Pith. sign in

REVIEW 3 major objections 7 minor 36 references

Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy

T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A synthetic-only pre-trained transformer detects zero-shot time series anomalies by comparing each window with its context, beating reconstruction-based foundation models on most benchmarks.

desk verdict A serious synthetic-data TSAD paper whose headline zero-shot SOTA claim is inflated by per-dataset test-set window tuning, but whose generator, disclosures, and core framing are real contributions worth engaging. read the letter →

arxiv 2509.21190 v5 pith:ONBVT3XU submitted 2025-09-25 cs.LG cs.AI

Tian Lan , Hao Duong Le , Jinbo Li , Wenjun He , Meng Wang , Chenghao Liu , Chen Zhang This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesanomalydetectionzero-shotlearningsyntheticpre-trainingrelativecontextdiscrepancytransformerfoundationmodelcontextualcausalpropagationTSB-ADbenchmark
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

The paper introduces TimeRCD, a foundation model for zero-shot time series anomaly detection that is pre-trained exclusively on synthetic data. Instead of reconstructing input windows, TimeRCD learns to score anomalies by detecting discrepancies between adjacent time windows, a principle the authors call Relative Context Discrepancy (RCD). The central claim is that this relational approach, trained on a large synthetic corpus with causal and contextual anomaly labels, generalizes better to unseen real-world series than existing zero-shot foundation models. On 56 univariate benchmark cases it ranks first in 41 and second in 6, and it remains competitive with full-shot baselines. If correct, the result shows that a purpose-built synthetic curriculum with explicit relational supervision is a viable path toward generalizable anomaly detectors.

What carries the argument

The central mechanism is Relative Context Discrepancy (RCD), the rule that an anomaly is defined by a significant discrepancy between adjacent time windows rather than by any absolute pattern. In TimeRCD, each non-overlapping window of the time series is linearly projected into a token and fed to a standard encoder-only Transformer; self-attention between these tokens computes the RCD features, and an anomaly head converts them into window-level scores. A reconstruction head is kept only as an auxiliary training task and discarded at inference. The second component is the synthetic generator: a four-stage pipeline that builds univariate context from trend, seasonality, and noise templates; weaves channels into a causal ARX system; injects both exogenous shocks and endogenous failures that propagate through the causal graph; and labels root causes and propagated effects at token level.

What would settle it

Construct a new anomaly benchmark whose anomaly types, durations, amplitudes, or causal structures are intentionally out of the generator's prior (for example, anomalies that appear only in low-frequency components or that involve feedback loops between channels) and measure TimeRCD's zero-shot performance on it; if performance collapses toward the reconstruction baselines, the transfer assumption fails.

Watch

Extended reading notes

Core claim

The paper argues that reconstruction-based objectives are fundamentally mismatched to zero-shot anomaly detection: they smooth over subtle anomalies and fail on complex normal patterns. TimeRCD replaces this with a pre-training objective that treats anomaly detection as comparing a query pattern with its surrounding context, implemented by treating each time window as a Transformer token and using self-attention to compute the discrepancy between windows. Trained on a 2.5-billion-point synthetic corpus with more than twenty anomaly archetypes, including endogenous anomalies that propagate through an ARX causal system, the model generalizes zero-shot to 14 public benchmarks. The paper's claim is that TimeRCD establishes a new state of the art for zero-shot TSAD on most of these datasets, and that the design of the synthetic data, not just model scale, is what drives the improvement.

Load-bearing premise

The hand-designed synthetic generator's range of normal patterns and anomaly templates is broad enough to cover the space of real-world anomalies, so that labels learned on it teach a universal context-discrepancy detector; this transfer assumption is never directly tested against generator priors outside the evaluation sets.

Editorial extensions

If this is right

  • If the transfer assumption holds, zero-shot TSAD no longer requires any target-domain labels or fine-tuning; a single pre-trained model can be dropped onto new sensors, servers, or industrial logs.
  • Reconstruction-based scoring becomes a secondary diagnostic rather than the primary signal for anomaly detection, at least for the types of anomalies covered by the synthetic curriculum.
  • The positive scaling law on synthetic data means that further enlarging the corpus—more anomaly archetypes, longer series, denser causal graphs—should keep improving zero-shot performance without touching the model architecture.
  • The RCD formulation gives a principled reason to use long context windows: the model's ability to detect contextual anomalies scales with its ability to see the whole series, which is why window size matters most on datasets with long-term patterns.
  • The ablations suggest that a pure anomaly head collapses without the auxiliary reconstruction task, so the two heads are not redundant: reconstruction stabilizes the learned representations that the anomaly head refines.

