Pith. sign in

REVIEW 3 major objections 8 minor 40 references

Synthetic-only model beats all TSC baselines on UCR benchmark

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 08:50 UTC pith:TFI5Y4ET

load-bearing objection TimEE: synthetic-pretrained ICL model for TSC with strong UCR results, but preprocessing ensemble confound needs ablation the 3 major comments →

arxiv 2607.07500 v1 pith:TFI5Y4ET submitted 2026-07-08 cs.LG cs.AI

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

Jaris K\"uken , Shi Bin Hoo , Martin Mr\'az , Frank Hutter , Lennart Purucker This is my paper
classification cs.LG cs.AI
keywords timeeseriestimelearningclassificationend-to-endmodelbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces TimEE, a 4.5M-parameter model that reframes time series classification as in-context learning: given a labeled support set and a query, it outputs a class distribution in a single forward pass with no per-dataset training. The key innovation is a synthetic data generation pipeline based on VARX (Vector Autoregressive with Exogenous inputs) processes, where classification tasks are constructed so that class identity reflects meaningful structural or signal-level differences in the data-generating process. Two mechanisms generate class structure: structural variation (dropping edges from the dependency graph between classes) and signal variation (warping the speed of exogenous inputs). Despite being trained exclusively on these synthetic tasks with no real time series data, TimEE achieves the lowest mean rank on ROC AUC across all 128 UCR benchmark datasets, outperforming both foundation models and supervised deep learning baselines, while also providing the best-calibrated probability estimates and lying on the Pareto front of inference speed versus performance.

Core claim

A model trained exclusively on synthetic classification tasks generated from a VARX-based prior, where class identity arises from controlled variation in the data-generating process itself, can match or exceed the discriminative performance of methods trained on real data across a broad benchmark, while simultaneously eliminating per-dataset training. The end-to-end in-context learning design, which allows the encoder to access labeled support examples during inference, produces better-calibrated probability estimates than two-stage pipelines that decouple representation learning from classification.

What carries the argument

The VARX-based synthetic prior with two class-generation mechanisms (structural variation via edge dropout and signal variation via exogenous input warping) is the central object. The transformer architecture processes all series jointly through temporal attention, cross-series attention, and in-context reasoning layers, with class embeddings injected into support series so the encoder builds class-aware representations that queries align to without receiving label signals.

Load-bearing premise

The VARX-based synthetic prior, with its two class-generation mechanisms (structural edge dropout and exogenous signal warping), generates classification tasks whose class structure is representative enough of real-world TSC diversity to enable generalization across 128 heterogeneous datasets including ECG, gesture, spectra, and image-derived series, despite the prior being grounded in linear autoregressive dynamics.

What would settle it

If TimEE were evaluated on time series classification benchmarks outside UCR that involve strongly nonlinear dynamics, long-range dependencies, or domain-specific patterns not well-approximated by VARX processes, and its mean rank dropped substantially below two-stage methods, this would suggest the prior's coverage is narrower than the UCR results imply.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If a linear autoregressive prior with two variation mechanisms suffices for state-of-the-art TSC, richer nonlinear generative priors could substantially widen the performance gap, suggesting prior design rather than model scale or real data volume is the primary lever.
  • The in-context paradigm eliminates the need for per-dataset classifier fitting, which could streamline deployment in settings where new classification tasks arrive continuously and labeled data is scarce.
  • The strong calibration results (best log-loss rank) suggest that end-to-end ICL produces more honest uncertainty estimates than two-stage pipelines, which matters for downstream decision-making under threshold-sensitive deployment.
  • The concurrent emergence of TiCT using a different synthetic prior (KernelSynth mixup) but the same ICL paradigm suggests the paradigm itself, not the specific prior, is the structural shift for TSC.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. This paper introduces TimEE, a 4.5M-parameter foundation model for time series classification (TSC) that operates via in-context learning (ICL). The model is meta-trained exclusively on synthetic classification tasks generated from a VARX-based prior, in which class identity arises from controlled structural or signal-level variation in the data-generating process. At test time, TimEE accepts a labeled support set and a query, and outputs a class distribution in a single forward pass with no per-dataset training. The authors evaluate on all 128 UCR benchmark datasets against 12 baselines spanning classical, deep learning, and foundation model categories, reporting the lowest mean rank on ROC AUC and third on accuracy. The paper also provides ablations on architecture components (Tables 3–4, Figure 8), a multivariate extension on 24 UEA datasets (Figure 7), and a calibration analysis via log loss (Figure 5).

