Pith. sign in

REVIEW 3 major objections 3 minor 52 references

RATFM: Retrieval-augmented Time Series Foundation Model for Anomaly Detection

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RATFM shows that a pretrained time-series foundation model fine-tuned once on diverse domains can use a retrieved normal example at test time to detect anomalies on an unseen domain, reaching performance comparable to in-domain fine-tuning.

desk verdict Retrieval-augmented fine-tuning for TSFM anomaly detection is a real contribution, but without a kNN copy baseline the paper can't claim the learned integration is what drives the gains. read the letter →

arxiv 2506.02081 v1 pith:WRRK5KFK submitted 2025-06-02 cs.LG cs.AI

Chihiro Maru , Shoetsu Sato This is my paper
classification cs.LGcs.AI
keywords timeseriesanomalydetectionretrieval-augmentedfoundationmodelstest-timeadaptationin-contextlearningcross-correlationretrievalUCRArchivemovingaveragepost-processing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

RATFM asks whether a pretrained time-series foundation model can be adapted to a brand-new domain at test time by being shown one good example, the way a large language model is adapted with an in-context example. The paper's answer is yes, provided the example-using skill is taught explicitly: fine-tuning the model on many domains with inputs that concatenate a retrieved similar series, its known future, and the target series makes the model treat the example's future as a template for the target's future. On the nine-domain UCR Anomaly Archive this lifts VUS-ROC from 68.7% (Time-MoE) and 70.6% (Moment) in zero-shot to 76.1% and 74.3%, approaching the in-domain fine-tuning upper bounds of 79.1% and 77.4% while requiring no domain-specific training. The paper also shows that a large share of the gain comes from a post-processing step: smoothing the raw deviation scores with a moving average whose window is the estimated period of the series.

What carries the argument

The carried mechanism is the example-conditioned forecasting input $X_{\text{example}}^{(d)}(1:T) \oplus X_{\text{example}}^{(d)}(T+1:T+H) \oplus X^{(d)}(1:T)$, trained to forecast $X^{(d)}(T+1:T+H)$ under a mean-squared-error loss across all training domains. The example is chosen by the cross-correlation coefficient between input segments, following the retrieval practice of earlier time-series work. After fine-tuning, the model treats the example's known future as a template for the target's unknown future; at test time the anomaly score is the absolute difference between forecast and observation. The second component is a scoring pipeline: the raw deviation scores are smoothed by a simple moving average whose window length is the Fourier-estimated period of the series, which removes the false-positive peaks that otherwise dominate the anomaly score.

What would settle it

On one domain from the UCR Anomaly Archive, keep RATFM's retrieval ranking by input cross-correlation but replace each retrieved example's future segment with a phase-shifted or unrelated segment from the same series; if VUS-ROC stays near the reported 76% rather than collapsing toward the zero-shot level, the claim that the example future carries the information is false.

Watch

Extended reading notes

Core claim

The central claim is that retrieval-augmented fine-tuning on diverse domains gives time-series foundation models a domain-independent ability to use a retrieved normal example as a forecasting reference, and that this test-time reference is enough for anomaly detection on unseen domains. The evidence is the UCR Anomaly Archive experiment in which Time-MoE and Moment with RATFM outperform both zero-shot and out-domain fine-tuning and approach in-domain fine-tuning, with bootstrap estimates of 76.2% ± 1.4% and 74.3% ± 1.5% VUS-ROC. The paper also establishes that the raw anomaly score, the absolute deviation of forecast from observation, is a poor detector on its own because periodic peaks generate false positives, and that smoothing the scores by the series' estimated period recovers most of the method's performance. A secondary finding is that vanilla pretrained models do not naturally use examples: feeding the same retrieval-augmented input without the fine-tuning step gives 65.9% VUS-ROC with Time-MoE, below zero-shot, so the example-interpretation ability has to be trained in.

Load-bearing premise

At test time there must be a pool of normal time series from the target domain from which a similar example with a reliable future segment can be retrieved, and the model's fine-tuned ability to use that example's future as a forecast template must transfer to the new domain.

