acceptodds
Under review as a conference paper at ICLR 2027

TimEE: End-to-end Time Series Classification via In-Context Learning

17 views

Abstract

Time series classification (TSC) is dominated by a two-stage paradigm: train a feature encoder—either from scratch on the target dataset or via pretraining on large corpora—and then fit a task-specific classifier on top. While effective, this decoupling optimizes representation learning independently of the classification objective, requires per-dataset training, and prevents the model from exploiting label information during inference. We introduce TIMEE, a 4.5M-parameter foundation model for end-to-end TSC via in-context learning. Given a labeled support set and a query time series, TIMEE outputs a predicted class distribution directly, with no encoder trained or classifier fitted on the target dataset. Following the prior-data fitted network (PFN) framework, TIMEE is meta-trained exclusively on synthetic TSC tasks, where each task contains time series with distinct class identities arising from structured distributional shifts in the generative process. Despite seeing no real time series during pre-training, TIMEE achieves the best mean rank in ROC AUC on the 128-dataset UCR archive, statistically tied with the strongest supervised baselines and ahead of every foundation model. It is also the best-calibrated and fastest of all compared methods. To our knowledge, TIMEE is the first purely synthetic-pretrained model to reach state-of-the-art performance. These results establish end-to-end ICL with synthetic priors as a compelling, largely unexplored direction for TSC, with scaling, prior design, and richer generation mechanisms as natural avenues for improvement.

Primary area:
learning on time series and dynamical systems
Keywords:
time series classification, in-context learning, prior-fitted networks, time series

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_G4UFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper introduces TimEE, an end-to-end time series classification model that uses in-context learning (ICL) within the prior-data fitted network (PFN) framework. Its main contributions are a structured synthetic data-generation process and a transformer architecture that conditions on labeled support examples. In the reported evaluation, TimEE substantially outperforms the prior time series ICL method TiCT and achieves competitive performance against strong supervised baselines. The synthetic prior is a central contribution, supported by ablation studies of its components.

Critical strengths
  1. The paper provides substantial empirical evidence of strong classification performance. Across 128 UCR datasets, TimEE achieves the best mean ROC AUC rank and is statistically tied with the strongest supervised baselines. These results support the practical value of synthetic pretraining for end-to-end time series classification.
  2. The paper is clear and easy to follow, and the proposed method is well motivated. The structured synthetic prior and architecture are described in sufficient detail to make the approach useful to researchers and practitioners developing time series ICL models.
Critical weaknesses

My main concern is that the connection between the synthetic prior and real-world classification tasks could be developed further. The paper provides intuitive motivation and component ablations, but these leave open which properties of the synthetic tasks explain successful transfer to real datasets. A more targeted analysis of this connection would strengthen the scientific contribution and help characterize the method's generalization limits.

Questions for authors
  1. Synthetic prior and generalization (clarification). The prior design is intuitively motivated, and the empirical results and ablations support its effectiveness for ICL training. Could the authors explain which statistical or discriminative properties of real time series classification tasks the prior is intended to capture, and how the existing evidence supports this connection? As a suggestion for further analysis, a focused comparison of synthetic tasks with UCR and/or UEA tasks, or theoretical justification where appropriate, could help explain the observed transfer. Broader validation beyond these benchmark families would be a useful direction for future work.
  2. One-vs-rest prediction (clarification). Section 3.3 states that, when the number of classes C exceeds 10, TimEE decomposes the task into C binary one-vs-rest (OvR) contexts. How are their outputs combined into a single multiclass prediction? For example, if several contexts favor their positive class, is the final label selected using the highest positive-class probability? Please also clarify how the C scores are normalized to obtain a multiclass probability distribution.
  3. TiCT preprocessing (clarification). Given TiCT's relatively weak performance in the reported evaluation, could the authors check whether its input time series were resampled to length 512 before inference? The reference implementation includes this preprocessing step, using scipy.signal.resample(data, 512, axis=2) for arrays shaped (samples, channels, time).
4 comments

Dear Reviewer G4UF,

Thank you for your positive assessment and constructive questions. We address your concerns and questions below.


Weaknesses and Questions

W1 & Q1: Synthetic Prior and Generalization

W1: "My main concern is that the connection between the synthetic prior and real-world classification tasks could be developed further."

Q1: "Synthetic prior and generalization (clarification)."

We agree that there is room to further develop the synthetic prior, and we do not claim that our current prior is optimal. We also acknowledge that the transfer from synthetic to real-world tasks was not sufficiently discussed in the paper.