Reading between the lines

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

  • The strongest testable implication is that the synthetic generator's priors are the true bottleneck: if the generator's amplitude, duration, and frequency priors are narrowed, TimeRCD's advantage should shrink exactly on the subtle contextual anomalies it is designed to catch. A sensitivity sweep over generator hyperparameters would directly test this.
  • A neighbouring problem that RCD might transfer to is change-point detection, since both tasks require comparing adjacent contexts. TimeRCD's contextual-discrepancy head could plausibly be fine-tuned to flag regime shifts, though the paper does not test this.
  • A hard out-of-distribution test would be to evaluate TimeRCD on anomaly types the generator cannot produce, such as anomalies spanning only a single variate in a system where the causal graph has feedback loops, which the ARX model cannot represent.
  • The paper's per-dataset window tuning (5k for most univariate, 13k for all multivariate) is a degree of freedom that a strict zero-shot evaluation could treat as a hyperparameter to be fixed before seeing the test set; a sensitivity analysis fixing a single global window would strengthen the zero-shot claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. This paper introduces TimeRCD, a zero-shot time-series anomaly detection (TSAD) foundation model that replaces reconstruction-error scoring with Relative Context Discrepancy (RCD). A standard encoder-only Transformer treats fixed-length windows of a multivariate series as tokens; self-attention captures discrepancies between a query window and its surrounding context, and a dedicated anomaly head converts the resulting embeddings into window-level anomaly scores. Pre-training uses only a synthetic corpus generated by a four-stage engine (Appx. C): univariate trend/seasonality/noise templates (Stage 1), coupling into a causal multivariate ARX system over a random DAG (Stage 2), injection of more than 20 exogenous, endogenous-propagating, and seasonal/contextual anomaly archetypes with token-level labels (Stages 3-4), for a total of 2.5B data points. The paper claims strict zero-shot SOTA: on 56 dataset-metric cases from 14 TSB-AD datasets, TimeRCD ranks first in 41 and second in 6 cases against zero-shot foundation-model baselines, and first in 28 and second in 5 cases against full-shot, dataset-specific baselines, across four metrics (Affiliation-F1, F1-T, Standard-F1, VUS-PR). Additional experiments cover specialized contextual-anomaly test sets, window-size sensitivity, an ablation of the auxiliary reconstruction head, an ablation comparing synthetic versus real pre-training data, and scaling-law results up to 2.5B points.

Significance. The RCD formulation is a conceptually clean departure from the reconstruction-based scoring used by nearly all existing TSAD foundation models, and the paper's motivation linking reconstruction failure modes (missed subtle anomalies, false alarms on complex normal patterns) to zero-shot performance is well argued. If the empirical claims hold, the paper would establish both a new pre-training objective for TSAD and a reusable large-scale synthetic benchmark with unusually detailed formal definitions of 20+ anomaly archetypes (Table 6). The evaluation is broader than typical TSAD work: four metrics, a conscious avoidance of point-adjusted F1 (which the authors correctly note is inflated), and comparisons against both zero-shot and full-shot baselines; the appendix also discloses the otherwise problematic window-selection protocol, and code is promised. The significance is currently conditional, however: the headline numbers conflate the method's transferable capability with per-dataset test-set-tuned context windows, and several table values are internally inconsistent, so the central SOTA claim needs a protocol-level cleanup before the contribution can be fully assessed.