Significance. The paper makes a genuine methodological contribution by introducing a structured synthetic prior for TSC in which class identity reflects differences in the underlying VARX data-generating process—a non-trivial advance over prior synthetic generation that produced unlabeled corpora or arbitrary labels. The PFN training objective (Eq. 1) is standard and the model sees zero real data during pre-training, making the UCR results a meaningful test of the prior's coverage. The evaluation is comprehensive: 128 UCR datasets, 12 baselines across three method categories, multiple metrics (ROC AUC, accuracy, log loss, inference speed), critical difference diagrams, and full per-dataset results (Tables 6–11). The authors provide ablations isolating temporal vs. cross-series attention (Table 3) and label conditioning sites (Table 4), and the representation evolution analysis (Figure 8) is a nice diagnostic. Code is stated to be publicly available. The core claim—that a purely synthetic-pretrained ICL model can reach state-of-the-art on UCR—is well-supported by the breadth of the evaluation, though one important confound in the headline ranking remains unaddressed (see Major Comments).

major comments (3)
  1. §3.4 and Appendix D.2: TimEE's reported results use an inference-time ensemble of 4 preprocessing variants (interpolation to {128,256,512} and first-order differencing, individually and in composition), with final predictions averaged. No ablation isolates the contribution of this ensemble to TimEE's headline performance. Meanwhile, baselines are not given the same preprocessing ensemble treatment. While some baselines like InceptionTime have their own internal ensembling (5 networks), the preprocessing ensemble is specific to TimEE and operates on a different axis (input transformations rather than model initialization diversity). The mean rank metric is sensitive to small per-dataset shifts, and TimEE's margin over the next-best method on ROC AUC (mean rank 4.66 vs. 5.44 for MiniRocket) could plausibly be accounted for by the ensemble alone. Without an ablation showing TimEE's single-变
  2. §4.1–4.2, Figure 1: The headline claim is that TimEE achieves the lowest mean rank on ROC AUC across all 128 UCR datasets. However, the critical difference diagram (Figure 16) shows that TimEE is not statistically significantly different from InceptionTime (5.39) or MiniRocket (5.44), as they are connected by a horizontal bar. The paper should explicitly acknowledge this in the main text (not only in the appendix) and temper the 'outperforming all compared methods' language accordingly. The win-rate analysis (Figure 4) showing TimEE at 51% vs. InceptionTime is also consistent with a statistical tie. This does not undermine the contribution but requires precise framing.
  3. §3.2, Algorithms 1–2: The VARX prior uses two class-generation mechanisms (structural variation via edge dropout; signal variation via exogenous input warping). The paper does not report what fraction of the 7M synthetic tasks come from each mechanism, nor whether both are necessary for the observed UCR performance. An ablation isolating the contribution of each mechanism (or at minimum reporting the mixture ratio) would clarify whether both are load-bearing or whether one dominates. This is directly relevant to the paper's central claim that the structured prior is the key enabler, since the reader cannot assess which aspect of the prior is responsible.
minor comments (8)
  1. §1, Figure 1 caption: The claim 'outperforming all compared methods' should be qualified given the CD diagram in Figure 16 showing statistical ties with InceptionTime and MiniRocket. Suggest 'achieving the lowest mean rank' rather than 'outperforming all' to avoid overstatement.
  2. §3.3: The architecture description mentions K=4 CLS tokens but does not explain why K=4 was chosen or whether this was tuned. A brief justification or reference to an ablation would help readers.
  3. §3.4: The training uses 7M synthetic tasks from 1.5M unique VARX datasets via augmentation. It would be informative to report the effective diversity (e.g., average number of unique tasks per unique dataset after augmentation) to contextualize the augmentation strategy.
  4. Table 6 (Appendix F.3): The Fungi dataset shows InceptionTime with ROC AUC of 0.008, which appears to be an anomaly. This should be investigated and, if a known failure mode, noted with a footnote.
  5. §4.3, Figure 7: The UEA multivariate results show TimEE (PV) achieving competitive but not top performance. The text attributes the marginal gain from variate pooling to prior limitations, but the per-variate variant itself underperforms several baselines (e.g., MiniRocket, Hydra). This should be acknowledged more directly.
  6. Appendix B.4, Tables 3–4: The ablation uses a scaled-down model (d_kv=16, 2 heads) and evaluates on train splits with cross-validation rather than the standard test splits. The paper should note whether the relative trends observed in ablations are expected to transfer to the full model configuration, or whether there are reasons to expect differences.
  7. §2.3: The related work discusses TiCT as the most directly related concurrent work. The differences are well-articulated in Appendix A.2, but a brief summary in the main text (one or two sentences) would benefit readers who may not consult the appendix.
  8. Figure 8: The cosine similarity heatmaps are informative but the color scale range (0.0–1.0) is not annotated on the axes. Adding scale bars or numeric annotations would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a thorough and constructive report. The referee correctly identifies three substantive issues: (1) the absence of an ablation isolating the preprocessing ensemble's contribution to TimEE's headline performance, (2) imprecise framing of statistical significance in the main text, and (3) missing details on the mixture ratio and individual contributions of the two class-generation mechanisms. We agree with all three points and will revise the manuscript accordingly. Below we address each comment in detail.

