Pith. sign in

REVIEW 2 major objections 6 minor 27 references

This paper claims that a pretrained tabular foundation model can replace the linear classifier in random-convolutional time series classification, matching the strongest published ensemble on univariate benchmarks without any task-specific

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 ↗

MASHT, which feeds MultiRocket+Hydra features into the pretrained TabPFN-3 model, matches HIVE-COTE 2.0 on univariate time-series classification and is competitive on multivariate benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A clean, honest empirical study showing a frozen tabular foundation model on random convolutional features can match HIVE-COTE 2.0 on univariate TSC; the causal claim about TabPFN-3 is less clean. the 2 major comments →

arxiv 2607.19234 v1 pith:6G3CI5UO submitted 2026-07-21 cs.LG

In-Context Time Series Classification with Random Convolutional Features

Joscha C\"uppers , Jilles Vreeken This is my paper
classification cs.LG
keywords time series classificationrandom convolutional featurestabular foundation modelin-context learningMultiRocketHydrano-training classificationbenchmark evaluation
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 reading

This paper asks whether a pretrained tabular foundation model can replace the simple linear classifiers normally paired with random convolutional time-series features. It proposes MASHT, which extracts MultiRocket and Hydra features and feeds them directly to TabPFN-3, requiring no task-specific training. On 112 univariate datasets MASHT matches the accuracy of HIVE-COTE 2.0, the strongest published ensemble, and significantly beats MR-Hydra, the same feature family with a ridge classifier. On 71 multivariate datasets it is competitive with leading methods but does not lead. The result matters because it suggests a modular, training-free route to accurate time series classification.

Core claim

The central claim is that the pairing of random convolutional feature transforms with an in-context tabular foundation model is itself a competitive time series classifier. Concretely, on the UTF-112 benchmark MASHT reaches a mean accuracy of 0.892 and a mean rank of 3.27, slightly above HIVE-COTE 2.0 (0.891, rank 3.41) and significantly better than MR-Hydra (0.884, rank 3.88) after Holm-corrected Wilcoxon testing. The authors present this as evidence that the tabular model captures nonlinear structure in the transformed features that a linear classifier misses, while the univariate/multivariate contrast suggests the representation is most effective for single-channel series.

What carries the argument

The pipeline's load-bearing components are two random convolutional transforms — MultiRocket, which pools kernel responses from raw and differenced series, and Hydra, which records competing-kernel extreme-response counts — concatenated into a fixed feature table, and TabPFN-3, a pretrained transformer that performs in-context classification on that table without fitting task-specific weights. The feature table turns temporal structure into a standard tabular problem; the foundation model supplies the learned decision rule.

Load-bearing premise

The central premise is that the published baseline results were produced under evaluation protocols and feature budgets comparable to MASHT's, so the measured accuracy differences reflect the classifier rather than the amount of features.

What would settle it

Run MR-Hydra under exactly MASHT's feature-budget protocol (equal Hydra/MultiRocket split, 10,000/2,000/200 features by dataset size) on UTF-112 and check whether its 0.008 accuracy gap to MASHT persists; if it vanishes, the gain attributed to the tabular foundation model is actually a feature-budget effect.

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

If this is right

  • On univariate benchmarks, MASHT matches the accuracy of HIVE-COTE 2.0 while requiring no task-specific training, only feature extraction and pretrained-model inference.
  • Because the classifier is pretrained, new datasets can be classified immediately at inference time once features are computed, which changes the deployment cost profile for time series classification.
  • The significant gain over MR-Hydra indicates that the choice of downstream classifier matters for random convolutional features, not just the transform.
  • On multivariate benchmarks, the pipeline remains competitive but not leading, so the tabular representation does not yet capture cross-channel structure as well as ensemble methods.
  • The reported runtimes suggest a favorable accuracy–runtime trade-off, though the paper notes the comparison is not controlled.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the feature-budget confound is resolved and the gain holds, the same recipe could be applied to other feature transforms and other pretrained tabular models, making 'transform + foundation model' a general template for time series classification.
  • The multivariate shortfall points to a testable extension: adding cross-channel convolution or a second-stage feature interaction layer before the tabular model might close the gap to the leading ensemble on the Multiverse benchmark.
  • The dataset-dependent feature budget (200 features for large data) implies accuracy may degrade at scale; investigating where that boundary bites would clarify the method's practical limits.
  • A controlled speed comparison on identical hardware and datasets would be the natural next evaluation; the paper leaves this open.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 6 minor