Editorial extensions

If this is right

  • Domain-specific fine-tuning becomes unnecessary for anomaly detection: one generic fine-tune on diverse domains plus a small pool of normal examples per deployment replaces per-domain retraining.
  • The example-using capability must be learned explicitly; retrieval-augmented inputs alone do not help a zero-shot time-series foundation model, since RATFM without training underperforms the zero-shot baseline.
  • Anomaly scoring matters as much as model capacity: smoothing deviation scores with the series' estimated period removes periodic-peak false positives and accounts for a substantial share of the reported gains.
  • Reconstruction-based use of the same recipe fails for Moment because the anomalous interval is included in the input and gets faithfully reproduced, so forecast-based conditioning is the direction that works.
  • Performance degrades gracefully as the candidate pool shrinks: at 25% of the original candidates, RATFM stays above the zero-shot and out-domain fine-tuning baselines in the tested domains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the paper attributes the gain to copying the example's future segment, a direct test is to hold retrieved inputs fixed but shuffle the example futures among candidates; if performance tracks future similarity rather than input similarity, the copying mechanism is confirmed.
  • The same retrieval-conditioned fine-tuning could extend beyond anomaly detection to other forecasting-facing tasks such as imputation or event detection wherever a reference window from the same domain is available; the paper lists this as future work but does not test it.
  • The need for a screened pool of normal candidates is an operational condition: retrieval by input similarity alone can pick an example whose future diverges from the target's normal pattern, and the paper's error analysis shows exactly that failure mode, implying that retrieval should ideally score future similarity too.
  • The Fourier-period-windowed moving average is domain-agnostic but depends on periodicity; for aperiodic or non-stationary series the period estimate is undefined, so an adaptive or learned smoothing window is a natural extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes RATFM, a retrieval-augmented test-time adaptation scheme for pretrained time series foundation models in anomaly detection. During training, the model is fine-tuned on out-domain data with inputs augmented by a retrieved same-domain example consisting of the example's input segment and its future segment. At test time, the model retrieves a similar example from other time series in the target domain and uses the concatenated input to forecast the target's future; anomalies are scored by absolute forecast error, post-processed with a simple moving average. Experiments on the UCR Anomaly Archive with Time-MoE and Moment report VUS-ROC values around 76.1% and 74.3%, respectively, which exceed zero-shot and out-domain fine-tuning and approach, but do not match, in-domain fine-tuning.

Significance. If the central claim is established, the work is a useful step toward adapting time series foundation models to new domains without gradient-based fine-tuning. The paper reports experiments across nine domains, includes bootstrap confidence intervals in Table 10, provides a detailed error analysis (Section 6.3), and commits to releasing code and results. The simple moving average post-processing is a practical and well-motivated contribution. However, the significance is currently limited by an incompletely controlled experimental design: the headline gains could be attributable to the retrieval pool itself rather than to the learned integration of examples, and the 'comparable to in-domain fine-tuning' wording is stronger than the reported numbers support.