major comments (3)
  1. [Section 3.2; Appx. E.1; Appx. E.3; Table 10] Section 3.2 claims that TimeRCD is 'strictly zero-shot in all settings, testing true out-of-the-box performance,' but Appx. E.1 states that for the main results in Table 1 'a context window size was selected for each dataset based on the analysis in Appendix E.3 to optimize performance,' and Appx. E.3 sweeps windows from 1k to 13k directly on the target benchmark test sets, with Table 10 reporting the resulting scores computed against the ground-truth labels of those test sets. The choice is consequential: Table 10 shows UCR Standard-F1 rising from 20.30 (1k) to 29.92 (13k), Power Affiliation-F from 69.30 (1k) to 79.73 (13k), and SMAP Affiliation-F from 80.16 (1k) to 84.10 (13k). Baselines, by contrast, use fixed windows chosen a priori (DADA 100, MOMENT 64, TimesFM 96, etc., Appx. D.2). Because the headline ranking counts (41 first / 6 second; 28 first / 5 second) are computed from Table 1, the comparison is not apples-to-apples and the 'strictly zero-shot' framing is internally contradicted. Please re-report the main results with a fixed window rule (or with windows chosen on a validation split that does not access the target test labels), and present the window-tuned numbers separately if they are kept.
  2. [Table 1 vs Table 10] Table 1 and Table 10 disagree for several configurations that should coincide. A subset of entries matches the appendix sweep at the stated windows (e.g., UCR and YAHOO at 5k), but others deviate substantially: at the 13k window that Appx. E.1 says was used for all multivariate datasets, Table 1 reports SMD Affiliation-F 92.58 and SMAP 87.73, while Table 10 gives 88.29 and 84.10, and Table 1's Power Affiliation-F of 85.51 exceeds the maximum Power value in the entire 1k-13k sweep (79.73). Table 1 also reports IOPS 83.28 and NAB 82.48 at the stated windows, versus 84.79 (5k) and 79.70 (5k) in Table 10. The manuscript does not explain whether Table 1 and Table 10 use different aggregation (pooled versus per-series) or different runs, and the Reproducibility Statement does not address this discrepancy. Because the ranking claims are built from Table 1, the authors must reconcile the two tables and state the exact scoring configuration used for the main results.
  3. [Appx. E.2; Appx. C.1/C.4; Reproducibility Statement] The specialized contextual-anomaly evaluation (Appx. E.2, Table 9) uses test sets generated by the same synthetic engine used for pre-training: although the sets are unseen, they are in-distribution with respect to the generator, so the strong contextual-anomaly advantage reported in Table 9 does not by itself demonstrate zero-shot transfer to unseen real-world anomaly types; the only out-of-distribution evidence is the TSB-AD table, which is confounded by the window-selection issue in the first comment. In addition, the Reproducibility Statement says the generator is 'fully specified' in Appx. C, but the anomaly taxonomy of Table 6 gives no sampling priors for the parameters that control anomaly salience and shape (amplitudes A, durations w, rise/fall times tau_r and tau_f, carrier frequencies f_h, and the modifiers r, rho, lambda), while Tables 4 and 5 specify priors only for the normal-context and causal-dynamics components. Please add the missing priors and add at least one genuinely out-of-distribution test (for example, a held-out generator variant or real anomalies from a domain absent from the pre-training curriculum) so that the transfer hypothesis is directly testable.
minor comments (7)
  1. [Section 2.1] The sentence 'The resulting windows are are then flattened and linearly projected' contains a doubled 'are'; please fix.
  2. [Appx. B; Appx. D.3] There are several typos, e.g., 'objective dismatch' (Appx. B) and 'the implmentation of Sarfraz et al.' (Appx. D.3); a careful proofread is needed.
  3. [Section 3.2] The phrase 'on 56 univariate cases' is inaccurate: the 56 cases consist of four metrics across all 14 datasets, including the five multivariate datasets; please write '56 dataset-metric cases' or similar.
  4. [Appx. D.4; Reproducibility Statement] The Reproducibility Statement says the masking strategy is detailed in Appx. D.4, but Appx. D.4 mentions only that an MSE loss is used for 'a masked reconstruction task' and does not specify the mask ratio, mask length, or masking scheme; please add this detail.
  5. [Section 2.1; Appx. C Stage 4; Appx. D.4] The anomaly head is trained with a Cross-Entropy loss on window-level scores, while the synthetic labels are token-level per channel (Appx. C, Stage 4); please specify how token-level labels are converted into window-level supervision (e.g., any-positive aggregation or per-token weighting).
  6. [Appx. D.4] The statement that the batch size 'starts at 64 but is adjusted dynamically' is imprecise; please state when and by what rule the batch size changes.
  7. [Section 2.2; Appx. E.2] Stage 4 of the generator produces separate root-cause and propagated-effect masks and says the corpus supports 'evaluation that distinguishes localization from detection,' but the evaluations in Section 3 and Appx. E.2 do not use this distinction; please clarify whether such an evaluation is performed.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'strictly zero-shot SOTA' ranking is partially constructed from test-set-tuned context windows, and the contextual-anomaly evidence is in-distribution by construction.