read point-by-point responses

  1. Referee: Major Comment 1 (§3.4, Appendix D.2): No ablation isolates the contribution of the 4-variant preprocessing ensemble to TimEE's headline performance, while baselines do not receive the same treatment. The margin over MiniRocket (4.66 vs 5.44 mean rank on ROC AUC) could plausibly be accounted for by the ensemble alone.

    Authors: The referee is correct that the preprocessing ensemble is currently unablated and that this is a meaningful confound in interpreting the headline ranking. We will address this in two ways in the revision. First, we will add an ablation reporting TimEE's single-variant performance (i.e., without the preprocessing ensemble) on all 128 UCR datasets, alongside the ensemble result, so readers can directly assess the ensemble's contribution. Second, we will add a discussion of why the preprocessing ensemble is not directly comparable to the internal ensembling used by baselines like InceptionTime (which averages over model initialization diversity rather than input transformations), and we will acknowledge that applying analogous input-variant ensembling to baselines could narrow the gap. We note that the preprocessing ensemble consists of only 4 lightweight variants (interpolation to {128, 256, 512} and first-order differencing), and the ensemble operates on input transformations rather than model diversity, but we agree the burden of proof is on us to show the ensemble is not the sole driver of the result. If the single-variant ablation shows a substantial drop, we will revise the headline framing accordingly. We cannot, however, retroactively apply the same preprocessing ensemble to all 12 baselines within the revision timeline, as this would require re-running all baselines across all 128 datasets with 4 preprocessing variants each. We will explicitly state this as a limitation. revision: yes

  2. Referee: Major Comment 2 (§4.1–4.2, Figure 1): The headline claim of 'outperforming all compared methods' is not supported by the CD diagram (Figure 16), which shows TimEE is not statistically significantly different from InceptionTime (5.39) or MiniRocket (5.44). The win-rate analysis (51% vs InceptionTime) is also consistent with a statistical tie. The paper should temper its language and acknowledge this in the main text.

    Authors: The referee is correct. The critical difference diagram (Figure 16) shows that TimEE, InceptionTime, and MiniRocket are connected by a horizontal bar and thus not statistically significantly different at alpha = 0.05. The win-rate of 51% against InceptionTime is likewise consistent with a tie. The current language in the main text ('outperforming all compared methods') overstates what the statistical tests support. We will revise the main text (Section 4.2 and the Introduction) to explicitly acknowledge that TimEE achieves the lowest mean rank on ROC AUC but is not statistically significantly different from InceptionTime or MiniRocket, and we will temper the 'outperforming' language to 'achieving the lowest mean rank' or similar. We will also move a reference to the CD diagram into the main text rather than relegating it to the appendix. This does not undermine the core contribution — a purely synthetic-pretrained ICL model achieving competitive performance with strongly tuned per-dataset methods is still a meaningful result — but the framing must be precise, as the referee notes. revision: yes

  3. Referee: Major Comment 3 (§3.2, Algorithms 1–2): The paper does not report what fraction of the 7M synthetic tasks come from each class-generation mechanism (structural variation vs. signal variation), nor whether both are necessary for the observed UCR performance. An ablation isolating each mechanism's contribution is needed.