major comments (3)
  1. [Section 3.1, Table 4] The paper does not include a kNN copy-forecast baseline, which is load-bearing for the central claim that RATFM's learned integration of retrieved examples is responsible for the improvement. In Eq. (2), the model receives the example's future segment X_example(T+1:T+H) directly, and Table 4 shows that this segment has average similarity 0.974 to the target future for Time-MoE. For periodic UCR series, simply outputting the retrieved example's future as the forecast would yield small errors on normal data and large errors on anomalies, hence a high VUS-ROC without any learned component. RATFM w/o training does not control for this because the zero-shot TSFM is not trained to copy the example; the gain of RATFM over RATFM w/o training could merely reflect learning to copy the supplied future. The authors should add a baseline that predicts by directly copying the retrieved example's future (appropriately rescaled to the target's standardization) and evaluate it with the same VUS-ROC/VUS-PR protocol, or otherwise demonstrate that the learned component outperforms such retrieval-only forecasts.
  2. [Section 5.2, Table 10] The claim that RATFM achieves performance 'comparable to that of In-domain FT' is overstated. In Table 10, Time-MoE RATFM attains VUS-ROC 76.2% ± 1.4% against 79.0% ± 1.3% for In-domain FT, and Moment 74.3% ± 1.5% against 77.4% ± 1.7%, a consistent gap of roughly 3 points. The domain-wise results in Table 8 show larger gaps in several domains (e.g., ECG 73.9 vs. 81.3 and Power Demand 68.4 vs. 78.4 for Time-MoE). 'Comparable' is too strong; the authors should either rephrase to 'approaches' or provide a paired significance test (e.g., paired bootstrap or Wilcoxon signed-rank over the 250 time series) demonstrating that the residual gap is not meaningful.
  3. [Section 4.3, Section 3.1] The 'unseen domain' protocol uses test-time retrieval from other time series in the same target domain, so the method is not fully domain-independent as the abstract and introduction suggest. The model is never fine-tuned on the target domain, but it does receive in-domain normal examples at test time; the practical requirement of a pool of normal target-domain series is acknowledged in Section 6.2 but should be stated explicitly in the contributions and abstract. This is not a fatal flaw, but it changes the scope of the claim: RATFM is a test-time adaptation method that requires in-domain normal data, not a zero-shot method that operates without any target-domain information. The authors should clarify this in the framing and in the comparison to in-domain fine-tuning.
minor comments (3)
  1. [Section 3.1, Eq. (1)] The definition of the cross-correlation coefficient is ambiguous: 'CC' is described as a sequence, while the name 'coefficient' suggests a scalar; the authors should specify whether the denominator normalizes the entire sequence and whether the maximum over lags is taken, as in the k-Shape similarity [30].
  2. [Table 8, Reference [8]] There is a domain-name inconsistency: Table 8 lists 'Atrial BP' while Table 5 and the text use 'Arterial BP'; also, reference [8] misspells 'Efron' as 'Efforn.'
  3. [Table 1] The notation for Moment 'RATFM (reconstruction)' input length (160 + 96) + (160 + 96) is cryptic; specifying which parts correspond to the example input, example future, target input, and target future would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RATFM's forecast target is not contained in the retrieved input, and no parameter is fitted to test labels.

full rationale

The derivation chain is self-contained. Eq. (2) defines the RATFM input as X_example(1:T) concatenated with X_example(T+1:T+H) and X(1:T), and Eq. (3) minimizes MSE against the target future X(T+1:T+H). The example future is an input, but the target future is not; retrieval selects an example by cross-correlation of the input segments, not by leaking the target future. Fine-tuning uses only the eight out-domain groups, and test-time retrieval supplies normal examples from the target domain, not labels or fitted values. The high similarity in Table 4 (0.974) between the example future and the forecast target is an empirical observation reported by the paper, and it does not make the target equal to the input by construction. The absence of a kNN copy-forecast baseline is a possible correctness/attribution gap, but it is not a circular step under the definitions used here. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known result; the SMA post-processing is a data-dependent smoothing choice, not a fitted prediction. Therefore the paper does not reduce to its inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two free parameters (SMA window and input allocation) and three domain assumptions about the benchmark and generalization. No new physical or conceptual entities are introduced. The tested setting assumes the availability of a target-domain example pool, which is a practical requirement not stated in the abstract.

free parameters (2)
  • SMA window size n = per-series period estimated via Fourier transform
    Used in Eq. (5) to smooth anomaly scores; the window is set to the estimated period of each time series, so it is a per-series adaptive parameter that materially affects results (Table 2).
  • Input length allocation = e.g., 512+96+512 for Time-MoE RATFM
    The split of the model's maximum context into example input, example future, and target input is a design choice that affects how much context the model sees; it is fixed to balance the components.
assumptions (3)
  • domain assumption The UCR Anomaly Archive provides sufficient normal training series per domain for retrieval candidates
    Section 4.1 and 4.3 rely on having multiple time series per domain to retrieve examples; if a domain had one series, retrieval would be impossible.
  • standard math Cross-correlation is an appropriate similarity measure for retrieving relevant examples
    Section 3.1 uses the cross-correlation coefficient to rank candidates, following prior work; this assumes shape similarity is indicative of predictive usefulness.
  • domain assumption The fine-tuned ability to use examples learned on out-domain data transfers to unseen domains
    Section 4.3 trains RATFM only on the eight non-target domains and evaluates on the held-out domain; this assumes generalization of the retrieval-augmented behavior across domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RATFM: Retrieval-augmented Time Series Foundation Model for Anomaly Detection." pith.science (2026). https://pith.science/paper/WRRK5KFK

@misc{pith2026250602081,
  author       = {Pith},
  title        = {Pith review of: RATFM: Retrieval-augmented Time Series Foundation Model for Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WRRK5KFK}},
  note         = {Machine review of arXiv:2506.02081}
}
read the original abstract