Summary. The paper proposes MASHT, a two-stage time-series classification pipeline that concatenates MultiRocket and Hydra random-convolutional features and feeds the resulting table to the pretrained TabPFN-3 model for in-context classification. The central empirical claim is that this training-free pipeline matches state-of-the-art TSC ensembles on univariate benchmarks: on UTF-112 it reports mean accuracy 0.892 (rank 3.27) vs. HIVE-COTE 2.0's 0.891 (rank 3.41), and it is significantly better than MR-Hydra (0.884, p=0.03 after Holm correction) while being statistically tied with HC2. On the 71-dataset Multiverse benchmark, MASHT is competitive but not leading (0.795 accuracy, rank 7.65 vs HC2 0.805/6.51). The paper also reports runtime and limitation analyses, and makes code available.

Significance. If the comparison to MR-Hydra were fully controlled, this would be an important result: it would show that a frozen in-context tabular model can replace a trained linear head on random-convolutional TSC features, with accuracy competitive with a much larger ensemble. The paper's statistical practice is careful: Wilcoxon signed-rank tests with Holm correction, critical-difference diagrams, and per-dataset scatterplots, and the limitations are honestly described. However, the attribution to TabPFN-3 is currently underdetermined because the key comparison changes both the classifier and the feature budget, and the provenance of some baseline metrics is not documented.

major comments (2)
  1. [§4 Experiment Setup and Table 2] MASHT uses a dataset-dependent feature budget (10,000/2,000/200 features, split equally between Hydra and MultiRocket), but the MR-Hydra baseline is imported from Middlehurst et al. (2024) and the manuscript never states that the baseline uses the same feature budget, the same realized feature dimensionality, or the same feature-extraction configuration. The text in §4 says 'obtain all MASHT results under the same dataset protocols', but this statement refers to datasets/resamples and does not control the feature budget. Since the comparison changes both the classifier (Ridge → TabPFN-3) and the representation (full/default → budgeted), the significant 0.008 accuracy gain and the associated Wilcoxon p=0.03 cannot be unambiguously attributed to TabPFN-3. This is load-bearing for the paper's research question and for the §5 statement that 'TabPFN-3 is a promising classifier for the MultiRo
  2. [§4 Table 1] Table 1 reports balanced accuracy and AUROC for all baselines, but the manuscript only says baseline results are 'from Middlehurst, Schäfer, and Bagnall (2024)'. It does not state whether those AUROC/balanced-accuracy values were taken directly from that publication or recomputed, and if recomputed, from which probability outputs and by which protocol. The AUROC gap between MASHT and MR-Hydra (0.970 vs. 0.913) is very large despite the same transform family, which raises the possibility of a protocol mismatch in how AUROC is averaged or which class is treated as positive. Please clarify the source and exact computation, and if the values were recomputed, make the procedure explicit for every baseline.
minor comments (6)
  1. [§1 Introduction] Typo: 'provide a provide broad temporal features' should be 'provide broad temporal features'.
  2. [§3 Feature Extraction] The formula for the number of Hydra groups, g = floor(B/(2kd)), uses d = (log2(L-1))/(9-1). The derivation is terse; please define all variables and explain why the 9-1 term appears (Hydra's kernel length). Also, the realized dimensionality may be 'slightly below' the target; for reproducibility, a table of achieved feature counts across dataset sizes would help.
  3. [§4 Table 5] The runtime table shows mean total runtime lower than the median for both benchmarks (e.g., 43.954 s mean vs. 59.743 s median on UTF-112). Given positive runtimes, this is unusual and suggests a reporting inconsistency. Please verify the numbers or explain the distribution.
  4. [§4 Evaluation] The footnote containing the data link has a line break and the word 'Direktlink' inserted mid-sentence. Please clean up the formatting.
  5. [Appendix A] Please specify the exact TabPFN-3 version/commit, the seed handling, and whether 'automatic estimator scaling' and 'automatic inference precision' produce deterministic results across runs. Also, since the feature-budget thresholds are fixed but ad hoc, a short sensitivity analysis (e.g., how accuracy varies with the budget) would strengthen the paper.
  6. [§5 Runtime] The runtime comparison is appropriately caveated as uncontrolled. However, the conclusion section says 'These results show that tabular foundation models are a promising classifier for random convolutional time-series features' — this should be tempered or explicitly conditioned on the feature-budget caveat.