Nevertheless, we view TimEE's competitive performance on real-world UCR tasks, despite being pretrained exclusively on synthetic tasks, as evidence that the prior captures at least some of the structure underlying real-world classification problems. Similar synthetic-to-real generalization has been observed in tabular foundation models (TabPFN, TabICL) and time-series models pretrained on CauKer [1] (e.g., MantisV2).

Our prior is designed for diversity rather than direct imitation of real-world distributions. Instead of targeting the characteristics of a particular domain or dataset, we aim to capture broad, generic time-series structures and diverse mechanisms through which classes may differ. These mechanisms are sampled from broad meta-distributions (Table 5, Appendix C.1).

Our prior ablations (Appendix C.5, Tables 6–7) provide further evidence for this design choice: each class-generation mechanism individually yields a competitive prior, while combining them achieves the strongest overall performance.

Domain-wise analysis on UCR. To better understand where TimEE transfers successfully and where limitations remain, we additionally break down UCR performance by the domain labels provided in the archive:

UCR typenTimEEMiniRocket
EOG2.910.868
Spectrum4.989.953
Motion17.962.954
Sensor30.954.946
ECG6.982.974
Image32.936.937
Device9.857.877
Hemodynamics3.901.997

TimEE matches or exceeds MiniRocket across several categories of recorded physical signals (Sensor, Motion, ECG, Spectrum) and performs comparably on image outlines. Its largest performance gaps occur in two domains:

  • Device: These datasets contain electrical appliance loads, often characterized by step-like on/off patterns that a stationary VARX process does not naturally generate.
  • Hemodynamics: This category includes a dataset with 52 classes and only two training examples per class, requiring our one-vs-rest strategy.

These gaps suggest opportunities to improve the prior. We also agree that evaluation beyond UCR/UEA would provide valuable additional evidence of generalization, and will explicitly identify this as future work.

Reference

[1] Xie, Shifeng, et al. "CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data." International Conference on Learning Representations (ICLR), 2026.


Q2: Combining One-vs-Rest Outputs

"Combining the one-vs-rest outputs."

For classification tasks with more than 10 classes (C>10C > 10), we construct CC binary contexts. In context cc, support examples belonging to class cc are labeled positive, while all remaining examples are labeled negative. A forward pass through TimEE then produces the positive-class probability pcp_c.

Because these CC contexts are evaluated independently, the resulting probabilities do not necessarily sum to one. We therefore normalize them across classes for each query:

p(y=c∣x)=pc∑j=1Cpj.p(y=c \mid x) = \frac{p_c}{\sum_{j=1}^{C} p_j}.

"If several contexts favor their positive class, is the final label selected using the highest positive-class probability?"

Yes. The predicted class is selected by taking the argmax of the normalized probability distribution (equivalently, the argmax of the unnormalized positive-class probabilities). ROC AUC and log loss are computed using the same normalized distribution.

We will clarify this in Section 3.3 of the paper.


Q3: TiCT Preprocessing

"TiCT preprocessing."

Yes, our TiCT evaluation follows the reference preprocessing procedure.

Each time series is first resampled to a length of 512 using scipy.signal.resample(X, 512, axis=2), where the input has shape (samples, channels, time). The resampled series are then z-normalized individually.

For context construction, we retrieve the 64 nearest neighbors from the training set for each query.

We will make these preprocessing details explicit in the revised manuscript.


We hope these responses clarify the scope of our work, the motivation behind the synthetic prior, and the evaluation methodology.

Thank you for your thoughtful feedback.

Kind regards,
The Authors

Reviewer_G4UFOpenReview ↗

Follow-up on the authors’ response

Thank you for the detailed response. The explanation of OvR normalization and the confirmation of TiCT’s preprocessing address my second and third questions.

The additional domain-wise analysis is helpful and strengthens the empirical evidence for synthetic-to-real transfer. My remaining question concerns which properties introduced by the prior’s structural and signal variations explain this transfer. A brief qualitative analysis would be sufficient—for example, visualizing representative synthetic and real-world series and discussing which class-discriminative patterns the prior captures or underrepresents. This could also help substantiate the proposed explanation for weaker performance on appliance-load datasets.

Overall, I retain my weak acceptance recommendation. The response resolves my implementation questions and provides useful additional evidence of transfer. A clearer qualitative connection between the synthetic prior and real-world task structure would further strengthen the paper.

Dear Reviewer G4UF,

Thank you for your constructive follow-up. Following your suggestion, we have added a qualitative comparison of synthetic and real-world time series in Appendix C.5 (Synthetic Prior and Real-World Data).