    Authors: The referee is correct that the mixture ratio is not reported and that no ablation isolates the contribution of each class-generation mechanism. This is a genuine gap given the paper's central claim that the structured prior is the key enabler. We will address this in the revision by: (1) reporting the mixture ratio between structural variation and signal variation tasks in the pre-training corpus (we will state the exact ratio used); and (2) conducting an ablation where we train separate models on each mechanism in isolation and evaluate on the UCR benchmark, so readers can assess whether both mechanisms are load-bearing or whether one dominates. We expect both mechanisms to contribute complementary signal — structural variation captures differences in inter-variable dependency structure, while signal variation captures differences in exogenous input dynamics — but this should be demonstrated empirically rather than assumed. If the ablation reveals that one mechanism dominates, we will revise the paper's framing accordingly. We note that this ablation requires training two additional models from scratch (approximately 40 GPU-hours each on an H200), which is feasible within the revision period. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; derivation is self-contained against external benchmarks

full rationale

The paper's derivation chain is straightforward and non-circular: (1) a VARX-based synthetic prior generates labeled TSC tasks from a structured stochastic process with no fitting to real data; (2) a standard PFN meta-learning objective (Eq. 1) trains the model on these synthetic tasks; (3) the model is evaluated on held-out UCR benchmarks it never saw during training. No prediction or first-principles result reduces to its inputs by construction. The self-citations to TabPFN [Hollmann et al., 2023, 2025] and TabICLv2 [Qu et al., 2026] — with which some authors overlap — are paradigm and architecture references, not load-bearing mathematical theorems invoked to forbid alternatives or force the central claim. The PFN training objective is standard and independently verifiable. The inference-time preprocessing ensemble (Section 3.4, Appendix D.2) raises legitimate fairness concerns (baselines do not receive the same treatment), but this is a correctness risk, not circularity — the ensemble does not make the prediction equivalent to a fitted input by definition. The Takens' theorem reference in Appendix C.1.3 is offered as intuition for the multivariate limitation, not as a derivation that forces the empirical result. Score 1 reflects the minor self-citation to the PFN paradigm, which is not load-bearing for the paper's specific claims.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 2 invented entities

See above.

free parameters (9)
  • Patch size p = 16
    Architectural hyperparameter chosen by the authors, not derived from theory.
  • Model dimension d = 128
    Architectural hyperparameter.
  • Temporal layers L_t = 5
    Architectural hyperparameter; ablation in Table 3 shows performance is stable across allocations.
  • Cross-series layers L_c = 5
    Architectural hyperparameter.
  • ICL layers L_icl = 2
    Architectural hyperparameter.
  • CLS tokens K = 4
    Number of summary tokens per series; chosen by design.
  • Max classes C_max = 10
    Training limit; OvR used beyond this. Covers majority of UCR datasets.
  • Number of ensemble members = 4
    Inference-time ensemble of preprocessing variants; chosen empirically.
  • VARX prior hyperparameters (p_max, s_max, d_min, d_max, q_max, etc.) = Not all explicitly stated in main text
    Control the synthetic data distribution; some values deferred to code.
axioms (4)
  • domain assumption VARX processes are a sufficiently general family to generate synthetic time series whose class structure generalizes to real TSC datasets.
    Section 3.2 introduces the VARX-based DGP. The paper assumes linear autoregressive dynamics with exogenous inputs can represent the diversity of patterns in real TSC data. This is the foundational assumption of the entire approach.
  • domain assumption Class identity arising from structural variation (edge dropout in dependency graph) or signal variation (warping of exogenous input) captures meaningful class differences in real TSC tasks.
    Section 3.2, 'Class Identity Mechanism'. The paper assumes these two mechanisms cover the space of discriminative patterns. No theoretical justification is provided for why these specific mechanisms generalize.
  • standard math The PFN training objective (amortized prediction over synthetic tasks) converges to a useful in-context classification prior.
    Section 3.1, Eq. 1. This is the standard PFN objective from Müller et al. 2024, well-established in the tabular ICL literature.
  • domain assumption Mean rank across UCR datasets is a reliable metric for comparing TSC methods.
    Section 4. The paper uses mean rank as the primary comparison metric. This is standard in the TSC literature but can be sensitive to the specific baseline set and dataset composition.
invented entities (2)
  • VARX structural variation prior independent evidence
    purpose: Generates synthetic labeled TSC tasks where class identity is determined by edge dropout in the VARX dependency graph.
    The prior generates data from a well-defined stochastic process (Algorithm 1) and its utility is tested on 128 real UCR datasets. The class-discriminative mechanism is falsifiable: if it doesn't generalize, UCR performance would be at chance.
  • VARX signal variation prior independent evidence
    purpose: Generates synthetic labeled TSC tasks where class identity is determined by warping speed of the exogenous input.
    Algorithm 2 specifies the generation procedure. Tested on real benchmarks. Falsifiable in the same way as the structural variation prior.

