the reviewer’s suggestion to provide more direct empirical evidence for the benefits of cross-scale compositional pre-training over conventional time-domain masked reconstruction.

Time-domain masked reconstruction does not explicitly supervise the multi-scale structure of time series, in which temporal patterns at different resolutions jointly form the observed signal. Learning such structures presents two key challenges: (1) explicitly distinguishing scale-specific temporal structures and (2) preserving their cross-scale correspondence, as the characteristics of each component depend on its interactions with other scales. However, conventional time-domain masked reconstruction does not explicitly address either challenge. By supervising the reconstructed signal as a whole, it aggregates reconstruction errors across scales, with each scale contributing in proportion to its signal energy (Theorem 3). Furthermore, time-domain masking removes all scale-specific information within the masked region, preventing direct supervision of missing-scale recovery from complementary scales.

TSComp is designed to explicitly preserve scale-specific temporal structures and promote learning of their cross-scale correspondence. The diagnostic study in Table 1 of the manuscript provides empirical evidence addressing both challenges. For the first challenge, TSComp achieves near-perfect reconstruction fidelity across all wavelet scales, whereas existing TSFMs exhibit substantially lower fidelity, particularly at finer scales. For the second challenge, TSComp achieves near-perfect cross-scale correspondence probe performance, outperforming all compared TSFMs in distinguishing temporally aligned from misaligned scale components. These findings support our claim that TSComp overcomes the representational limitations observed in conventional time-domain masked-reconstruction TSFMs by preserving scale-specific temporal structures and encoding cross-scale correspondence more effectively.

To further examine whether these advantages translate into downstream performance, we introduce Capacity-matched MOMENT, a baseline pre-trained by time-domain masked reconstruction (MOMENT). For a controlled comparison, Capacity-matched MOMENT uses the same pre-training corpus, masking ratio, encoder layer, hidden dimension, and number of training epochs as TSComp-small. We compare this baseline with TSComp-small without scale-aware downstream adaptation (w/o Scale), reducing the influence of the proposed task-specific adaptation on the comparison. As shown in the Table 1, TSComp-small w/o Scale shows superior performance over the capacity-matched MOMENT across all three evaluated tasks. These improvements are achieved without the proposed scale-aware downstream adaptation, indicating that the performance gains are attributable to cross-scale compositional pre-training. Although the models differ in their backbone architectures, the comparison controls for pre-training data and model capacity, providing additional empirical support for the effectiveness of cross-scale compositional pre-training over conventional time-domain reconstruction-based pre-training.

Table 1. Comparison of TSComp without scale-aware adaptation with a capacity-matched MOMENT baseline pre-trained on the same corpus

VariantParamsETTh1 Imp. MSEClassification Avg. Acc.Anomaly VUSPR
Capacity-matched MOMENT11.38M0.302447.670.0266
TSComp-small w/o Scale12.17M0.188567.000.0343

Matched-scale Comparison with Time-domain Masked Reconstruction

(W2-a) No model with the same encoder size, corpus and compute is trained with conventional time-domain patch masking. The comparisons are against released MOMENT, GTM, TSPulse and SymTime checkpoints, which differ in corpus, parameters and pre-training compute (none reported for baselines...
(W2-c) Appendices E.1 and E.5 ablate only inside the wavelet framework, so they cannot show that wavelet-domain pre-training beats time-domain pre-training.
(Q4) Please add a controlled baseline with the same encoder, corpus and budget trained with time-domain patch masking, and one with wavelet inputs but no cross-scale embedding.

We thank the reviewer for emphasizing the importance of controlled comparisons between cross-scale compositional pre-training and conventional time-domain masked reconstruction approaches.

In existing TSFM studies, comparisons with other foundation models typically rely on publicly released pre-trained checkpoints under a common downstream evaluation protocol, without retraining every baseline on an identical corpus with matched model capacity and computational budgets [1, 2]. Such controlled comparisons are challenging because foundation models differ substantially in their architectures, pre-training objectives, and data-processing pipelines, and retraining multiple foundation models requires considerable computational resources.

Nevertheless, to provide more controlled comparisons, we conduct two additional experiments. First, we pre-train a capacity-matched MOMENT model (11.38M parameters) using conventional time-domain masked reconstruction on the same corpus as TSComp-small, matching the masking ratio, number of encoder layers, hidden dimension, and training epochs. Second, to examine the contribution of explicit cross-scale interactions beyond wavelet decomposition, we additionally evaluate TSComp-small without cross-scale embedding, retaining its wavelet-based representation while removing the proposed cross-scale interaction mechanism.