In the new Figure 13, we compare representative series from 15 UCR datasets with their nearest synthetic counterparts among 153,600 generated series. We observe similarities across diverse patterns, including smooth trends, localized events, and noisy oscillations. This illustrates the diversity of time-series patterns covered by our prior.

Regarding class-discriminative patterns, our structural and signal variation mechanisms introduce differences in temporal dynamics and time scale, respectively, as illustrated in Figures 10–12. Our prior ablations (Appendix C.6, Table 6) further show that pretraining with either mechanism independently yields competitive performance on real-world UCR datasets.

While our work describes the class-discriminative mechanisms generated by the prior and provides empirical evidence of their transferability, we agree that a deeper understanding of which specific properties enable this transfer to real-world tasks would be valuable. Such an investigation could involve analyzing the discriminative structures of real-world datasets alongside controlled synthetic experiments. We see this as a promising direction for future work.

Thank you again for your helpful suggestion. We hope the additional analysis and clarification are useful.

Kind regards,

The Authors

Thank you for the update. I have increased the soundness score from 2 to 3.

Reviewer_Mx2XOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TimEE, an end-to-end in-context learning framework for time series classification. It uses VARX-generated synthetic tasks for meta-training and adopts a class-aware transformer architecture, enabling query classification from labeled support examples without per-dataset parameter updates. The method achieves competitive performance on the UCR benchmark and reports promising probabilistic calibration and inference efficiency.

Critical strengths
  1. The paper presents an interesting and practically relevant formulation of time series classification as in-context learning, requiring no task-specific fine-tuning at inference.
  2. The structured synthetic VARX prior provides a principled mechanism for generating labeled tasks for meta-training.
  3. The class-aware encoder is designed to capture relationships between support examples and queries in a unified forward pass.
  4. Results on 128 UCR datasets demonstrate competitive performance, while the reported log-loss and inference-speed results highlight potential practical benefits.
  5. The paper provides useful analyses and openly acknowledges important limitations.
Critical weaknesses
  1. TimEE is pretrained exclusively on a single family of synthetic classification tasks generated from a VARX-based prior. Although its performance on the UCR benchmark is encouraging, the extent to which the learned classification procedure generalizes to real-world data with substantially different underlying dynamics remains insufficiently established.

Question: How sensitive is TimEE's performance to the choice of synthetic data-generating process? Can the authors evaluate the model on tasks generated from alternative priors, particularly those with dynamics or class-discriminative structures that differ substantially from those induced by the VARX prior?

  1. The proposed method combines a structured synthetic prior, a class-aware encoder, and a meta-training procedure. However, the respective contributions of these components to the reported performance are not fully isolated. It therefore remains unclear whether the gains primarily arise from the prior design, the architecture, the scale of meta-training, or their interaction.

Question: Can the authors provide more systematic ablation experiments that independently evaluate the effects of the synthetic prior, class-aware architecture, and meta-training configuration? Such experiments would help clarify which components are essential to TimEE's performance and substantiate the claimed benefits of the proposed design.

  1. The paper uses representation visualizations to illustrate how cross-series attention promotes class-discriminative representations. However, these visualizations alone do not establish that the model effectively exploits the labeled support set during inference, rather than relying primarily on representations acquired during pretraining.

Question: How much does cross-series attention contribute to classification performance compared with temporal attention alone? Can the authors conduct controlled experiments, such as removing cross-series attention, shuffling support labels, or varying the number of labeled support examples, to verify that predictions genuinely depend on the support set and its label information?

  1. The multivariate extension yields only marginal gains over the per-variate baseline. The paper attributes this limitation partly to the VARX prior, which tends to concentrate class-discriminative information in individual variates, and acknowledges that the architecture was originally designed for univariate classification. These results leave the model's ability to exploit cross-variate dependencies insufficiently demonstrated.

Question: Can TimEE effectively classify tasks in which discriminative information is encoded primarily in cross-variate dependencies rather than in individual channels? Controlled experiments on such tasks would help determine whether the limitation stems from the synthetic prior, the architecture, or both, and clarify the model's suitability for general multivariate classification.

  1. The reported inference-speed advantage is evaluated on datasets with at most 10 classes. For datasets with more classes, TimEE relies on a one-vs-rest decomposition, which introduces additional inference overhead. Furthermore, although the absence of per-dataset training is an important practical advantage, the reported inference-time comparisons do not fully characterize the computational cost of offline meta-training or the amortization of this cost across downstream tasks.