Inspired by the success of large language models (LLMs) in natural language processing, recent research has explored the building of time series foundation models and applied them to tasks such as forecasting, classification, and anomaly detection. However, their performances vary between different domains and tasks. In LLM-based approaches, test-time adaptation using example-based prompting has become common, owing to the high cost of retraining. In the context of anomaly detection, which is the focus of this study, providing normal examples from the target domain can also be effective. However, time series foundation models do not naturally acquire the ability to interpret or utilize examples or instructions, because the nature of time series data used during training does not encourage such capabilities. To address this limitation, we propose a retrieval augmented time series foundation model (RATFM), which enables pretrained time series foundation models to incorporate examples of test-time adaptation. We show that RATFM achieves a performance comparable to that of in-domain fine-tuning while avoiding domain-dependent fine-tuning. Experiments on the UCR Anomaly Archive, a multi-domain dataset including nine domains, confirms the effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 2506.02081 by the authors.

Figure 1
Figure 1. Overview of the retrieval-augmented time series founda [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Training and test split for In-domain FT. The pretrained model was fine-tuned from the target domain data. Specifically, the time series in each domain were divided into two groups, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Anomaly scores of Time-MoE after applying SMA. The time series is the same as that in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Target (bottom) and retrieved example (top) in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Segments used for comput￾ing similarity with the forecast target. 67.7 70.4 71.2 71.4 73.0 81.3 60.8 65.8 66.3 70.4 74.1 80.9 50 60 70 80 Zero-shot Out-domain FT RATFM 25% RATFM 50% RATFM 75% RATFM 100% VUS -ROC% ECG Gait [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Three error cases in anomaly detection using [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Example of a prompt in GPT-4o. 6 4 2 0 2 Value Retrieved example Input time series 0 50 100 150 200 250 Time step 6 4 2 0 2 Value Target Input time series Reconstruction 2 1 0 1 2 3 Value Retrieved example Input time series 0 50 100 150 200 250 Time step 2 1 0 1 2 3 Va…
Figure 10
Figure 10. Figure 10: Target (bottom) and retrieved example (top) in [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Target (bottom) and retrieved example (top) under the [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Time series forecasting results (top) and anomaly scores after applying SMA (bottom) [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

  1. [1]

    Springer International Publishing, 2 edn

    Aggarwal, C.C.: Outlier analysis. Springer International Publishing, 2 edn. (2017)

    work page 2017

  2. [2]

    Transactions on Machine Learning Research (2024)

    Ansari, A.F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S.S., Pineda Arango, S., Kapoor, S., Zschiegner, J., Maddix, D.C., Mahoney, M.W., Torkkola, K., Gordon Wilson, A., Bohlke-Schneider, M., Wang, Y .: Chronos: Learning the language of time series. Transactions on Machine Learning Research (2024)

    work page 2024

  3. [3]

    arXiv preprint arXiv:2005.14165 (2020)

    Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-V oss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A...

    arXiv 2020

  4. [4]

    CRC Press, 6th edn

    Chatfield, C.: The analysis of time series: an introduction. CRC Press, 6th edn. (2004)

    work page 2004

  5. [5]

    In: Advances in Neural Information Processing Systems 37 (2024)

    Dai, Z., He, L., Yang, S., Leeke, M.: Sarad: Spatial association-aware anomaly detection and diagnosis for multivariate time series. In: Advances in Neural Information Processing Systems 37 (2024)

    work page 2024

  6. [6]

    In: Proceedings of the 41st International Conference on Machine Learning (2024)

    Das, A., Kong, W., Sen, R., Zhou, Y .: A decoder-only foundation model for time-series forecasting. In: Proceedings of the 41st International Conference on Machine Learning (2024)

    work page 2024

  7. [7]

    IFAC Proceedings V olumes46(20), 12–17 (2013)

    Ding, Z., Fei, M.: An anomaly detection approach based on isolation forest algorithm for streaming data using sliding window. IFAC Proceedings V olumes46(20), 12–17 (2013)

    work page 2013

  8. [8]

    The Annals of Statistics 7, 1–26 (1979)

    Eforn, B.: Bootstrap methods: another look at the jackknife. The Annals of Statistics 7, 1–26 (1979)

    work page 1979

Show all 52 references
  1. [9]