As shown in Table 2, TSComp-small substantially outperforms the capacity-matched MOMENT across all three downstream tasks, providing additional empirical support for the effectiveness of cross-scale compositional pre-training compared with conventional time-domain masked reconstruction. Furthermore, removing cross-scale embedding consistently degrades performance across imputation, classification, and anomaly detection, despite retaining the wavelet-based representation. These results suggest that TSComp's performance gains are not attributable to wavelet decomposition alone, but also benefit from explicitly modeling cross-scale interactions during pre-training.

Table 2. Comparison with capacity-matched MOMENT and TSComp without cross-scale embedding

VariantETTh1 Imp. MSEClassification Avg. Acc.Anomaly VUSPR
Capacity-matched MOMENT0.302447.670.0266
TSComp-small w/o cross-scale emb0.185767.620.0297
TSComp-small0.181368.180.0421

[1] He, C. et al., GTM: a general time-series model for enhanced representation learning of time-series data. ICLR 2026.

[2] Ekambaram, V. et al., TSPulse: tiny pre-trained models with disentangled representations for rapid time-series analysis. ICLR 2026.

Efficiency Comparison with Foundation Model Baselines

(W2-b) ... with no FLOPs or throughput comparison. 
(Q6-b) Please give trainable parameter counts, FLOPs and throughput, and the data, parameters and compute of each baseline checkpoint.

In response to the reviewer's concern, we additionally compare the computational requirements of TSComp with existing TSFM baselines, reporting model parameters, pre-training data scale, and forward FLOPs. Forward FLOPs are measured on ETTh1 using a 512-step lookback window, a 96-step prediction horizon, and seven input channels.

As shown in Table 3, TSComp-small requires approximately 81% fewer forward FLOPs than MOMENT and SymTime, while achieving competitive performance across diverse downstream tasks, particularly forecasting and imputation. These results highlight the favorable balance between computational efficiency and downstream performance achieved by TSComp-small. GTM and TSPulse require fewer forward FLOPs, whereas TSComp-base incurs higher computational costs due to its larger architecture. Regarding pre-training data, TSComp uses 4.12 billion observations, exceeding the corpus sizes reported for MOMENT, GTM, and TSPulse, while SymTime uses a larger synthetic corpus.

Table 3. Computational cost comparison of TSFM baselines. Forward FLOPs are measured on ETTh1 with a 512-step lookback, a 96-step forecast horizon, and seven input channels. Total Params measure the number of parameters of the backbone and the forecasting head.

ModelTotal Params (M)Pre-training Data ScaleForward FLOPs (G/sample) ↓
MOMENT114.3541.23B77.230
TSPulse0.519~1.00B0.276
SymTime43.45150.00B (Synthetic)76.863
GTM37.495~1.02B4.275
TSComp-small13.7374.12B14.723
TSComp-base51.0724.12B116.096

Detailed Probe Protocol and Result Analysis

(W3) The diagnostic study does not conclusively support the motivation. 
(W3-a) The misaligned inputs are built with the same 3-level Sym4 DWT that TSComp tokenizes (Section 2), so TSComp is probed in its native domain. 
(W3-b) MOMENT has the highest probe score among baselines (balanced accuracy 0.877) yet near-zero detail-scale correlations and a time-domain correlation of 0.043 (Table 1). This suggests the probe detects any input change rather than cross-scale correspondence, and no control is given (raw-input probe, untrained TSComp, TSComp without cross-scale embedding). 
(W3-c) The reconstruction protocol (masked or unmasked, mask ratio) is not stated (Section 2). L_TIME supervises the whole stream, including unmasked patches (Eq. 20, Appendix G.3), so near-perfect fidelity (0.999) may follow from the objective (Table 1). A correlation of 0.043 for MOMENT is implausibly low for a working reconstruction.

(Q3) Please state the Table 1 protocol (masked or unmasked reconstruction, mask ratio, probe architecture) and add controls: a raw-input probe, an untrained TSComp, and TSComp without cross-scale embedding. Why is MOMENT's time-domain correlation 0.043?