Question: How does inference latency scale with the number of classes, support examples, sequence length, and input channels? Would the efficiency advantage remain substantial for many-class and high-dimensional multivariate tasks? It would also be useful to report the offline training cost and discuss how many downstream tasks are needed to amortize this cost relative to conventional methods.

  1. TimEE achieves competitive results on UCR, but its mean ROC AUC rank is statistically tied with several strong conventional baselines, while its multivariate extension remains less effective. These findings support the promise of synthetic-task meta-training for time series classification but provide more limited evidence for broad generalization across diverse task distributions and for the model's characterization as a general time series classification foundation model.

Question: Can the authors more precisely distinguish the settings in which TimEE consistently outperforms existing methods from those in which it achieves comparable performance? Further clarification of the scope of its demonstrated generalization would help ensure that the claims about its effectiveness and foundation-model capabilities are commensurate with the experimental evidence.

Questions for authors

Please see the Weaknesses section for details.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Mx2X (Part 1/3): Synthetic Prior and Ablations

Dear Reviewer Mx2X,

Thank you for your detailed and constructive review. We appreciate that you found the ICL formulation practically relevant, the VARX prior principled, and our analyses useful, and that you recognized our discussion of limitations.

We address your questions and concerns across three responses.

WQ1: Sensitivity to the Synthetic Data-Generating Process

"How sensitive is TimEE's performance to the choice of synthetic data-generating process?"

Since TimEE is pretrained exclusively on synthetic data, evaluation on all 128 real-world UCR datasets and 27 UEA datasets constitutes an out-of-distribution evaluation relative to the VARX prior. These datasets span diverse dynamics, including image outlines, spectrograms, ECG, motion, and sensor signals, and TimEE remains competitive across them (Section 4.2, Appendix F.3). We consider this strong evidence that the learned classification procedure transfers beyond the prior's dynamics.

We also analyze sensitivity to prior components in Appendix C.5. Table 6 evaluates each class-generation mechanism separately (structural vs. signal variation), with and without mixup. Each mechanism alone yields a competitive prior, while their combination performs best.

We agree that relying on a single prior family remains a limitation, as discussed in Section 5. Developing more diverse priors is an important direction for future research.


WQ2: Independent Effects of the Prior, Architecture, and Meta-Training

"Independently evaluate the effects of the synthetic prior, class-aware architecture, and meta-training configuration?"

Our ablations isolate the main components of TimEE. The prior and architecture ablations use a fixed, reduced training setup (Appendix B.4, averaged over three seeds), while the inference ablation uses the full model across all 128 UCR datasets.

1. Synthetic prior (Appendix C.5, Tables 6–7). Each class-generation mechanism alone yields a competitive prior, while combining both achieves the best performance (91.92 ROC AUC, 77.30% accuracy). Mixup is the most influential prior component, improving both mechanisms, particularly structural variation (+3.5 ROC AUC, +5.7 percentage points in accuracy), by increasing intra-class diversity. Per-series augmentations have smaller effects, and no individual augmentation outperforms their combination.

2. Architecture (Appendix B.4, Tables 3–4). Both attention mechanisms contribute to performance:

  • Removing cross-series attention produces the largest drop among encoder allocations (80.32 → 78.21 ROC AUC), indicating that cross-series comparisons within the encoder provide information that the ICL stage alone cannot recover.
  • Conditioning on labels only at the ICL stage reduces ROC AUC to 78.55.
  • Removing label conditioning entirely reduces accuracy to approximately 50%.

3. Inference (Appendix G, Table 13). All four preprocessing views perform similarly in isolation (93.8–94.0 ROC AUC), suggesting that neither differencing nor interpolation length is individually decisive. Averaging across views improves ROC AUC by 0.9 and accuracy by 2.5 percentage points over the best individual view.

In summary, TimEE's gains primarily arise from label-aware cross-series encoding and a diverse synthetic prior, with mixup playing an important role. The test-time ensemble provides a smaller additional improvement.

AuthorsOpenReview ↗

Response to Reviewer Mx2X (Part 2/3): Support Conditioning and Multivariate Classification

WQ3: Contribution of Cross-Series Attention and Support-Set Dependence

"How much does cross-series attention contribute to classification performance compared with temporal attention alone?"

We conducted two controlled ablations in Appendix B.4:

  • Removing cross-series attention (Lc=0L_c = 0, Table 3) reduces ROC AUC from 80.32 to 78.21, the largest drop among encoder allocations.
  • Removing label conditioning (Table 4) reduces accuracy from 71.26% to 50.13%. Without support labels, the model cannot classify reliably, demonstrating that predictions depend on label information rather than pretrained representations alone.