Circularity Check

0 steps flagged

No circularity: MASHT is an empirical pipeline with external benchmarks, frozen pretrained classifier, and no self-citation loop.

full rationale

The paper makes no derivation-equals-input claim. MASHT combines fixed random-convolutional features (MultiRocket, Hydra) with the frozen pretrained TabPFN-3 and evaluates on established external benchmarks (UTF-112, Multiverse). The feature budget is a fixed function of dataset size, not fitted to benchmark outcomes; TabPFN-3 is not trained on the evaluated datasets; and no parameter is tuned against the reported accuracies. There are no self-citations by the authors, no imported uniqueness theorem, and no ansatz smuggled in via citation to make the conclusion forced. The potential feature-budget mismatch between MASHT and the MR-Hydra baseline is a legitimate experimental-confounding concern, but it does not make the comparison circular: the MR-Hydra results come from an independent external study, and the MASHT results are not constructed from those numbers. The runtime limitation is explicitly acknowledged as not a controlled speed comparison, which is an honest caveat rather than evidence of circularity. Accordingly, the paper is self-contained against external benchmarks and warrants a circularity score of 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

MASHT introduces no new theory or entities; it composes MultiRocket, Hydra, and TabPFN-3. The load-bearing support comes from (i) external baseline comparability, (ii) TabPFN-3's out-of-the-box generalization to these feature tables, and (iii) the information content of the random-convolutional features. The only hand-set values are the feature-budget thresholds and the equal budget split.

free parameters (3)
  • Feature budget thresholds = 10,000 (n<1,000); 2,000 (n<100,000); 200 (otherwise)
    Hand-chosen cutoffs on the combined train+test instance count to bound feature dimensionality; not derived from theory and not ablated, so their effect on the MASHT-vs-baseline comparison is unknown.
  • Hydra/MultiRocket budget split = 50/50
    The feature budget is divided equally between Hydra and MultiRocket; an arbitrary design choice that shapes the representation and could affect which datasets MASHT handles well.
  • TabPFN-3 estimator count = 8
    Number of ensemble estimators in TabPFNClassifier; a fixed implementation setting with no sensitivity analysis reported.
axioms (3)
  • domain assumption Published baseline results (HC2, MR-Hydra, and others) are accurate and directly comparable to MASHT's runs.
    Invoked in Section 4: 'The reference results are taken from the corresponding studies, and obtain all MASHT results under the same dataset protocols'; all comparisons depend on this.
  • domain assumption The frozen TabPFN-3 model generalizes in-context to MultiRocket+Hydra feature tables without adaptation.
    Core hypothesis of the pipeline (Section 3, Eq. 1); the paper provides benchmark evidence but no independent verification of feature-distribution coverage.
  • domain assumption MultiRocket and Hydra features retain sufficient temporal information for competitive TSC.
    Taken from prior work (Tan et al. 2022; Dempster, Schmidt, and Webb 2023); the paper does not re-derive or justify this beyond citation.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of In-Context Time Series Classification with Random Convolutional Features." pith.science (2026). https://pith.science/paper/6G3CI5UO

@misc{pith2026260719234,
  author       = {Pith},
  title        = {Pith review of: In-Context Time Series Classification with Random Convolutional Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6G3CI5UO}},
  note         = {Machine review of arXiv:2607.19234}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Time series classification is central to domains like medical signal analysis, industrial monitoring, and sensor-based activity recognition, where class information manifests as localized shapes, specific frequencies, temporal shifts, or complex cross-channel interactions. Random convolutional transforms efficiently map these sequences to fixed-dimensional tabular features but are traditionally paired with simple linear classifiers. We investigate whether a pretrained tabular foundation model can more effectively harness these rich representations. We propose MASHT, a pipeline that marries MultiRocket and Hydra features with the power of in-context tabular foundation models. By leveraging a pretrained tabular foundation model, our approach completely bypasses task-specific model training, requiring only feature extraction and direct inference. Extensive experiments demonstrate that MASHT matches state-of-the-art time series classification baselines on univariate tasks, achieving a lower average rank than HIVE-COTE 2.0. On multivariate datasets, MASHT remains highly competitive with the strongest reference methods.