(W3-a) Construction of Misaligned Inputs:

We acknowledge that constructing misaligned inputs using the same decomposition method as TSComp may favor representations that explicitly operate in the wavelet domain. The purpose of this probe is not to evaluate the role of the wavelet decomposition mechanism itself, but to assess whether the learned representations encode information about cross-scale temporal correspondence, as measured by their ability to distinguish temporally aligned from misaligned scale components.

To assess the contribution of pre-training beyond the wavelet-based architecture itself, we additionally evaluate a randomly initialized TSComp model. As shown in Table 4, pre-trained TSComp achieves a balanced accuracy of 0.9989, substantially outperforming randomly initialized TSComp (0.9109). Although the randomly initialized model already exhibits considerable discriminability, the improvement after pre-training suggests that the learned representations capture additional information relevant to cross-scale correspondence.

Table 4. Cross-scale correspondence probe results, averaged across eight datasets.

ModelB.AccAUROCF1
TSComp0.99891.00000.9989
TSComp (Initialized)0.91090.95610.9164
MOMENT-base0.87730.94890.8859
GTM0.81500.88450.8228
TSPulse0.65760.70180.6689
Reviewer_T3jnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TSComp is a wavelet-based pre-training framework that combines scale-specific representations, cross-scale attention, and directional masking. Experiments on forecasting, imputation, classification, and anomaly detection demonstrate competitive downstream performance.

Critical strengths

1.The paper consistently incorporates scale-specific information into tokenization, pre-training, and downstream adaptation. These components follow a clear motivation and provide a coherent approach to learning and transferring multi-scale time-series representations. The theoretical discussion also helps explain the behavior and limitations of the proposed objectives. 2.The experiments cover four substantially different tasks and show competitive results, particularly for imputation and PR-based anomaly detection. Table 3 supports the usefulness of cross-scale interaction and scale-aware adaptation, while the supplementary ablations report task-dependent trade-offs. The presentation and implementation details generally make the approach easy to follow.

Critical weaknesses
  1. PhysioWave [1] combines wavelet representations with frequency-guided masked reconstruction, WaveToken [2] pre-trains an autoregressive model to predict wavelet coefficients, and MMR [3] reconstructs masked coefficients across wavelet scales. TSComp differs in its aligned scale tokens, factorized cross-scale attention, directional masking, and combination of reconstruction and energy objectives. However, the paper does not sufficiently distinguish which of these elements provides a new capability beyond prior wavelet-based pre-training, or demonstrate that the proposed cross-scale interaction is necessary rather than one effective implementation choice. A more explicit comparison of objectives and mechanisms would make the methodological contribution clearer. 2.Table 3(a) shows that removing either component of SCA hurts performance, but these ablations do not compare SCA with a capacity- and compute-matched Transformer using the same wavelet tokens, or with independently encoded scales followed by learned fusion. The benefits of other components are also task dependent: in Table 3(c), the complete three-objective combination is not the best configuration for either reported ETTm1 setting, and Appendix Table 25 shows that random masking achieves higher anomaly-detection VUSPR than directional masking. These results do not negate the strong overall performance, but leave unclear which parts of the full design are essential for broad transfer. [1] Chen, Yanlong, et al. "Physiowave: A multi-scale wavelet-transformer for physiological signal representation." Advances in Neural Information Processing Systems 38 (2026): 169357-169399. [2] Masserano, Luca, et al. "Enhancing foundation models for time series forecasting via wavelet-based tokenization." ICML, 2025. [3] Thukral M, Tanade C, Lee S A, et al. Wavelet-driven masked multiscale reconstruction for ppg foundation models[J]. arXiv preprint arXiv:2601.12215, 2026.
Questions for authors
  1. Which additional capability or empirical finding most clearly distinguishes TSComp from MMR, PhysioWave, and WaveToken? Are there capacity-matched comparisons against conventional attention or independently encoded scales followed by learned fusion that could further establish the value of SCA?
  2. How are datasets, sequences, and windows sampled during pre-training? Please clarify the handling of sequences shorter than 512 steps and whether large data sources are reweighted or capped, given the substantial differences in corpus sizes across domains.
  3. Could you clarify the data-splitting and model-selection protocols in Tables 18–20? Which forecasting horizon or imputation setting do the window counts in Table 18 represent? How are the validation sets in Table 19 constructed, and are they disjoint from the test sets? How is the final model selected when no validation samples are available in Table 20? Are these procedures consistent across baselines and independent of test data for tuning and checkpoint selection?
5 comments