The attention mask (Appendix B.2) further ensures that each query is classified using only the support set, independently of other queries.

"Can the authors conduct controlled experiments, such as removing cross-series attention, shuffling support labels, or varying the number of labeled support examples, to verify that predictions genuinely depend on the support set and its label information?"

To address this directly, we conducted two additional experiments, both averaged over three seeds. We will include them in the appendix of the revised paper.

Experiment 1: Shuffling support labels.

We randomly permute support labels while preserving class counts across all 128 UCR datasets. Mean ROC AUC drops from 0.948 to 0.492, demonstrating that the model relies on the correspondence between support examples and their labels.

Support setROC AUCAccuracy
Correct labels0.9480.841
Shuffled labels0.4920.308

Experiment 2: Varying the number of labeled examples.

We evaluate kk labeled examples per class on the 71 datasets containing at least 20 training examples per class.

Examples per class (kk)ROC AUCAccuracy
10.6840.456
50.8300.674
100.8780.740
200.9030.784

ROC AUC increases monotonically with the number of labeled examples, from 0.684 at k=1k=1 to 0.903 at k=20k=20.

Together, these experiments provide direct evidence that TimEE genuinely exploits the labeled support set, rather than relying solely on representations learned during pretraining.


WQ4: Cross-Variate Dependencies in Multivariate Classification

"Can TimEE effectively classify tasks in which discriminative information is encoded primarily in cross-variate dependencies rather than in individual channels?"

We agree that this is an important open question, which we discuss in Section 4.3 and Section 5 (Limitation 2).

Our analysis in Appendix C.1.2 (Figure 9) reveals a limitation of the current VARX prior: class-discriminative information is predominantly concentrated in individual variates, providing little incentive for the model to learn cross-variate aggregation.

Architecturally, the two multivariate variants have different limitations:

  • Per-variate: By construction, this variant cannot exploit purely cross-variate dependencies.
  • Variate-pooling: This variant could, in principle, capture such dependencies, but the current prior provides insufficient training signal to learn them effectively.

TimEE is primarily designed and positioned as a univariate classification model, while the multivariate extension is explored as an ablation (Section 4.3). We will clarify this scope in the revised manuscript.

Developing priors and architectures that better capture cross-variate dependencies is an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer Mx2X (Part 3/3): Computational Complexity and Scope of Generalization

WQ5: Inference Latency and Computational Scaling

"How does inference latency scale with the number of classes, support examples, sequence length, and input channels?"

We summarize the scaling behavior along each dimension:

  • Number of classes: The speed comparison in Figure 6 covers datasets with at most 10 classes. Figure 13 (Appendix F.1) reports results for more than 10 classes, where one-vs-rest decomposition introduces substantial overhead. We acknowledge this in Section 5 (Limitation 1).
  • Sequence length: Inputs are interpolated to fixed lengths of 256 and 512 time steps (Section 3.4, Appendix D.2), making the model's computational cost independent of the original sequence length after preprocessing.
  • Support and query size: Temporal attention operates within individual series, while cross-series attention operates across series at each patch position. Queries attend only to the support set (Appendix B.2). Therefore, inference cost scales linearly with the number of queries and quadratically with support size.
  • Input channels: Both multivariate variants require one encoder pass per variate, so computational cost scales linearly with the number of channels (Section 5, Limitation 2).
  • Training cost: Pretraining takes approximately 40 hours on a single H200 GPU (Section 3.4), a one-time cost amortized across downstream tasks. Compared with InceptionTime, which requires approximately 1,600 seconds of training per dataset (single H100 GPU, Figure 6), this cost is recovered after roughly 90 datasets, within the UCR archive alone. Compared with fast feature-based methods such as MiniRocket, TimEE's primary advantage is predictive quality, particularly calibration (Figure 5), rather than computational cost.

Computational complexity.

Let:

  • nn denote the number of support series and mm the number of queries, with N=n+mN=n+m.
  • PP denote the number of patches (sequence length divided by patch size 16).
  • K=4K=4 denote the number of CLS tokens.
  • d=128d=128 denote the embedding dimension and dclsd_{\mathrm{cls}} the class-token embedding dimension.
  • Lt=4L_t=4, Lc=4L_c=4, and LICL=2L_{\mathrm{ICL}}=2 denote the numbers of temporal, cross-series, and ICL attention layers.