Figures

Figures reproduced from arXiv: 2607.19234 by Jilles Vreeken, Joscha C\"uppers.

Figure 1
Figure 1. Figure 1: Critical-difference diagram for accuracy on UTF-112. Methods connected by a horizontal bar are not significantly [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-dataset UTF-112 accuracy comparisons with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Critical-difference diagram for accuracy on the Multiverse benchmark. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-dataset Multiverse accuracy comparisons with [PITH_FULL_IMAGE:figures/full_fig_p005_4.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

27 extracted references · 1 linked inside Pith

  1. [1]

    Bagnall, A.; Lines, J.; Bostrom, A.; Large, J.; and Keogh, E. 2017. The Great Time Series Classification Bake Off: A Review and Experimental Evaluation of Recent Algorithmic Advances. Data Mining and Knowledge Discovery, 31(3): 606--660

    2017

  2. [2]

    Bostrom, A.; and Bagnall, A. 2017. Binary Shapelet Transform for Multiclass Time Series Classification. Transactions on Large-Scale Data- and Knowledge-Centered Systems, 32: 24--46

    2017

  3. [3]

    A.; Bagnall, A.; Kamgar, K.; Yeh, C.-C

    Dau, H. A.; Bagnall, A.; Kamgar, K.; Yeh, C.-C. M.; Zhu, Y.; Gharghabi, S.; Ratanamahatana, C. A.; and Keogh, E. 2019. The UCR Time Series Archive. IEEE/CAA Journal of Automatica Sinica, 6(6): 1293--1305

    2019

  4. [4]

    Dempster, A.; Fran c ois, P.; and Webb, G. I. 2020. ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery, 34(5): 1454--1495

    2020

  5. [5]

    F.; and Webb, G

    Dempster, A.; Schmidt, D. F.; and Webb, G. I. 2021. MiniRocket : A Very Fast (Almost) Deterministic Transform for Time Series Classification. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 248--257

    2021

  6. [6]

    F.; and Webb, G

    Dempster, A.; Schmidt, D. F.; and Webb, G. I. 2023. Hydra: Competing Convolutional Kernels for Fast and Accurate Time Series Classification. Data Mining and Knowledge Discovery, 37: 1779--1805

    2023

  7. [7]

    F.; and Webb, G

    Dempster, A.; Schmidt, D. F.; and Webb, G. I. 2024. quant: a minimalist interval method for time series classification. Data Mining and Knowledge Discovery, 38(4): 2377--2402

    2024

  8. [8]

    o ge, K.; Key, O.; Birkel, F.; Jund, P.; Roof, B.; Manium, M.; Hoo, S. B.; B \

    Grinsztajn, L.; Fl \"o ge, K.; Key, O.; Birkel, F.; Jund, P.; Roof, B.; Manium, M.; Hoo, S. B.; B \"u hler, M.; Garg, A.; et al. 2026. TabPFN-3 : Technical Report. arXiv preprint arXiv:2605.13986

    Pith/arXiv arXiv 2026

  9. [9]

    Guillaume, A.; Vrain, C.; and Elloumi, W. 2022. Random dilated shapelet transform: A new approach for time series shapelets. In International Conference on Pattern Recognition and Artificial Intelligence, 653--664. Springer

    2022

  10. [10]

    B.; Schirrmeister, R

    Hollmann, N.; Muller, S.; Purucker, L.; Krishnakumar, A.; Korfer, M.; Hoo, S. B.; Schirrmeister, R. T.; and Hutter, F. 2025. Accurate Predictions on Small Data with a Tabular Foundation Model. Nature, 637: 319--326

    2025

  11. [11]

    Holm, S. 1979. A simple sequentially rejective multiple test procedure. Scandinavian journal of statistics, 65--70

    1979

  12. [12]

    Ismail-Fawaz, A.; Devanne, M.; Berretti, S.; Weber, J.; and Forestier, G. 2025. Look into the LITE in Deep Learning for Time Series Classification. International Journal of Data Science and Analytics, 20: 4029--4049

    2025

  13. [13]

    Ismail Fawaz, H.; Forestier, G.; Weber, J.; Idoumghar, L.; and Muller, P.-A. 2019. Deep Learning for Time Series Classification: A Review. Data Mining and Knowledge Discovery, 33(4): 917--963

    2019

  14. [14]