Dear Reviewer T3jn,

We thank the reviewer for their careful evaluation and constructive feedback. In response to their comments, we have (1) clarified the methodological distinctions between TSComp and existing wavelet-based pre-training approaches, (2) conducted capacity-matched comparisons to evaluate the effectiveness of Scale Coefficient Attention (Table 1), (3) added diagnostic ablations to examine how the pre-training objectives contribute to scale-specific reconstruction and cross-scale correspondence learning (Table 2), and (4) clarified the pre-training data sampling strategy and downstream evaluation protocols. We will incorporate these additional analyses and methodological clarifications into the revised manuscript. Detailed responses are provided below.


Distinction of TSComp Pre-training over Existing Wavelet-based Pre-training

(W1-1) However, the paper does not sufficiently distinguish ...
(Q1-1) Which additional capability ... distinguishes TSComp from MMR, PhysioWave, and WaveToken? 

We thank the reviewer for highlighting the need to clarify the methodological distinctions between TSComp and existing wavelet-based pre-training approaches. While these methods can learn cross-scale dependencies, they do not explicitly supervise cross-scale composition, which is the recomposition process of scale-specific components into the original time-domain signal during pre-training.

PhysioWave models cross-scale dependencies in a directional manner through hierarchical feature fusion with learnable wavelet decomposition. In contrast, TSComp enables bidirectional cross-scale interactions through SCA, while using a fixed orthonormal wavelet basis. Furthermore, TSComp supervises time-domain signal recomposition for cross-scale compositional pre-training, whereas PhysioWave focuses on individual masked coefficient-feature reconstruction.

MMR applies conventional self-attention over flattened wavelet coefficient patches, allowing cross-scale dependencies to be learned through masked coefficient reconstruction. In contrast, TSComp explicitly separates cross-scale interactions from within-scale temporal dependencies and supervises both coefficient recovery and the reconstruction of the original time series from the recovered scale-specific components.

WaveToken employs wavelet-based tokenization for autoregressive forecasting, learning dependencies across scales through next-token prediction. Unlike WaveToken, which optimizes coefficient-token prediction, TSComp jointly supervises scale-specific coefficient reconstruction and their recomposition into the time-domain signal.

Importantly, TSComp distinguishes itself from existing wavelet-based pre-training approaches through its cross-scale compositional pre-training framework, which explicitly enables bidirectional cross-scale interactions and supervises time-domain signal recomposition. Beyond these methodological differences, TSComp is designed as a general-purpose, multi-task TSFM, whereas MMR focuses on PPG representation learning, PhysioWave targets physiological signal modalities, and WaveToken is tailored to forecasting. TSComp demonstrates effective transfer of a shared pre-trained backbone across four distinct downstream tasks—forecasting, imputation, classification, and anomaly detection—rather than focusing on a particular signal domain or task. Furthermore, TSComp achieves strong performance on a real-world industrial regression task, extending its evaluation beyond standard benchmarks and providing further evidence of its broad applicability. Together, explicit supervision of cross-scale composition and broad downstream transfer constitute the central distinctions of TSComp from existing wavelet-based pre-training approaches.

Comparison with the Variants Without SCA

(W1-2) demonstrate that the proposed cross-scale interaction is necessary ...

The cross-scale interaction in Scale Coefficient Attention (SCA) is designed to capture a fundamental property of inherently multi-scale time series: temporal structures at different resolutions are not independent, but jointly compose the observed signal through their cross-scale relationships. Accordingly, learning scale-specific structures alone is insufficient; their interactions must also be modeled to capture how temporal patterns at different resolutions complement one another. To this end, SCA explicitly models interactions among temporally aligned scale-specific representations, allowing each scale to incorporate information from other resolutions describing the same temporal region. These interactions are bidirectional: coarse-scale structures provide broader temporal context for interpreting fine-scale variations, while fine-scale structures provide localized information that refines coarse-scale representations. Unlike independently encoded scales followed by learned fusion, SCA incorporates cross-scale context before within-scale temporal attention, enabling scale-specific representations to be progressively refined throughout encoding.