The attention complexity of each component is:

Temporal attention (within each series):

O(LtN(P+K)2d)O\left(L_t N(P+K)^2d\right)

Cross-series attention (across series at each patch position):

O(Lc(P+K)nNd)O\left(L_c(P+K)nNd\right)

ICL attention:

O(LICLnNdcls)O\left(L_{\mathrm{ICL}}nNd_{\mathrm{cls}}\right)

Consequently, inference cost grows linearly with the number of queries, quadratically with the number of support examples, and quadratically with the processed sequence length through temporal attention.

We will include this complexity analysis in the revised manuscript.


WQ6: Scope of Performance and Generalization Claims

"Can the authors more precisely distinguish the settings in which TimEE consistently outperforms existing methods from those in which it achieves comparable performance?"

We agree that the scope of our claims should reflect the experimental evidence. We will clarify the following distinctions in the revision.

Where TimEE leads:

  • ROC AUC: Best mean rank across 128 UCR datasets, statistically tied with InceptionTime, MiniRocket, and Hydra (Figure 14).
  • Efficiency: On the speed–accuracy Pareto front for datasets with at most 10 classes (Figure 6).
  • Log loss: Best mean log-loss rank by a clear margin on datasets with at most 10 classes (Figure 5), indicating strong probabilistic predictions and calibration.

Where TimEE is competitive but not optimal:

  • Accuracy: Third-best rank on UCR (Section 4.2).
  • Multivariate classification: Competitive but limited performance (Figure 7).

Where TimEE underperforms:

  • More than 10 classes: One-vs-rest decomposition incurs additional inference overhead (Figure 13) and reduces predictive performance. Across the 25 datasets with more than 10 classes, TimEE achieves 0.748 mean accuracy versus MiniRocket's 0.797. In contrast, on the 103 datasets with at most 10 classes, their performance is nearly identical (0.864 vs. 0.866).

We will therefore scope our primary claims to univariate time-series classification with at most 10 classes, the setting for which TimEE is pretrained.


We hope these additional experiments and clarifications address your concerns and better establish both the contributions and limitations of our approach.

Thank you again for your constructive feedback.

Kind regards,
The Authors

Reviewer_14tvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TimEE, an end-to-end ICL(in-context learning) framework for time series classification. It combines a VARX-based synthetic classification prior with a Transformer architecture that leverages labeled support examples through cross-series attention and label conditioning. Experiments on 128 UCR datasets demonstrate competitive classification performance, probabilistic prediction quality, and inference efficiency.

Critical strengths

S1. A promising alternative paradigm for time series classification.

  • TimEE explores e2e ICL as an alternative way to conventional two-stage time series classification. It combines synthetic task generation with context-aware classification design. This provides a useful exploration of how labeled support examples can be directly leveraged at inference time.

S2. Competitive performance and inference efficiency.

  • TimEE achieves competitive classification performance across 128 UCR datasets. On datasets with at most 10 classes, it matches MiniRocket's accuracy while achieving approximately 25× faster inference, and is over 1,000× faster than InceptionTime.
Critical weaknesses

W1. Limited evidence of methodological advances.

  • Although TimEE introduces a VARX-based synthetic prior and a Transformer architecture with cross-series attention, its methodological advantages over TiCT (the paper itself discusses and compares against), which also adopts synthetic pretraining for end-to-end in-context time series classification, remain unclear. More targeted comparisons are needed to establish the benefits of these design choices.
Questions for authors

Q1. Baseline selection protocol.

  • How are encoder layers selected for the tuned foundation model baselines? Is the selection based exclusively on independent validation data, without using test-set labels or statistics?

Q2. Evaluation under imbalanced class labels.

  • Can the authors report macro-averaged precision, recall, and F1-score to provide a more comprehensive assessment of classification performance, particularly on datasets with imbalanced class distributions?

Q3. Consistency of hyper-parameter settings.

  • Section 3.3 reports four temporal attention layers and four cross-series attention layers, whereas Appendix B.1 (Table 2) lists five for each. Which configuration is used in the reported experiments?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 14tv (Part 1/2): Methodological Contributions

Dear Reviewer 14tv,

Thank you for your valuable feedback. We appreciate that you found end-to-end ICL a promising alternative paradigm and recognized TimEE's competitive performance and inference efficiency. We address the methodological concerns below and the remaining questions in Part 2.

W1: Limited Evidence of Methodological Advances

"Limited evidence of methodological advances"

We agree that TiCT is our closest related work. We highlight the differences between our approach and TiCT across three aspects:

1. Architecture and representations. A core difference lies in the architecture and where series labels enter the computation. TiCT encodes each series independently with ResNet and introduces labels only after forming compressed per-series representations. In contrast, TimEE constructs label-aware series representations and refines them using information from the labeled context.