pith-pipeline@v1.1.0-glm · 43265 in / 5525 out tokens · 305519 ms · 2026-07-09T08:50:57.109741+00:00 · methodology

0 comments
read the original 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 directly outputs a predicted class distribution in a single forward pass with no per-dataset training required. 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 ranks first in ROC AUC (and third on accuracy) on the UCR benchmark among all compared methods, which include both foundation models and supervised deep learning baselines. To our knowledge, TimEE is the first purely synthetic-pretrained model to reach state-of-the-art performance on the UCR benchmark. 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. Code is publicly available at http://github.com/automl/timee.

Figures

Figures reproduced from arXiv: 2607.07500 by Frank Hutter, Jaris K\"uken, Lennart Purucker, Martin Mr\'az, Shi Bin Hoo.

Figure 1
Figure 1. Figure 1: We compare mean rank across all 128 UCR benchmark datasets. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Class generation mechanisms in the TIMEE prior: Left – Structural Variation: samples from Class 1 (blue) are generated under the full graph, while samples from Class 2 (orange) are generated under a graph with a subset of edges dropped (×); and Right – Signal Variation: Class 1 (blue) and Class 2 (orange) share the same underlying signal pattern, and class identity is determined by the warping speed of the… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of TIMEE. Label Conditioning. A key advantage of the end-to-end paradigm is that class labels are available throughout the forward pass. We exploit this by conditioning the en￾coder directly on label identity. Concretely, following TabICLv2 [Qu et al., 2026], a learned class embedding is added to each training series’ patch representations, while query series receive no such signal. This allow… view at source ↗
Figure 4
Figure 4. Figure 4: Win rate over ROC AUC across 128 UCR benchmark datasets. TIMEE ranks joint first alongside INCEP￾TIONTIME, outperforming all other com￾pared methods. Notably, TIMEE outper￾forms all existing pre-trained foundation models by a significant margin. We evaluate TIMEE against a broad set of state-of-the-art baselines for time series classification, assessing predictive performance, calibration, and inference ef… view at source ↗
Figure 5
Figure 5. Figure 5: Mean rank on log loss across UCR datasets with maximum 10 classes (n = 103). TIMEE achieves the lowest mean rank, indicating well-calibrated pre￾dictive distributions across datasets. 10 0 10 1 10 2 10 3 Mean Total Time (seconds, log scale) 76 78 80 82 84 86 88 Mean Raw Accuracy (%) 10 Classes (n=103) Catch22 MiniRocket Hydra InceptionTime TS2Vec TICT Chronos-2 TiRex Moment NuTime MantisV2 T EE Statistical… view at source ↗
Figure 7
Figure 7. Figure 7: Mean rank over ROC AUC on 24 UEA datasets. We compare two vari￾ants of TIMEE multivariate extensions: (i) TIMEE (PV) denotes per-variate vari￾ant; (ii) TIMEE (VP) denotes variate￾pooling variant. TIMEE reaches compet￾itive results despite being primary devel￾oped for univariate setting. Inference Speed. TIMEE’s end-to-end ICL mechanism enables fast inference without any per-dataset training. While many oth… view at source ↗
Figure 8
Figure 8. Figure 8: Representation alignment across encoder stages. The leftmost panel shows six time series [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Dataset generated using structural variation mechanism. Lift gain from using all variates over best single variate [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Dataset generated using structural variation mechanism. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Dataset generated using structural variation mechanism. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Dataset generated using signal variation mechanism. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Dataset generated using signal variation mechanism. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Pareto front of inference speed vs. predictive performance on UCR datasets with more [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Critical difference diagram over ROC AUC across 128 UCR datasets (Wilcoxon–Holm [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Critical difference diagram over ROC AUC across 128 UCR datasets (Wilcoxon–Holm [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 12 internal anchors