    In: Proceedings of the 40th IEEE International Conference on Data Engineering

    Fang, Y ., Xie, J., Zhao, Y ., Chen, L., Gao, Y ., Zheng, K.: Temporal-frequency masked au- toencoders for time series anomaly detection. In: Proceedings of the 40th IEEE International Conference on Data Engineering. pp. 1228–1241 (2024)

    2024

  2. [10]

    In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Feng, Y ., Zhang, W., Fu, Y ., Jiang, W., Zhu, J., Ren, W.: Sensitivehue: Multivariate time series anomaly detection by enhancing the sensitivity to normal patterns. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 782–793 (2024)

    2024

  3. [11]

    In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Ghorbani, R., Reinders, M.J., Tax, D.M.: Pate: Proximity-aware time series anomaly evaluation. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 872–883 (2024)

    2024

  4. [12]

    Circulation 101(23), e215–e220 (2000)

    Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E.: Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. Circulation 101(23), e21...

    2000

  5. [13]

    In: Proceedings of the 41st International Conference on Machine Learning (2024)

    Goswami, M., Szafer, K., Choudhry, A., Cai, Y ., Li, S., Dubrawski, A.: Moment: A family of open time-series foundation models. In: Proceedings of the 41st International Conference on Machine Learning (2024)

    2024

  6. [14]

    O’Reilly Media (2019)

    Grus, J.: Data science from scratch: first principles with Python. O’Reilly Media (2019)

    2019

  7. [15]

    arXiv preprint arXiv:2002.08909 (2020)

    Guu, K., Lee, K., Tung, Z., Pasupat, P., Chang, M.: Realm: retrieval-augmented language model pre-training. arXiv preprint arXiv:2002.08909 (2020)

    2020 arXiv

  8. [16]

    In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Huet, A., Navarro, J.M., Rossi, D.: Local evaluation of time series anomaly detection algorithms. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 635–645 (2022)

    2022

  9. [17]

    In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery and data mining

    Hundman, K., Constantinou, V ., Laporte, C., Colwell, I., Soderstrom, T.: Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery and data mining. pp. 387–395 (2018)

    2018

  10. [18]

    In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management

    Hwang, W.S., Yun, J.H., Kim, J., Kim, H.C.: Time-series aware precision and recall for anomaly detection: Considering variety of detection result and addressing ambiguous labeling. In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management...

    2019

  11. [19]

    Hwang, W.S., Yun, J.H., Kim, J., Min, B.G.: Do you know existing accuracy metrics overrate time-series anomaly detections? In: Proceedings of the 37th ACM/SIGAPP Symposium on Applied Computing. pp. 403–412 (2022)

    2022

  12. [20]

    Scientific Data 3(1), 1–9 (2016)

    Johnson, A.E., Pollard, T.J., Shen, L., Lehman, L.w.H., Feng, M., Ghassemi, M., Moody, B., Szolovits, P., Anthony Celi, L., Mark, R.G.: Mimic-iii, a freely accessible critical care database. Scientific Data 3(1), 1–9 (2016)

    2016

  13. [21]

    In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (2021)

    Lai, K.H., Zha, D., Xu, J., Zhao, Y .: Revisiting time series outlier detection: definitions and benchmarks. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (2021)

    2021

  14. [22]

    In: Proceedings of the 14th International Conference on Machine Learning and Applications

    Lavin, A., Ahmad, S.: Evaluating real-time anomaly detection algorithms–the numenta anomaly benchmark. In: Proceedings of the 14th International Conference on Machine Learning and Applications. pp. 38–44 (2015)

    2015

  15. [23]

    arXiv preprint arXiv:2005.11401 (2020)

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V ., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge- intensive NLP tasks. arXiv preprint arXiv:2005.11401 (2020)

    2020 arXiv

  16. [24]

    Proceedings of the VLDB Endowment 17(12), 4229–4232 (2024)

    Liu, Q., Boniol, P., Palpanas, T., Paparrizos, J.: Time-series anomaly detection: Overview and new trends. Proceedings of the VLDB Endowment 17(12), 4229–4232 (2024)

    2024

  17. [25]

    In: Advances in Neural Information Processing Systems 35 (2024)

    Liu, Q., Paparrizos, J.: The elephant in the room: Towards a reliable time-series anomaly detection benchmark. In: Advances in Neural Information Processing Systems 35 (2024)