Our ablations (Appendix B.4, Tables 3–4) demonstrate the importance of label conditioning and cross-series attention:

  • Without cross-series attention (Lc=0L_c = 0, Table 3), ROC AUC drops from 80.32 to 78.21.
  • With label conditioning only at the ICL stage, as in TiCT (Table 4), ROC AUC drops from 80.32 to 78.55.

Figure 8 further illustrates the role of label-aware representations: query series, which receive no labels, align with their correct class clusters only after cross-series attention.

2. Synthetic prior. TiCT's prior generates natively binary tasks by mixing two KernelSynth templates, with class labels determined by whether the mixing ratio crosses a random threshold. Class identity is thus defined by a cutoff along a single interpolation axis.

In contrast, TimEE's VARX-based prior generates natively multi-class tasks (up to 10 classes), where classes differ in their dynamics through either dependency structure (structural variation) or the speed of exogenous signals (signal variation).

3. Context handling. TiCT retrieves 64 nearest neighbors per test sample using Euclidean distance. TimEE instead attends over the full training set, removing dependence on a fixed similarity metric for retrieval.

Downstream performance. To ensure a fair comparison, we directly adopt TiCT's evaluation protocol. Under this protocol, TimEE outperforms TiCT on all reported metrics while using roughly 10× fewer parameters (4.5M vs. 47M):

MetricTimEETiCT
UCR ROC AUC (mean / mean rank).948 / 4.98.889 / 14.56
UCR accuracy (mean / mean rank).841 / 5.32.727 / 13.88
UCR log-loss rank (10 classes)3.2215.60
Pairwise ROC AUC win rate94%6%
UEA ROC AUC rank8.98 (PV)14.65

We attribute these gains to the design choices outlined above: label-aware series representations, a natively multi-class prior with distinct class dynamics, and attention over the full labeled context. Together, these choices and their empirical benefits constitute our main methodological contribution relative to TiCT.

AuthorsOpenReview ↗

Response to Reviewer 14tv (Part 2/2): Questions and Clarifications

We address the reviewer's three questions below.

Q1: Baseline Selection Protocol

"Baseline selection protocol"

We evaluated every possible exit layer for each foundation model that benefits from layer tuning (MantisV2, MOMENT, NuTime, Chronos-2, TiRex), using the validation setup described in Appendix B.4. The best layer was selected based on mean aggregated performance across the validation datasets only. No test-set data or statistics were used in this selection.

When re-running this procedure, the selected layer shifts slightly for two models (MOMENT: 13 → 14; TiRex: 6 → 5). These differences arise from randomness in training and evaluation, as performance differences between adjacent layers are within run-to-run noise. In both cases, the newly selected layer is statistically tied with the originally chosen one.

We report the per-layer results below and will include the full table in the revised manuscript. We will also reevaluate these two tuned baselines and update the corresponding main results, figures, and tables.

LayerChronos-2TiRexMantisV2NuTimeMOMENT
192.3291.6092.9890.1989.78
293.3792.7293.7591.2690.20
393.8393.2493.7891.9090.58
493.8693.5093.3292.1890.94
593.7493.7892.8992.1991.23
693.6493.5292.5292.1991.54
793.5893.25––91.66
893.0993.08––91.79
992.9192.80––91.93
1092.4192.69––91.99
1192.0392.42––92.10
1291.1991.65––92.03
13––––92.10
14––––92.14
15––––92.02
16––––92.03
17––––92.01
18––––92.02
19––––91.97
20––––91.90
21––––91.90
22––––91.89
23––––91.92
24––––91.89

Q2: Evaluation Under Imbalanced Class Labels

"Evaluation under imbalanced class labels"

We report macro-averaged precision, recall, and F1 across all 128 UCR datasets (macro recall is equivalent to balanced accuracy).

To specifically examine class imbalance, we additionally evaluate the 34 datasets with imbalanced training sets, defined as those where the largest class contains more than twice as many examples as the smallest class. Ranks are computed per dataset among the models listed below.

Consistent with Section 4.2, TimEE ranks closely with the strongest baselines (MiniRocket, Hydra, InceptionTime) and clearly outperforms all other foundation-model baselines. Notably, TimEE achieves the second-best macro-F1 rank on imbalanced datasets, behind only Hydra.