    F.; Weber, J.; Webb, G

    Ismail Fawaz, H.; Lucas, B.; Forestier, G.; Pelletier, C.; Schmidt, D. F.; Weber, J.; Webb, G. I.; Idoumghar, L.; Muller, P.-A.; and Petitjean, F. 2020. InceptionTime : Finding AlexNet for Time Series Classification. Data Mining and Knowledge Discovery, 34(6): 1936--1962

    2020

  15. [15]

    H.; Sethi, S

    Lubba, C. H.; Sethi, S. S.; Knaute, P.; Schultz, S. R.; Fulcher, B. D.; and Jones, N. S. 2019. catch22: CAnonical Time-series CHaracteristics selected through highly comparative time-series analysis. bioRxiv

    2019

  16. [16]

    Lucas, B.; Shifaz, A.; Pelletier, C.; O'Neill, L.; Zaidi, N.; Goethals, B.; Petitjean, F.; and Webb, G. I. 2019. Proximity Forest: An Effective and Scalable Distance-Based Classifier for Time Series. Data Mining and Knowledge Discovery, 33(3): 607--635

    2019

  17. [17]

    Middlehurst, M.; and Bagnall, A. 2022. The FreshPRINCE : A Simple Transformation Based Pipeline Time Series Classifier. In International Conference on Pattern Recognition and Artificial Intelligence, 150--161

    2022

  18. [18]

    Middlehurst, M.; and Bagnall, A. 2023. Extracting features from random subseries: A hybrid pipeline for time series classification and extrinsic regression. In International Workshop on Advanced Analytics and Learning on Temporal Data, 113--126. Springer

    2023

  19. [19]

    Middlehurst, M.; Large, J.; and Bagnall, A. 2020. The Canonical Interval Forest (CIF) Classifier for Time Series Classification. In IEEE International Conference on Big Data, 188--195

    2020

  20. [20]

    Middlehurst, M.; Large, J.; Cawley, G.; and Bagnall, A. 2020. The temporal dictionary ensemble (TDE) classifier for time series classification. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 660--676. Springer

    2020

  21. [21]

    Middlehurst, M.; Large, J.; Flynn, M.; Lines, J.; Bostrom, A.; and Bagnall, A. 2021. HIVE-COTE 2.0: a new meta ensemble for time series classification. Machine Learning, 110(11): 3211--3243

    2021

  22. [22]

    I.; Holder, C.; and Bagnall, A

    Middlehurst, M.; Rushbrooke, A.; Ismail-Fawaz, A.; Devanne, M.; Forestier, G.; Dempster, A.; Webb, G. I.; Holder, C.; and Bagnall, A. 2026. The Multiverse of Time Series Machine Learning: An Archive for Multivariate Time Series Classification. arXiv preprint arXiv:2603.20352

    arXiv 2026

  23. [23]

    Middlehurst, M.; Sch \"a fer, P.; and Bagnall, A. 2024. Bake off redux: a review and experimental evaluation of recent time series classification algorithms. Data Mining and Knowledge Discovery, 38(4): 1958--2031

    2024

  24. [24]

    P.; Flynn, M.; Large, J.; Middlehurst, M.; and Bagnall, A

    Ruiz, A. P.; Flynn, M.; Large, J.; Middlehurst, M.; and Bagnall, A. 2021. The Great Multivariate Time Series Classification Bake Off: A Review and Experimental Evaluation of Recent Algorithmic Advances. Data Mining and Knowledge Discovery, 35(2): 401--449

    2021

  25. [25]

    Sch \"a fer, P.; and Leser, U. 2023. WEASEL 2.0: A Random Dilated Dictionary Transform for Fast, Accurate and Memory Constrained Time Series Classification. Machine Learning, 112(12): 4763--4788

    2023

  26. [26]

    Shokoohi-Yekta, M.; Hu, B.; Jin, H.; Wang, J.; and Keogh, E. 2017. Generalizing DTW to the multi-dimensional case requires an adaptive approach. Data mining and knowledge discovery, 31(1): 1--31

    2017

  27. [27]

    W.; Dempster, A.; Bergmeir, C.; and Webb, G

    Tan, C. W.; Dempster, A.; Bergmeir, C.; and Webb, G. I. 2022. MultiRocket : Multiple Pooling Operators and Transformations for Fast and Effective Time Series Classification. Data Mining and Knowledge Discovery, 36: 1623--1646

    2022

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.