  1. fitted input called prediction [Section 3.2 (RQ1) and Appendix E.1]
    "We stress that TimeRCD is strictly zero-shot in all settings, testing true out-of-the-box performance. ... In zero-shot comparisons, TimeRCD achieves clear SOTA: on 56 univariate cases it ranks first in 41 and second in 6. ... For the main results presented in Table 1, a context window size was selected for each dataset based on the analysis in Appendix E.3 to optimize performance."

    The paper's central 'strictly zero-shot SOTA' claim is reported after selecting a per-dataset context window by optimizing performance on the target test sets (Appendix E.3 sweeps windows from 1k to 13k), while baselines use fixed a priori windows (DADA=100, MOMENT=64, TimesFM/Time-MOE=96, Appx D.2). Table 10 shows the chosen window materially changes scores (e.g., UCR Standard-F1 is 20.30 at 1k vs 29.92 at 13k; IOPS Affiliation-F is 85.64 at 1k vs 83.98 at 13k). Thus the 41/6 rank advantage is not out-of-the-box; a test-set-informed hyperparameter is fitted and then presented as a zero-shot prediction.

  2. self definitional [Appendix E.2 (Table 9) and Section 3.3]
    "Each dataset was generated using the framework described in the main paper, with a crucial modification: only one type of anomaly was injected per set. ... These datasets were not used in any part of the pre-training process."

    The specialized contextual-anomaly test sets used to demonstrate TimeRCD's 'substantially superior' contextual capability are generated by the same four-stage synthetic engine and the same anomaly taxonomy (trend/seasonality templates, ARX causal model, and the Section C.4 anomaly archetypes) used to build the 2.5B pre-training corpus. The 'contextual anomaly' definition in the test set is therefore the paper's own construction, identical to the supervised signal used during pre-training. The claimed contextual advantage (Standard-F1 0.827 vs collapse of other models) is an in-distribution, same-generator evaluation, so it is partly defined by the training input rather than providing independent out-of-distribution evidence for the RCD principle.

full rationale

The core evaluation on the external TSB-AD benchmarks is genuinely independent of the synthetic pre-training distribution, and the model is not trained on TSB-AD labels, so the RCD pre-training idea itself is not circular. However, the flagship 'strictly zero-shot SOTA' ranking is partially constructed: the per-dataset context windows used for Table 1 were selected by optimizing performance on the same target test sets (Appendix E.1/E.3), whereas the baselines use fixed a priori windows, so the 41/6 rank counts are test-set-informed rather than truly out-of-the-box. Additionally, the contextual-anomaly evidence (Figure 5/Table 9) is a same-generator, in-distribution evaluation: the test sets are produced by the same synthetic engine and anomaly taxonomy used for pre-training, so the contextual-advantage claim is partly defined by the paper's own construction. Self-citations (e.g., Moirai's any-variate attention) are architectural building blocks and are not load-bearing. No uniqueness theorem or derived equivalence makes the method's success follow by definition; the partial circularity comes from the test-set-tuned protocol and the same-generator contextual benchmark. Overall: partial circularity, score 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on two hand-built inductive biases. The first is the synthetic generator's prior distribution, which defines what 'normal' and 'anomalous' mean for all pre-training labels; its anomaly template priors are not fully specified numerically. The second is the set of per-dataset context windows, which are selected using target benchmark performance. Standard background math (ARX stability, transformer expressivity) is comparatively safe. No new entities are introduced; the evaluation also assumes TSB-AD labels are truthful.