ModelMacro PrecisionMacro RecallMacro F1AccuracyMacro F1 rankMacro F1 (imbalanced)Macro F1 rank (imbalanced)
TimEE (ours).834.823.819.8413.56.7693.35
MiniRocket.844.835.833.8523.30.7703.96
Hydra.842.831.830.8503.39.7793.26
InceptionTime.839.827.825.8433.41.7643.40
TiRex (Untuned).799.778.775.8035.30.7125.47
NuTime (Untuned).760.742.738.7667.02.6746.96
MOMENT (Untuned).760.739.736.7626.97.6836.62
Catch22.755.738.733.7627.27.6716.93
Chronos-2.769.737.732.7646.84.6517.37
TiCT.715.705.698.7277.95.6307.69

We will include these additional results in the appendix of the revised manuscript.


Q3: Consistency of Hyperparameter Settings

"Consistency of hyper-parameter settings"

Thank you for spotting this inconsistency. Table 2 in Appendix B.1 contains an error.

The model used for all reported experiments has Lt=4L_t = 4 temporal attention layers and Lc=4L_c = 4 cross-series attention layers, consistent with Section 3.3 and the base configuration in Table 3.

We will correct Table 2 in the revised manuscript.


Thank you again for your constructive feedback. We hope the additional experiments and clarifications address your concerns and questions.

Kind regards,
The Authors

Thank you for the detailed rebuttal.

  • The clarifications on baseline selection and hyperparameter consistency error correction address my Q1 and Q3.

  • The additional precision, recall, and F1 results correspond to my Q2. However, these results show that TimEE does not outperform SOTA baselines such as MiniRocket, Hydra, and InceptionTime on these classification metrics.

Considering the rebuttal and the remaining concerns about methodological contributions and the classification performance, I'll keep my original rating.

Dear Reviewer 14tv,

Thank you for considering our rebuttal.

We appreciate your remaining concerns regarding methodological contributions and classification performance. We would like to offer two brief clarifications regarding these points.

Regarding TiCT: We agree that both approaches share the high-level paradigm of synthetic pretraining for end-to-end time-series classification. Our contribution lies in advancing this paradigm through a different approach to synthetic task construction and label-aware contextual representation learning. As detailed in our rebuttal, the architectural ablations (Appendix B.4, Tables 3–4) show measurable benefits from conditioning temporal representations on support labels and enabling cross-series interactions. Together with our natively multiclass synthetic prior, the resulting model substantially outperforms TiCT across the reported metrics while using approximately 10× fewer parameters.

Regarding SOTA classification performance: We acknowledge that TimEE does not consistently outperform the strongest supervised baselines on accuracy or macro-F1, and we do not intend to claim otherwise. Nevertheless, our additional critical-difference analyses (Appendix F.2, Figures 17–19) show no statistically significant differences between TimEE and MiniRocket, Hydra, or InceptionTime on balanced accuracy, macro-F1, and macro precision. Notably, TimEE achieves this competitive performance despite being pretrained exclusively on synthetic classification tasks, without requiring any target-dataset training or classifier fitting.

We believe these results support the methodological and empirical contributions of TimEE to end-to-end ICL for time-series classification, while laying the groundwork for future advances in synthetic priors, contextual reasoning, and multivariate classification.

Thank you again for your time and consideration.

Kind Regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair and Reviewers,

Thank you for your constructive feedback and thoughtful engagement. We would like to briefly summarize the main outcomes of our rebuttal.

Methodological contributions. We further clarified TimEE's contributions relative to TiCT, particularly its natively multiclass synthetic prior and label-aware contextual representation learning. Our existing ablations in Appendix B.4 (Tables 3–4) support these design choices. New support-label shuffling and support-set size experiments further demonstrate that TimEE genuinely exploits labeled examples during inference. Also, TimEE substantially outperforms TiCT while using approximately 10× fewer parameters.

Classification performance. We provided additional macro-averaged metrics, including an evaluation on imbalanced datasets, where TimEE achieves the second-best macro-F1 mean rank. While TimEE does not consistently outperform the strongest supervised baselines, the additional critical-difference analyses (Appendix F.2, Figures 17–19) show no statistically significant differences on balanced accuracy, macro-F1, and macro precision.

Synthetic-to-real transfer. We extended our analysis with domain-wise evaluations and qualitative comparisons (Appendix C.5), illustrating the diversity of time-series patterns covered by our prior. While these analyses provide further empirical evidence of transfer, understanding which specific class-discriminative properties enable this transfer remains an important direction for future work.

We believe these additional experiments and clarifications strengthen the evidence for end-to-end ICL with synthetic pretraining as a promising direction for time-series classification.

Thank you again for your time and consideration.

Kind regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.