  1. [1]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Ali Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Bernie Wang. Chronos: Learning the Langu...

    work page 2024

  2. [2]

    Chronos-2: From Univariate to Universal Forecasting

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael B...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Pre-trained forecasting models: Strong zero-shot feature extractors for time series classification, 2025 a

    Andreas Auer, Daniel Klotz, Sebastinan Böck, and Sepp Hochreiter. Pre-trained forecasting models: Strong zero-shot feature extractors for time series classification, 2025 a . URL https://arxiv.org/abs/2510.26777

    work page arXiv 2025

  4. [4]

    Tirex: Zero-shot forecasting across long and short horizons with enhanced in-context learning, 2025 b

    Andreas Auer, Patrick Podest, Daniel Klotz, Sebastian Böck, Günter Klambauer, and Sepp Hochreiter. Tirex: Zero-shot forecasting across long and short horizons with enhanced in-context learning, 2025 b . URL https://arxiv.org/abs/2505.23719

    work page arXiv 2025

  5. [5]

    The UEA multivariate time series classification archive, 2018

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. The uea multivariate time series classification archive, 2018, 2018. URL https://arxiv.org/abs/1811.00075

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    The UCR Time Series Classification Archive , July 2015

    Yanping Chen, Eamonn Keogh, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen, and Gustavo Batista. The UCR Time Series Classification Archive , July 2015

    work page 2015

  7. [7]

    ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels

    Angus Dempster, Fran c ois Petitjean, and Geoffrey I Webb. Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. arXiv preprint arXiv:1910.13051, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  8. [8]

    Minirocket: A very fast (almost) deterministic transform for time series classification

    Angus Dempster, Daniel F Schmidt, and Geoffrey I Webb. Minirocket: A very fast (almost) deterministic transform for time series classification. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 248--257, 2021

    work page 2021

  9. [9]

    Hydra: competing convolutional kernels for fast and accurate time series classification: A

    Angus Dempster, Daniel F Schmidt, and Geoffrey I Webb. Hydra: competing convolutional kernels for fast and accurate time series classification: A. dempster et al. Data Mining and Knowledge Discovery, 37 0 (5): 0 1779--1805, 2023

    work page 2023

  10. [10]

    Rethinking Zero - Shot Time Series Classification : From Task -specific Classifiers to In - Context Inference , January 2026

    Juntao Fang, Shifeng Xie, Shengbin Nie, Yuhui Ling, Yuming Liu, Zijian Li, Keli Zhang, Lujia Pan, Themis Palpanas, and Ruichu Cai. Rethinking Zero - Shot Time Series Classification : From Task -specific Classifiers to In - Context Inference , January 2026. URL http://arxiv.org/abs/2602.00620. arXiv:2602.00620 [cs]

    work page arXiv 2026

  11. [11]

    Mantis: Lightweight Foundation Model for Time Series Classification

    Vasilii Feofanov, Songkang Wen, Marius Alonso, Romain Ilbert, Hongbo Guo, Malik Tiomoko, Lujia Pan, Jianfeng Zhang, and Ievgen Redko. Mantis: Lightweight Calibrated Foundation Model for User - Friendly Time Series Classification , February 2025. URL http://arxiv.org/abs/2502.15637. arXiv:2502.15637 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Mantisv2: Closing the zero-shot gap in time series classification with synthetic data and test-time strategies, 2026

    Vasilii Feofanov, Songkang Wen, Jianfeng Zhang, Lujia Pan, and Ievgen Redko. Mantisv2: Closing the zero-shot gap in time series classification with synthetic data and test-time strategies, 2026. URL https://arxiv.org/abs/2602.17868

    work page arXiv 2026

  13. [13]