free parameters (3)
  • Per-dataset context window size = 5k default univariate; 1k NAB; 13k Power and all multivariate
    Appx E.1 states the window was chosen per dataset 'based on the analysis in Appendix E.3 to optimize performance', i.e., using the target test sets; window choice changes Affiliation-F by up to 17 points (Power) and is a test-set-tuned hyperparameter.
  • Anomaly template priors (amplitudes A, durations w, tau_r, tau_f, carrier frequencies) = not numerically specified
    Stage 3 and Appx C.4 give formulas for 20+ archetypes but not the sampling distributions for amplitudes and durations; these priors define the label distribution the RCD head is trained on and are the main control over what the model can detect.
  • ARX causal priors (sigma_b^2, l_max, alpha_i, edge probability) = a_i ~ U[-0.8,0.8]; lags ~ U{0..l_max}; b_ij ~ N(0, sigma_b^2) with sigma_b^2 and l_max unstated; alpha_i in [0,1] with…
    Stage 2 and Appx C.2 shape causal propagation of endogenous anomalies; the paper states ranges but not the distributions, so the multivariate curriculum is not fully reproducible from the text alone.
assumptions (4)
  • domain assumption Real-world anomalies are reliably detectable as discrepancies between a window and its surrounding context (the RCD premise)
    Section 2.1 motivates the entire method; if a large class of real anomalies (e.g., global level shifts, distribution-wide drifts without local contrast) does not disrupt local context, the scoring premise fails.
  • ad hoc to paper The hand-designed synthetic distribution (trend/seasonality/noise templates, ARX coupling, 20+ archetypes) covers the real-world anomaly space well enough for zero-shot transfer
    Section 2.2 and Appx C; the only direct contextual evidence (Table 9) is in-distribution (same generator), and the external benchmarks are confounded by window tuning, so transfer coverage is assumed rather than demonstrated.
  • domain assumption TSB-AD labels and the four chosen metrics are faithful for comparing detectors
    Section 3.1 and Appx D.1, D.3; the paper rejects PA-F1 citing Huet 2022 and Wang 2023a, but the remaining metrics still encode specific tolerance to delay, and the headline ranks aggregate them without weighting.
  • standard math A standard encoder transformer can represent and learn context-discrepancy scoring from supervised window labels
    Section 2.1; supported empirically by the ablations, not by a derivation; residual risk is capacity and optimization rather than logic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy." pith.science (2026). https://pith.science/paper/ONBVT3XU

@misc{pith2026250921190,
  author       = {Pith},
  title        = {Pith review of: Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ONBVT3XU}},
  note         = {Machine review of arXiv:2509.21190}
}
read the original abstract

Time series anomaly detection (TSAD) is a critical task, but developing models that generalize to unseen data in a zero-shot manner remains challenging. Existing foundation models for TSAD often rely on reconstruction-error scoring at inference time, which can miss subtle anomalies that are well reconstructed and can falsely flag complex but normal patterns in unseen domains. We introduce TimeRCD, a foundation model for TSAD built on Relative Context Discrepancy (RCD), a pre-training paradigm that trains the model to detect anomalies by comparing a query pattern with its surrounding context. This relational formulation, implemented with a standard Transformer architecture, enables the model to infer normality from the input context rather than relying on fixed global normal patterns. We further construct a large-scale synthetic corpus with context-dependent anomaly labels to provide supervised pre-training signals for RCD. Experiments across diverse benchmarks show that TimeRCD outperforms existing general-purpose and anomaly-specific foundation models in most zero-shot TSAD settings, while remaining competitive with dataset-specific full-shot baselines. These results provide empirical evidence that RCD is an effective direction for building robust and generalizable TSAD models.

Figures

Figures reproduced from arXiv: 2509.21190 by the authors.