As shown in Table 3(a) of the manuscript, removing either cross-scale embedding or coefficient attention consistently degrades performance across all four downstream tasks. These results suggest that both components make complementary contributions to downstream performance, supporting the effectiveness of explicitly modeling cross-scale interactions in multi-scale time series.

(W1-3) do not compare SCA with a capacity- and compute-matched Transformer ... 
(Q1-2) Are there capacity-matched comparisons against ...?

To empirically validate the effectiveness of SCA, we compare with two capacity-matched variants of TSComp: (1) TSComp-Flatten, which replaces SCA with conventional self-attention over flattened wavelet coefficient patches, allowing unrestricted interactions across scales and time; and (2) TSComp-Ind-Fusion, which independently encodes each scale using scale-specific Transformers and combines their representations through learned fusion after encoding. These variants retain the wavelet-based representation while examining whether the explicit alignment-aware interactions in SCA provide benefits over conventional self-attention and independent scale encoding.

As shown in the Table 1, TSComp achieves the lowest MSE across all four evaluated settings, outperforming both capacity-matched variants. Although conventional self-attention can also model cross-scale interactions, TSComp consistently achieves better results in these settings, suggesting that explicitly structuring interactions among temporally aligned scale-specific representations is beneficial. Furthermore, the improvements over TSComp-Ind-Fusion support the advantage of incorporating cross-scale context throughout encoding rather than combining independently learned representations only at the final stage.

Table 1. Comparison of TSComp with capacity-matched attention variants across forecasting, imputation, and classification tasks. Forecasting and imputation are evaluated using MSE.

ModelForecasting ETTm1Forecasting WeatherImputation ETTh1Imputation ETTm1Classification Avg. Acc.
TSComp0.36090.23400.18130.084968.18
TSComp-Flatten0.36510.23410.19200.085468.06
TSComp-Ind-Fusion0.36700.25040.21680.102163.21

Validation of TSComp Design for Braod Transfer

(W1-4) The benefits of other components are also task dependent: ... but leave unclear which parts of the full design are essential for broad transfer.

We acknowledge that the task-dependent ablation results do not establish that every component of TSComp is essential for broad transfer. However, the primary goal of TSComp is to learn transferable representations of multi-scale temporal structures and their cross-scale relationships through a unified pre-training framework. We do not claim that the complete configuration is optimal for every downstream task. Our pre-training design is motivated by the hypothesis that understanding time series requires not only capturing temporal structures at different scales but also modeling the relationships through which these components jointly form the observed signal.

TSComp implements this learning principle through Multi-scale Directional Masking (MDM) and three complementary objectives for cross-scale compositional pre-training. MDM encourages the recovery of missing scale-specific components from complementary cross-scale context in both coarse-to-fine and fine-to-coarse directions. The three objectives provide distinct learning signals: Scale Energy Distribution Calibration supervises the relative energy contributions of individual scales; Time-domain Signal Reconstruction enforces the fidelity of the signal recomposed from scale-specific coefficients; and Scale-aware Coefficient Reconstruction supervises masked coefficient recovery, weighted by the predicted scale-energy distribution.

Our additional diagnostic ablations in Table 2 show that the complete pre-training objective achieves the strongest combination of scale-specific reconstruction fidelity and cross-scale correspondence discriminability. While using only LTIMEL_{\mathrm{TIME}} preserves high scale-wise reconstruction fidelity, it yields lower correspondence probe performance. Conversely, using only LCOEFL_{\mathrm{COEF}} leads to weaker fine-scale reconstruction (D1 correlation drops from 0.999 to 0.665). Removing individual objectives also reduces correspondence discriminability, even when reconstruction fidelity remains largely unchanged. These findings suggest that the objectives provide complementary learning signals for the representational properties targeted by TSComp.

The complete configuration therefore supports the intended multi-scale representation learning objectives without requiring additive performance gains from every component on individual downstream tasks. Although certain ablations achieve better results in specific settings, TSComp maintains competitive performance across forecasting, imputation, classification, and anomaly detection using a unified pre-training configuration. These results do not establish the necessity of every objective for broad transfer, but show that the complete design jointly achieves strong scale-specific reconstruction and cross-scale correspondence discriminability while supporting diverse downstream tasks.

Table 2. Diagnostic analysis of pre-training objective ablations. Reconstruction fidelity is measured by Pearson correlation in the time domain and at individual wavelet scales.