    MOMENT : a family of open time-series foundation models

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. MOMENT : a family of open time-series foundation models. In Proceedings of the 41st International Conference on Machine Learning , volume 235 of ICML '24 , pages 16115--16152, Vienna, Austria, July 2024. JMLR.org

    work page 2024

  14. [14]

    TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

    Léo Grinsztajn, Klemens Flöge, Oscar Key, Felix Birkel, Philipp Jund, Brendan Roof, Benjamin Jäger, Dominik Safaric, Simone Alessi, Adrian Hayler, Mihir Manium, Rosen Yu, Felix Jablonski, Shi Bin Hoo, Anurag Garg, Jake Robertson, Magnus Bühler, Vladyslav Moroshan, Lennart Purucker, Clara Cornu, Lilly Charlotte Wehrhahn, Alessandro Bonetto, Bernhard Schölk...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Harris, K

    C. Harris, K. Millman, S. van der Walt, R. Gommers, P. Virtanen, D. Cournapeau, E. Wieser, J. Taylor, S. Berg, N. Smith, R. Kern, M. Picus, S. Hoyer, M. van Kerkwijk, M. Brett, A. Haldane, J. del Río, M. Wiebe, P. Peterson, P. Gérard-Marchant, K. Sheppard, T. Reddy, W. Weckesser, H. Abbasi, C. Gohlke, and T. Oliphant. Array programming with numpy. Nature,...

    work page 2020

  16. [16]

    Hollmann, S

    N. Hollmann, S. M \"u ller, K. Eggensperger, and F. Hutter. Tab PFN : A transformer that solves small tabular classification problems in a second. In The Eleventh International Conference on Learning Representations ( ICLR '23) . ICLR, 2023

    work page 2023

  17. [17]

    u ller, L. Purucker, A. Krishnakumar, M. K \

    N. Hollmann, S. M \"u ller, L. Purucker, A. Krishnakumar, M. K \"o rfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model. Nature, 637 0 (8045): 0 319--326, 2025

    work page 2025

  18. [18]

    Schmidt, Jonathan Weber, Geoffrey I

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F. Schmidt, Jonathan Weber, Geoffrey I. Webb, Lhassane Idoumghar, Pierre-Alain Muller, and François Petitjean. Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery, 34 0 (6): 0 1936–1962, 2020. ISSN 1573-756X. doi:10.1007/s1061...

    work page doi:10.1007/s10618-020-00710-y 1936

  19. [19]

    Muon: An optimizer for hidden layers in neural networks, 2024

    Keller Jordan, Yuchen Jin, Vlado Boza, You Jiacheng, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks, 2024. URL https://kellerjordan. github. io/posts/muon, 6 0 (3): 0 4, 2024

    work page 2024

  20. [20]

    Data augmentation with suboptimal warping for time-series classification

    Krzysztof Kamycki, Tomasz Kapuscinski, and Mariusz Oszust. Data augmentation with suboptimal warping for time-series classification. Sensors, 20 0 (1): 0 98, 2019

    work page 2019

  21. [21]

    NuTime: Numerically Multi-Scaled Embedding for Large-Scale Time-Series Pretraining

    Chenguo Lin, Xumeng Wen, Wei Cao, Congrui Huang, Jiang Bian, Stephen Lin, and Zhirong Wu. Nutime: Numerically multi-scaled embedding for large-scale time-series pretraining, 2024. URL https://arxiv.org/abs/2310.07402

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URL https://arxiv.org/abs/1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  23. [23]

    catch22: CAnonical Time-series CHaracteristics

    Carl H Lubba, Sarab S Sethi, Philip Knaute, Simon R Schultz, Ben D Fulcher, and Nick S Jones. catch22: Canonical time-series characteristics, 2019. URL https://arxiv.org/abs/1901.10200

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  24. [24]