Figure 1
Figure 1. Limitations of Reconstruction-based TSAD and Our Proposed TimeRCD. (a) The model ac￾curately reconstructs a smooth anomaly, resulting in a low error score and a missed detection (false negative). (b) The model fails to reconstruct a complex normal pattern unseen in the training dataset, leading to a high error score and a false alarm (false positive). (c) TimeRCD adopts RCD with a large-scale corpus and a standard T… view at source ↗
Figure 2
Figure 2. TimeRCD Architec￾ture Problem Definition For the zero-shot time series anomaly detec￾tion problem, we observe a multivariate d-channel time series X = (x1, . . . , xn) with xt ∈ R d for each time step t ∈ [n] := {1, 2, . . . , n}. The objective is to produce a binary annotation sequence yˆ = (ˆy1, . . . , yˆn) ∈ {0, 1} n such that yˆt = 1 if and only if time t is anomalous. In the zero-shot setting, the model must d… view at source ↗
Figure 3
Figure 3. Synthetic data generation procedure. 2.2 SYNTHETIC DATA GENERATION We design a synthetic engine to generate multi￾variate time series with rich, controllable contex￾tual structures, creating a TSAD benchmark that en￾courages models to recognize anomalies relative to context. The pipeline proceeds in three hierarchical stages: first, defining univariate contextual patterns (Phase 1); next, integrating them into a mul… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of anomaly scores. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison on specially-created datasets containing either point or contextual anomalies. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance on univariate (top) and multivariate (bottom) datasets as a function of the input [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Demonstration of positive scaling laws. The figure shows weighted average performance across [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Distribution of time series lengths within our 2.5B point synthetic pre-training dataset. The his [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Affiliation-F performance across all 14 datasets as a function of the input window size from 1k to [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]
Figure 10
Figure 10. Figure 10: F1T performance across all 14 datasets as a function of the input window size from 1k to 13k. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_10.png]
Figure 11
Figure 11. Figure 11: Standard-F1 performance across all 14 datasets as a function of the input window size from 1k to [PITH_FULL_IMAGE:figures/full_fig_p032_11.png]
Figure 12
Figure 12. Figure 12: VUS-PR performance across all 14 datasets as a function of the input window size from 1k to [PITH_FULL_IMAGE:figures/full_fig_p033_12.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [1]

    •Increase after downward spike

    =1 [t1,∞)(t). •Increase after downward spike. Using a negative spike template∆ nspk(t)and a positive shift B >0: ∆(t) = ∆nspk(t) +BH(t;t 1). •Increase after upward spike. Use ∆(t) = ∆spk(t) +BH(t;t 1). •Decrease after downward spike. Use ∆(t) = ∆nspk(t)−BH(t;t 1). High-frequency burst •Shake(localized high-frequency vibration). Parameters: carrier frequen...

    work page 2026

  2. [7]

    Monash time series forecasting archive

    Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I Webb, Rob J Hyndman, and Pablo Montero-Manso. Monash time series forecasting archive. Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. Moment: A family of open time-series foundation models.arXiv preprint arXiv:2402.03885,

    arXiv

  3. [8]

    From tables to time: How tabpfn-v2 outper- forms specialized time series forecasting models.arXiv preprint arXiv:2501.02945,

    Shi Bin Hoo, Samuel M¨uller, David Salinas, and Frank Hutter. From tables to time: How tabpfn-v2 outper- forms specialized time series forecasting models.arXiv preprint arXiv:2501.02945,

    arXiv

  4. [9]

    The table is grouped by anomaly type, with metrics listed vertically for each

    28 Under review as a conference paper at ICLR 2026 Table 9: Quantitative comparison on specialized Point and Contextual anomaly datasets. The table is grouped by anomaly type, with metrics listed vertically for each. Our model demonstrates competitive performance on point anomalies and substantially superior performance on contextual anomalies across all ...

    work page 2026

  5. [10]

    Breaking the time-frequency granularity discrepancy in time-series anomaly detection

    Youngeun Nam, Susik Yoon, Yooju Shin, Minyoung Bae, Hwanjun Song, Jae-Gil Lee, and Byung Suk Lee. Breaking the time-frequency granularity discrepancy in time-series anomaly detection. InProceedings of the ACM Web Conference 2024, pp. 4204–4215,

    work page 2024

  6. [11]

    A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    arXiv

  7. [12]

    Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders

    Qichao Shentu, Beibu Li, Kai Zhao, Yang Shu, Zhongwen Rao, Lujia Pan, Bin Yang, and Chenjuan Guo. Towards a general time series anomaly detector with adaptive bottlenecks and dual adversarial decoders. arXiv preprint arXiv:2405.15273,

    work page Pith review arXiv

  8. [13]

    Time-moe: Billion-scale time series foundation models with mixture of experts.arXiv preprint arXiv:2409.16040,

    11 Under review as a conference paper at ICLR 2026 Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts.arXiv preprint arXiv:2409.16040,

    arXiv 2026

Show all 36 references
  1. [14]

    Cauker: classification time series foundation models can be pretrained on synthetic data only.arXiv preprint arXiv:2508.02879,

    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.arXiv preprint arXiv:2508.02879,

  2. [15]

    Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning

    Zhe Xie, Zeyan Li, Xiao He, Longlong Xu, Xidao Wen, Tieying Zhang, Jianjun Chen, Rui Shi, and Dan Pei. Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning. arXiv preprint arXiv:2412.03104,

  3. [16]

    Anomaly transformer: Time series anomaly detection with association discrepancy.arXiv preprint arXiv:2110.02642,

    Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy.arXiv preprint arXiv:2110.02642,

    arXiv

  4. [17]

    A USE OFLLMS In preparing our manuscript, we used large language models (LLMs) only to aid or polish the writing. Their use was restricted to improving grammar, readability, and style, without contributing to the methodology, theoretical results, algorithmic implementation, or...