VariantTimeA3D3D2D1B.AccAUROCF1
TSComp0.9991.0001.0000.9990.9990.9991.0000.999
w/o LTIMEL_{\mathrm{TIME}}0.9950.9940.9980.9980.9980.9901.0000.990
w/o LCOEFL_{\mathrm{COEF}}0.9990.9990.9980.9960.9960.9910.9990.991
w/o LSEDL_{\mathrm{SED}}0.9991.0001.0000.9990.9990.9810.9980.981
Only LTIMEL_{\mathrm{TIME}}0.9970.9980.9950.9930.9960.9190.9620.920
Only LCOEFL_{\mathrm{COEF}}0.9270.9500.8730.8130.6650.9860.9990.985

As a complementary analysis, we examine the effects of scale-energy weighting and gradient coupling within LCOEFL_{\mathrm{COEF}}. Table 3 compares TSComp with variants that either detach the predicted scale-energy weights or replace them with uniform weights in LCOEFL_{\mathrm{COEF}}. The original configuration achieves the best performance across the evaluated imputation, classification, and anomaly detection settings. Although this experiment examines a specific aspect of the loss design, the results provide some empirical support for the current loss formulation as a reasonable design choice.

Table 3. Ablation study on LCOEFL_{\mathrm{COEF}} across downstream tasks.

ModelImputation ETTh1 MSEClassification Acc. (%)Anomaly VUSPR
TSComp0.181368.180.0421
Detached r^s\hat{r}_s0.212966.640.0320
Uniform Weighting0.204663.680.0364

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their detailed and constructive feedback throughout the discussion period. Their comments have directly contributed to substantial improvements in our work, strengthening the methodological clarity, experimental rigor, and empirical validation of TSComp's contributions. The table below summarizes how we have addressed each reviewer's key comments through additional experiments, analyses, and corresponding revisions to the manuscript.

ReviewerCommentHow AddressedManuscript Ref
T3jnDistinction from wavelet-based TSFMsExtended comparison with PhysioWave, MMR, WaveTokenSection 3
T3jnReal-world industrial generalizationIndustrial regression and forecasting on manufacturing dataApp. D, Tab. 22–23
T3jn, cTe8Benefit of cross-scale interactionControlled SCA comparisons (Flatten, LateFusion)App. E.7, Tab. 28
T3jn, EBa5, cTe8Validation of broad transferDiagnostic ablations, capacity-matched MOMENTSection 6.2, Tab. 4, 30
T3jn, EBa5Downstream evaluation protocolPre-training/downstream protocols documentedApp. C.2–C.3
T3jn, EBa5Test-set checkpoint selectionHeld-out validation on 26 UEA datasets; ranking fully preserved, confirming evaluation robustness.App. J.1, Tab. 42
EBa5Pre-training loss weight rationaleLoss trajectory analysisApp. F.3, Fig. 8
EBa5, cTe8Validity of cross-scale correspondence probeAdded raw-input, random-initialization, and cross-scale embedding ablation controlsSection 6.2, Tab. 4
EBa5Data overlapRe-training without Electricity/TrafficApp. E.10, Tab. 31
EBa5Baseline reliabilityAudited baselines, statistical significance testsApp. F.2,I.2,I.4, Tab. 35, 39, 41
EBa5, cTe8Wavelet design sensitivityAnalyses across families, depths, decomposition strategiesApp. E.1–E.3
EBa5, cTe8Contributions of pre-training, downstream adaptationControlled ablations, pre-trainingSection 6.1–6.2, App. E.9 Tab. 3–4, 30
cTe8Robustness under challenging conditionsNon-stationary perturbation, irregular samplingApp. E.11–E.12, Tab. 32–33

Beyond addressing individual comments, the revised manuscript incorporates additional experiments and analyses, substantially strengthening the empirical support for TSComp's contributions. We note that Reviewer cTe8 has confirmed that our responses addressed their main concerns, and Reviewer EBa5 previously acknowledged several of our additional experiments as "convincing" and our design defense as "fair." TSComp's key contribution lies in its unified, task-agnostic pre-training framework that explicitly learns scale-specific temporal structures and cross-scale compositional relationships, enabling competitive transfer across four benchmark task families and real-world industrial applications from a single pre-training configuration. We respectfully hope that the Area Chair will consider the scope of these revisions and how they address the reviewers' specific comments in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.