    2024

  18. [26]

    In: Proceedings of the 39th AAAI Conference on Artificial Intelligence

    Liu, Z., Wang, Y ., Zhang, Y .: Gcad: Anomaly detection in multivariate time series from the perspective of granger causality. In: Proceedings of the 39th AAAI Conference on Artificial Intelligence. pp. 19041–19049 (2025)

    2025

  19. [27]

    IEEE Transactions on Fuzzy Systems 23(3), 688– 700 (2014)

    Moshtaghi, M., Bezdek, J.C., Leckie, C., Karunasekera, S., Palaniswami, M.: Evolving fuzzy rules for anomaly detection in data streams. IEEE Transactions on Fuzzy Systems 23(3), 688– 700 (2014)

    2014

  20. [28]

    arXiv preprint arXiv:2303.08774 (2024)

    OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V ., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I., Berdine, J., Bernadett-Shapi...

    2024 arXiv

  21. [29]

    Proceedings of the VLDB Endowment 15(11), 2774–2787 (2022)

    Paparrizos, J., Boniol, P., Palpanas, T., Tsay, R.S., Elmore, A., Franklin, M.J.: V olume under the surface: A new accuracy evaluation measure for time-series anomaly detection. Proceedings of the VLDB Endowment 15(11), 2774–2787 (2022)

    2022

  22. [30]

    In: Proceedings of the ACM SIGMOD International Conference on Management of Data

    Paparrizos, J., Gravano, L.: k-shape: Efficient and accurate clustering of time series. In: Proceedings of the ACM SIGMOD International Conference on Management of Data. pp. 1855–1870 (2015)

    2015

  23. [31]

    Proceedings of the VLDB Endowment 15(8), 1697–1711 (2022)

    Paparrizos, J., Kang, Y ., Boniol, P., Tsay, R.S., Palpanas, T., Franklin, M.J.: Tsb-uad: An end-to-end benchmark suite for univariate time-series anomaly detection. Proceedings of the VLDB Endowment 15(8), 1697–1711 (2022)

    2022

  24. [32]

    In: Proceedings of the ACM SIGMOD International Conference on Management of Data (2020)

    Paparrizos, J., Liu, C., Elmore, A.J., Franklin, M.J.: Debunking four long-standing misconcep- tions of time-series distance measures. In: Proceedings of the ACM SIGMOD International Conference on Management of Data (2020)

    2020

  25. [33]

    Proceedings of the VLDB Endowment 15(9), 1779–1797 (2022)

    Schmidl, S., Wenig, P., Papenbrock, T.: Anomaly detection in time series: a comprehensive evaluation. Proceedings of the VLDB Endowment 15(9), 1779–1797 (2022)

    2022

  26. [34]

    In: Pro- ceedings of the 13th International Conference on Learning Representations (2025)

    Shen, K.Y .: Learn hybrid prototypes for multivariate time series anomaly detection. In: Pro- ceedings of the 13th International Conference on Learning Representations (2025)

    2025

  27. [35]

    In: Proceedings of the 13th International Conference on Learning Representations (2025)

    Shentu, Q., Li, B., Zhao, K., Shu, Y ., Rao, Z., Pan, L., Yang, B., Guo, C.: Towards a general time series anomaly detector with adaptive bottlenecks and dual adversarial decoders. In: Proceedings of the 13th International Conference on Learning Representations (2025)

    2025

  28. [36]

    In: Proceedings of the 12th International Conference on Learning Representations (2024)

    Shi, X., Wang, S., Nie, Y ., Li, D., Ye, Z., Wen, Q., Jin, M.: Time-moe: Billion-scale time series foundation models with mixture of experts. In: Proceedings of the 12th International Conference on Learning Representations (2024)

    2024

  29. [37]

    In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery and data mining

    Su, Y ., Zhao, Y ., Niu, C., Liu, R., Sun, W., Pei, D.: Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery and data mining. pp. 2828–2837 (2019)

    2019

  30. [38]

    In: Proceedings of the 40th IEEE International Conference on Data Engineering

    Sun, Y ., Pang, G., Ye, G., Chen, T., Hu, X., Yin, H.: Unraveling the ‘anomaly’ in time series anomaly detection: A self-supervised tri-domain solution. In: Proceedings of the 40th IEEE International Conference on Data Engineering. pp. 981–994 (2024)