    2026

  5. [18]

    16 Under review as a conference paper at ICLR 2026 Table 5: Multivariate and causal-dynamics hyperparameters and intended roles. Name Typical range / rule Applies to Intended role Number of nodes Uniformly in 1 to 50 system graph Dimensionality of the multivariate system.Numbe...

    2026

  6. [19]

    For any scalaru, the logistic sigmoid isσ(u) = 1 1+e−u

    =1 [t0,∞)(t). For any scalaru, the logistic sigmoid isσ(u) = 1 1+e−u . For a rectangular window we useh(t;t s,te) =1 [ts,te)(t). All amplitudesA,B,{A m} are real scalars; all durations and lags (e.g.,w,τ r,τf,d) are positive integers unless otherwise stated. 17 Under review as...

    2026

  7. [20]

    19 Under review as a conference paper at ICLR 2026 •Sudden decrease(permanent level down-shift)

    , σ(u) = 1 1 +e−u. 19 Under review as a conference paper at ICLR 2026 •Sudden decrease(permanent level down-shift). Use ∆(t) =−Aσ κ(t−t

    2026

  8. [22]

    Fort∈[t s,te), scale the duty cycle by a factorλ >0:d7→d ′ = min{1,max{0,λd}}

    21 Under review as a conference paper at ICLR 2026 •Pulse width modulation. Fort∈[t s,te), scale the duty cycle by a factorλ >0:d7→d ′ = min{1,max{0,λd}}. Wavelet-based (atoms and families)Let a wavelet-based seasonal component be represented as a finite sum of atoms S(t) = LX...

    2024

  9. [23]

    Zero-Shot ModelsThese models are pre-trained on large-scale datasets and can be applied directly to new time series without fine-tuning

    and the DADA (Shentu et al., 2024). Zero-Shot ModelsThese models are pre-trained on large-scale datasets and can be applied directly to new time series without fine-tuning. •DADA: This model, from the paper by (Shentu et al., 2024), is a pre-trained general anomaly detector. W...

    2024

  10. [27]

    Its window size is

    •Chronos: A generative model that frames time series analysis as a language modeling task, using a transformer-based architecture to learn and predict time series values (Ansari et al., 2024). Its window size is

    2024

  11. [28]

    It is pre-trained on a large time series corpus for forecasting and multi-task learning (Shi et al., 2024)

    •Time MOE: A decoder-only transformer model with a sparse Mixture-of-Experts (MoE) architec- ture. It is pre-trained on a large time series corpus for forecasting and multi-task learning (Shi et al., 2024). We set the window size to

    2024

  12. [29]

    •TranAD: A transformer-based model that uses a reconstructive approach to detect anomalies by comparing original and reconstructed time series (Tuli et al., 2022)

    Full-Shot ModelsThese models require training on the target dataset. •TranAD: A transformer-based model that uses a reconstructive approach to detect anomalies by comparing original and reconstructed time series (Tuli et al., 2022). It is configured with a window size of

    2022

  13. [30]

    We set the window size to

    •USAD: An autoencoder-based model that employs an adversarial training mechanism to enhance its reconstruction capability and anomaly detection (Audibert et al., 2020). We set the window size to

    2020

  14. [31]

    The model is configured with a window size of

    •OmniAnomaly: A deep learning model that uses a Variational Autoencoder (V AE) with a Gated Recurrent Unit (GRU) to learn normal patterns and detect deviations (Su et al., 2019). The model is configured with a window size of

    2019

  15. [32]

    For univariate datasets, we set n neighbors to 50, and for multivariate, we set n neighbors to 50 with metric as euclidean

    •LOF: A traditional statistical method that measures the local deviation of a data point from its neighbors, identifying anomalies with lower local density (Breunig et al., 2000). For univariate datasets, we set n neighbors to 50, and for multivariate, we set n neighbors to 50...