    New Introduction to Multiple Time Series Analysis

    Helmut Lütkepohl. New Introduction to Multiple Time Series Analysis . Springer, Berlin, Heidelberg, 2005. ISBN 978-3-540-40172-8 978-3-540-27752-1. doi:10.1007/978-3-540-27752-1. URL http://link.springer.com/10.1007/978-3-540-27752-1

    work page doi:10.1007/978-3-540-27752-1 2005

  25. [25]

    aeon: a python toolkit for learning from time series

    Matthew Middlehurst, Ali Ismail-Fawaz, Antoine Guillaume, Christopher Holder, David Guijo-Rubio, Guzal Bulatova, Leonidas Tsaprounis, Lukasz Mentel, Martin Walter, Patrick Sch \"a fer, and Anthony Bagnall. aeon: a python toolkit for learning from time series. Journal of Machine Learning Research, 25 0 (289): 0 1--10, 2024. URL http://jmlr.org/papers/v25/2...

    work page 2024

  26. [26]

    Transformers Can Do Bayesian Inference

    Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers can do bayesian inference, 2024. URL https://arxiv.org/abs/2112.10510

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Paszke, S

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic differentiation in PyTorch . In A. Wiltschko, B. van Merriënboer , and P. Lamblin, editors, N eur IPS Autodiff Workshop , 2017

    work page 2017

  28. [28]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in P ython. 12: 0 2825--2830, 2011

    work page 2011

  29. [29]

    Provost, Tom Fawcett, and Ron Kohavi

    Foster J. Provost, Tom Fawcett, and Ron Kohavi. The case against accuracy estimation for comparing induction algorithms. In Proceedings of the Fifteenth International Conference on Machine Learning, ICML '98, page 445–453, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. ISBN 1558605568

    work page 1998

  30. [30]

    Structural Classification of Locally Stationary Time Series Based on Second-order Characteristics

    Chen Qian, Xiucai Ding, and Lexin Li. Structural classification of locally stationary time series based on second-order characteristics. arXiv preprint arXiv:2507.04237, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    TabICL: A Tabular Foundation Model for In-Context Learning on Large Data

    Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. Tabicl: A tabular foundation model for in-context learning on large data, 2025. URL https://arxiv.org/abs/2502.05564

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Tabiclv2: A better, faster, scalable, and open tabular foundation model, 2026

    Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. Tabiclv2: A better, faster, scalable, and open tabular foundation model, 2026. URL https://arxiv.org/abs/2602.11139

    work page arXiv 2026

  33. [33]

    Dynamic programming algorithm optimization for spoken word recognition

    Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26 0 (1): 0 43--49, 1978

    work page 1978

  34. [34]

    The trade-off between universality and label efficiency of representations from contrastive learning

    Zhenmei Shi, Jiefeng Chen, Kunyang Li, Jayaram Raghuram, Xi Wu, Yingyu Liang, and Somesh Jha. The trade-off between universality and label efficiency of representations from contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=rvsbw2YthH_

    work page 2023

  35. [35]

    Roformer: Enhanced transformer with rotary position embedding, 2021

    Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021

    work page 2021

  36. [36]

    Detecting strange attractors in turbulence

    Floris Takens. Detecting strange attractors in turbulence. In Dynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80, pages 366--381. Springer, 2006

    work page 1980

  37. [37]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art...

    work page 2020

  38. [38]

    CauKer : classification time series foundation models can be pretrained on synthetic data only, August 2025

    Shifeng Xie, Vasilii Feofanov, Marius Alonso, Ambroise Odonnat, Jianfeng Zhang, Themis Palpanas, and Ievgen Redko. CauKer : classification time series foundation models can be pretrained on synthetic data only, August 2025. URL http://arxiv.org/abs/2508.02879. arXiv:2508.02879 [cs]

    work page arXiv 2025

  39. [39]

    Tict: A synthetically pre-trained foundation model for time series classification, 2025

    Chin-Chia Michael Yeh, Uday Singh Saini, Junpeng Wang, Xin Dai, Xiran Fan, Jiarui Sun, Yujie Fan, and Yan Zheng. Tict: A synthetically pre-trained foundation model for time series classification, 2025. URL https://arxiv.org/abs/2511.19694

    work page arXiv 2025

  40. [40]

    TS2Vec: Towards Universal Representation of Time Series

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. Ts2vec: Towards universal representation of time series, 2022. URL https://arxiv.org/abs/2106.10466

    work page internal anchor Pith review Pith/arXiv arXiv 2022