    2024

  31. [39]

    In: Advances in Neural Information Processing Systems 31 (2018)

    Tatbul, N., Lee, T.J., Zdonik, S., Alam, M., Gottschlich, J.: Precision and recall for time series. In: Advances in Neural Information Processing Systems 31 (2018)

    2018

  32. [40]

    In: Advances in Neural Information Processing Systems 30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems 30 (2017)

    2017

  33. [41]

    In: Proceedings of the 41st International Conference on Machine Learning (2024)

    Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., Sahoo, D.: Unified training of universal time series forecasting transformers. In: Proceedings of the 41st International Conference on Machine Learning (2024)

    2024

  34. [42]

    IEEE Transactions on Knowledge and Data Engineering 35(3), 2421–2429 (2021)

    Wu, R., Keogh, E.J.: Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress. IEEE Transactions on Knowledge and Data Engineering 35(3), 2421–2429 (2021)

    2021

  35. [43]

    In: Proceedings of the 13th International Conference on Learning Representations (2025)

    Wu, X., Qiu, X., Li, Z., Wang, Y ., Hu, J., Guo, C., Xiong, H., Yang, B.: Catch: Channel-aware multivariate time series anomaly detection via frequency patching. In: Proceedings of the 13th International Conference on Learning Representations (2025)

    2025

  36. [44]

    In: Proceedings of the ACM on Web Conference

    Xu, H., Chen, W., Zhao, N., Li, Z., Bu, J., Li, Z., Liu, Y ., Zhao, Y ., Pei, D., Feng, Y ., et al.: Unsupervised anomaly detection via variational auto-encoder for seasonal kpis in web applications. In: Proceedings of the ACM on Web Conference. pp. 187–196 (2018) 12

    2018

  37. [45]

    In: Proceedings of the 10th International Conference on Learning Representations (2022)

    Xu, J., Wu, H., Wang, J., Long, M.: Anomaly transformer: Time series anomaly detection with association discrepancy. In: Proceedings of the 10th International Conference on Learning Representations (2022)

    2022

  38. [46]

    In: Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Yamanishi, K., Takeuchi, J.i.: A unifying framework for detecting outliers and change points from non-stationary time series data. In: Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 676–681 (2002)

    2002

  39. [47]

    In: Proceedings of the 16th IEEE International Conference on Data Mining

    Yeh, C.C.M., Zhu, Y ., Ulanova, L., Begum, N., Ding, Y ., Dau, H.A., Silva, D.F., Mueen, A., Keogh, E.: Matrix profile i: all pairs similarity joins for time series: a unifying view that includes motifs, discords and shapelets. In: Proceedings of the 16th IEEE International Co...

    2016

  40. [48]

    ACM Computing Surveys 57(1), 1–42 (2024)

    Zamanzadeh Darban, Z., Webb, G.I., Pan, S., Aggarwal, C., Salehi, M.: Deep learning for time series anomaly detection: A survey. ACM Computing Surveys 57(1), 1–42 (2024)

    2024

  41. [49]

    In: Advances in Neural Information Processing Systems 35 (2022)

    Zhang, X., Zhao, Z., Tsiligkaridis, T., Zitnik, M.: Self-supervised contrastive pre-training for time series via time-frequency consistency. In: Advances in Neural Information Processing Systems 35 (2022)

    2022

  42. [50]

    In: Proceedings of the 13th International Conference on Learning Representations (2025)

    Zhong, G., Wang, P., Yuan, J., Li, Z., Chen, L.: Multi-resolution decomposable diffusion model for non-stationary time series anomaly detection. In: Proceedings of the 13th International Conference on Learning Representations (2025)

    2025

  43. [51]

    Zhou, Z., Yu, R.: Can llms understand time series anomalies? In: Proceedings of the 13th International Conference on Learning Representations (2025)

    2025

  44. [52]

    start": ...,

    Zhuang, J., Yan, L., Zhang, Z., Wang, R., Zhang, J., Gu, Y .: See it, think it, sorted: Large multi- modal models are few-shot time series anomaly analyzers. arXiv preprint arXiv:2411.02465 (2024) 13 Table 5: Details of the UCR Anomaly Archive. Domain # Time Series # Data Poin...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.