    2000

  16. [33]

    For Temporal F1, we follow the implmentation of Sarfraz et al

    for Affiliation F1 (F1 A), Standard F1 (F1), and VUS-PR. For Temporal F1, we follow the implmentation of Sarfraz et al. (2024). •Standard-F1The Standard F1-score is a widely used metric that provides a harmonic mean of precision and recall. It is calculated using point-wise Tr...

    2024

  17. [34]

    The total loss is a sum of two components: a Mean Squared Error loss for a masked reconstruction task and a Cross-Entropy loss for an anomaly detection task

    The model is trained using theAdamWoptimizer with a learning rate of 5e-4 and a weight decay of 1e-5. The total loss is a sum of two components: a Mean Squared Error loss for a masked reconstruction task and a Cross-Entropy loss for an anomaly detection task. Training runs for...

    2026

  18. [36]

    generally appears less sensitive to changes in window size compared to the other three metrics. This suggests that while a larger context may be crucial for the precise, point-wise accuracy captured by metrics like ‘Standard-F1‘, the broader, proximity-based evaluation of ‘Aff...

    2024

  19. [64]

    We set the window size to

    •TimesFM: A decoder-only transformer model from Google trained on a large time series corpus using a patching strategy, enabling strong zero-shot generalization (Das et al., 2024). We set the window size to

    2024

  20. [96]

    We use a window size of

    •MOMENT: A foundation model that utilizes a patch-based pre-training strategy to learn rich rep- resentations from diverse time series data (Goswami et al., 2024). We use a window size of

    2024

  21. [100]

    It is pre-trained on a massive corpus for multi-task applications, including anomaly detection

    24 Under review as a conference paper at ICLR 2026 •TS-Pulse: An IBM foundation model that uses a TSMixer architecture (Ekambaram et al., 2025). It is pre-trained on a massive corpus for multi-task applications, including anomaly detection. Fol- lowing the suggestion from the ...

    2026

  22. [2000]

    Timeseriesexam: A time series understanding exam.arXiv preprint arXiv:2410.14752,

    Yifu Cai, Arjun Choudhry, Mononito Goswami, and Artur Dubrawski. Timeseriesexam: A time series understanding exam.arXiv preprint arXiv:2410.14752,

    arXiv

  23. [2016]

    Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815,

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Olek- sandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, et al. Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815,

    arXiv

  24. [2019]

    Cicada: Cross-domain interpretable coding for anomaly detection and adaptation in multivariate time series.arXiv preprint arXiv:2505.00415,

    Tian Lan, Yifei Gao, Yimeng Lu, and Chen Zhang. Cicada: Cross-domain interpretable coding for anomaly detection and adaptation in multivariate time series.arXiv preprint arXiv:2505.00415,

    arXiv

  25. [2020]

    Mamba4cast: Efficient zero-shot time series forecasting with state space models.arXiv preprint arXiv:2410.09385,

    Sathya Kamesh Bhethanabhotla, Omar Swelam, Julien Siems, David Salinas, and Frank Hutter. Mamba4cast: Efficient zero-shot time series forecasting with state space models.arXiv preprint arXiv:2410.09385,

    arXiv

  26. [2023]

    Tspulse: Dual space tiny pre-trained models for rapid time- series analysis.arXiv preprint arXiv:2505.13033,

    Vijay Ekambaram, Subodh Kumar, Arindam Jati, Sumanta Mukherjee, Tomoya Sakai, Pankaj Dayama, Wesley M Gifford, and Jayant Kalagnanam. Tspulse: Dual space tiny pre-trained models for rapid time- series analysis.arXiv preprint arXiv:2505.13033,

  27. [2024]

    Lof: identifying density-based local outliers

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and J ¨org Sander. Lof: identifying density-based local outliers. InProceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 93–104,

    2000

  28. [2025]

    Units: A unified multi-task time series model.Advances in Neural Information Processing Systems, 37: 140589–140631,

    10 Under review as a conference paper at ICLR 2026 Shanghua Gao, Teddy Koker, Owen Queen, Tom Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. Units: A unified multi-task time series model.Advances in Neural Information Processing Systems, 37: 140589–140631,

    2026

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.